[RFC,v2,2/2] scheduler: add scheduler level for clusters

Commit Message

Song Bao Hua (Barry Song) Dec. 1, 2020, 2:59 a.m. UTC

ARM64 server chip Kunpeng 920 has 6 clusters in each NUMA node, and each
cluster has 4 cpus. All clusters share L3 cache data, but each cluster
has local L3 tag. On the other hand, each clusters will share some
internal system bus. This means cache coherence overhead inside one cluster
is much less than the overhead across clusters.

+-----------------------------------+                          +---------+
|  +------+    +------+            +---------------------------+         |
|  | CPU0 |    | cpu1 |             |    +-----------+         |         |
|  +------+    +------+             |    |           |         |         |
|                                   +----+    L3     |         |         |
|  +------+    +------+   cluster   |    |    tag    |         |         |
|  | CPU2 |    | CPU3 |             |    |           |         |         |
|  +------+    +------+             |    +-----------+         |         |
|                                   |                          |         |
+-----------------------------------+                          |         |
+-----------------------------------+                          |         |
|  +------+    +------+             +--------------------------+         |
|  |      |    |      |             |    +-----------+         |         |
|  +------+    +------+             |    |           |         |         |
|                                   |    |    L3     |         |         |
|  +------+    +------+             +----+    tag    |         |         |
|  |      |    |      |             |    |           |         |         |
|  +------+    +------+             |    +-----------+         |         |
|                                   |                          |         |
+-----------------------------------+                          |   L3    |
                                                               |   data  |
+-----------------------------------+                          |         |
|  +------+    +------+             |    +-----------+         |         |
|  |      |    |      |             |    |           |         |         |
|  +------+    +------+             +----+    L3     |         |         |
|                                   |    |    tag    |         |         |
|  +------+    +------+             |    |           |         |         |
|  |      |    |      |            ++    +-----------+         |         |
|  +------+    +------+            |---------------------------+         |
+-----------------------------------|                          |         |
+-----------------------------------|                          |         |
|  +------+    +------+            +---------------------------+         |
|  |      |    |      |             |    +-----------+         |         |
|  +------+    +------+             |    |           |         |         |
|                                   +----+    L3     |         |         |
|  +------+    +------+             |    |    tag    |         |         |
|  |      |    |      |             |    |           |         |         |
|  +------+    +------+             |    +-----------+         |         |
|                                   |                          |         |
+-----------------------------------+                          |         |
+-----------------------------------+                          |         |
|  +------+    +------+             +--------------------------+         |
|  |      |    |      |             |   +-----------+          |         |
|  +------+    +------+             |   |           |          |         |
|                                   |   |    L3     |          |         |
|  +------+    +------+             +---+    tag    |          |         |
|  |      |    |      |             |   |           |          |         |
|  +------+    +------+             |   +-----------+          |         |
|                                   |                          |         |
+-----------------------------------+                          |         |
+-----------------------------------+                         ++         |
|  +------+    +------+             +--------------------------+         |
|  |      |    |      |             |  +-----------+           |         |
|  +------+    +------+             |  |           |           |         |
|                                   |  |    L3     |           |         |
|  +------+    +------+             +--+    tag    |           |         |
|  |      |    |      |             |  |           |           |         |
|  +------+    +------+             |  +-----------+           |         |
|                                   |                          +---------+
+-----------------------------------+

This patch adds the sched_domain for clusters. On kunpeng 920, without
this patch, domain0 of cpu0 would be MC for cpu0-cpu23 with
min_interval=24, max_interval=48; with this patch, MC becomes domain1,
a new domain0 "CL" including cpu0-cpu3 is added with min_interval=4 and
max_interval=8.
This will affect load balance. For example, without this patch, while cpu0
becomes idle, it will pull a task from cpu1-cpu15. With this patch, cpu0
will try to pull a task from cpu1-cpu3 first. This will have much less
overhead of task migration.

On the other hand, while doing WAKE_AFFINE, this patch will try to find
a core in the target cluster before scanning the llc domain.
This means it will proactively use a core which has better affinity with
target core at first.

Not much benchmark has been done yet. but here is a rough hackbench
result.
we run the below command with different -g parameter to increase system load
by changing g from 1 to 4, for each one of 1-4, we run the benchmark ten times
and record the data to get the average time:

First, we run hackbench in only one NUMA node(cpu0-cpu23):
$ numactl -N 0 hackbench -p -T -l 100000 -g $1

g=1 (seen cpu utilization around 50% for each core)
Running in threaded mode with 1 groups using 40 file descriptors
Each sender will pass 100000 messages of 100 bytes
w/o: 7.689 7.485 7.485 7.458 7.524 7.539 7.738 7.693 7.568 7.674=7.5853
w/ : 7.516 7.941 7.374 7.963 7.881 7.910 7.420 7.556 7.695 7.441=7.6697
performance improvement w/ patch: -1.01%

g=2 (seen cpu utilization around 70% for each core)
Running in threaded mode with 2 groups using 40 file descriptors
Each sender will pass 100000 messages of 100 bytes
w/o: 10.127 10.119 10.070 10.196 10.057 10.111 10.045 10.164 10.162 9.955=10.1006
w/ : 9.694 9.654 9.612 9.649 9.686 9.734 9.607 9.842 9.690 9.710=9.6878
performance improvement w/ patch: 4.08%

g=3 (seen cpu utilization around 90% for each core)
Running in threaded mode with 3 groups using 40 file descriptors
Each sender will pass 100000 messages of 100 bytes
w/o: 15.885 15.254 15.932 15.647 16.120 15.878 15.857 15.759 15.674 15.721=15.7727
w/ : 14.974 14.657 13.969 14.985 14.728 15.665 15.191 14.995 14.946 14.895=14.9005
performance improvement w/ patch: 5.53%

g=4
Running in threaded mode with 4 groups using 40 file descriptors
Each sender will pass 100000 messages of 100 bytes
w/o: 20.014 21.025 21.119 21.235 19.767 20.971 20.962 20.914 21.090 21.090=20.8187
w/ : 20.331 20.608 20.338 20.445 20.456 20.146 20.693 20.797 21.381 20.452=20.5647
performance improvement w/ patch: 1.22%

After that, we run the same hackbench in both NUMA nodes(cpu0-cpu47):
g=1
w/o: 7.351 7.416 7.486 7.358 7.516 7.403 7.413 7.411 7.421 7.454=7.4229
w/ : 7.609 7.596 7.647 7.571 7.687 7.571 7.520 7.513 7.530 7.681=7.5925
performance improvement by patch: -2.2%

g=2
w/o: 9.046 9.190 9.053 8.950 9.101 8.930 9.143 8.928 8.905 9.034=9.028
w/ : 8.247 8.057 8.258 8.310 8.083 8.201 8.044 8.158 8.382 8.173=8.1913
performance improvement by patch: 9.3%

g=3
w/o: 11.664 11.767 11.277 11.619 12.557 12.760 11.664 12.165 12.235 11.849=11.9557
w/ : 9.387 9.461 9.650 9.613 9.591 9.454 9.496 9.716 9.327 9.722=9.5417
performance improvement by patch: 20.2%

g=4
w/o: 17.347 17.299 17.655 18.775 16.707 18.879 17.255 18.356 16.859 18.515=17.7647
w/ : 10.416 10.496 10.601 10.318 10.459 10.617 10.510 10.642 10.467 10.401=10.4927
performance improvement by patch: 40.9%

g=5
w/o: 27.805 26.633 24.138 28.086 24.405 27.922 30.043 28.458 31.073 25.819=27.4382
w/ : 13.817 13.976 14.166 13.688 14.132 14.095 14.003 13.997 13.954 13.907=13.9735
performance improvement by patch: 49.1%

It seems the patch can bring a huge increase on hackbench especially when
we bind hackbench to all of cpu0-cpu47, comparing to 5.53% while running
on single NUMA node(cpu0-cpu23)

Signed-off-by: Barry Song <[e-mail]>
---
 arch/arm64/Kconfig       |  7 +++++++
 arch/arm64/kernel/smp.c  | 17 +++++++++++++++++
 include/linux/topology.h |  7 +++++++
 kernel/sched/fair.c      | 35 +++++++++++++++++++++++++++++++++++
 4 files changed, 66 insertions(+)

Comments

Valentin Schneider Dec. 1, 2020, 4:04 p.m. UTC | #1

On 01/12/20 02:59, Barry Song wrote:
> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> index 1a68a05..ae8ec910 100644
> --- a/kernel/sched/fair.c
> +++ b/kernel/sched/fair.c
> @@ -6106,6 +6106,37 @@ static inline int select_idle_smt(struct task_struct *p, int target)
>  
>  #endif /* CONFIG_SCHED_SMT */
>  
> +#ifdef CONFIG_SCHED_CLUSTER
> +/*
> + * Scan the local CLUSTER mask for idle CPUs.
> + */
> +static int select_idle_cluster(struct task_struct *p, int target)
> +{
> +	int cpu;
> +
> +	/* right now, no hardware with both cluster and smt to run */
> +	if (sched_smt_active())
> +		return -1;
> +
> +	for_each_cpu_wrap(cpu, cpu_cluster_mask(target), target) {

Gating this behind this new config only leveraged by arm64 doesn't make it
very generic. Note that powerpc also has this newish "CACHE" level which
seems to overlap in function with your "CLUSTER" one (both are arch
specific, though).

I think what you are after here is an SD_SHARE_PKG_RESOURCES domain walk,
i.e. scan CPUs by increasing cache "distance". We already have it in some
form, as we scan SMT & LLC domains; AFAICT LLC always maps to MC, except
for said powerpc's CACHE thingie.

*If* we are to generally support more levels with SD_SHARE_PKG_RESOURCES,
we could say frob something into select_idle_cpu(). I'm thinking of
something like the incomplete, untested below: 

---
diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
index ae7ceba8fd4f..70692888db00 100644
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -6120,7 +6120,7 @@ static inline int select_idle_smt(struct task_struct *p, struct sched_domain *sd
 static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, int target)
 {
 	struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
-	struct sched_domain *this_sd;
+	struct sched_domain *this_sd, *child = NULL;
 	u64 avg_cost, avg_idle;
 	u64 time;
 	int this = smp_processor_id();
@@ -6150,14 +6150,22 @@ static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, int t
 
 	time = cpu_clock(this);
 
-	cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
+	do {
+		/* XXX: sd should start as SMT's parent */
+		cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
+		if (child)
+			cpumask_andnot(cpus, cpus, sched_domain_span(child));
+
+		for_each_cpu_wrap(cpu, cpus, target) {
+			if (!--nr)
+				return -1;
+			if (available_idle_cpu(cpu) || sched_idle_cpu(cpu))
+				break;
+		}
 
-	for_each_cpu_wrap(cpu, cpus, target) {
-		if (!--nr)
-			return -1;
-		if (available_idle_cpu(cpu) || sched_idle_cpu(cpu))
-			break;
-	}
+		child = sd;
+		sd = sd->parent;
+	} while (sd && sd->flags & SD_SHARE_PKG_RESOURCES);
 
 	time = cpu_clock(this) - time;
 	update_avg(&this_sd->avg_scan_cost, time);

Vincent Guittot Dec. 2, 2020, 8:27 a.m. UTC | #2

On Tue, 1 Dec 2020 at 04:04, Barry Song <[e-mail]> wrote:
>
> ARM64 server chip Kunpeng 920 has 6 clusters in each NUMA node, and each
> cluster has 4 cpus. All clusters share L3 cache data, but each cluster
> has local L3 tag. On the other hand, each clusters will share some
> internal system bus. This means cache coherence overhead inside one cluster
> is much less than the overhead across clusters.
>
> +-----------------------------------+                          +---------+
> |  +------+    +------+            +---------------------------+         |
> |  | CPU0 |    | cpu1 |             |    +-----------+         |         |
> |  +------+    +------+             |    |           |         |         |
> |                                   +----+    L3     |         |         |
> |  +------+    +------+   cluster   |    |    tag    |         |         |
> |  | CPU2 |    | CPU3 |             |    |           |         |         |
> |  +------+    +------+             |    +-----------+         |         |
> |                                   |                          |         |
> +-----------------------------------+                          |         |
> +-----------------------------------+                          |         |
> |  +------+    +------+             +--------------------------+         |
> |  |      |    |      |             |    +-----------+         |         |
> |  +------+    +------+             |    |           |         |         |
> |                                   |    |    L3     |         |         |
> |  +------+    +------+             +----+    tag    |         |         |
> |  |      |    |      |             |    |           |         |         |
> |  +------+    +------+             |    +-----------+         |         |
> |                                   |                          |         |
> +-----------------------------------+                          |   L3    |
>                                                                |   data  |
> +-----------------------------------+                          |         |
> |  +------+    +------+             |    +-----------+         |         |
> |  |      |    |      |             |    |           |         |         |
> |  +------+    +------+             +----+    L3     |         |         |
> |                                   |    |    tag    |         |         |
> |  +------+    +------+             |    |           |         |         |
> |  |      |    |      |            ++    +-----------+         |         |
> |  +------+    +------+            |---------------------------+         |
> +-----------------------------------|                          |         |
> +-----------------------------------|                          |         |
> |  +------+    +------+            +---------------------------+         |
> |  |      |    |      |             |    +-----------+         |         |
> |  +------+    +------+             |    |           |         |         |
> |                                   +----+    L3     |         |         |
> |  +------+    +------+             |    |    tag    |         |         |
> |  |      |    |      |             |    |           |         |         |
> |  +------+    +------+             |    +-----------+         |         |
> |                                   |                          |         |
> +-----------------------------------+                          |         |
> +-----------------------------------+                          |         |
> |  +------+    +------+             +--------------------------+         |
> |  |      |    |      |             |   +-----------+          |         |
> |  +------+    +------+             |   |           |          |         |
> |                                   |   |    L3     |          |         |
> |  +------+    +------+             +---+    tag    |          |         |
> |  |      |    |      |             |   |           |          |         |
> |  +------+    +------+             |   +-----------+          |         |
> |                                   |                          |         |
> +-----------------------------------+                          |         |
> +-----------------------------------+                         ++         |
> |  +------+    +------+             +--------------------------+         |
> |  |      |    |      |             |  +-----------+           |         |
> |  +------+    +------+             |  |           |           |         |
> |                                   |  |    L3     |           |         |
> |  +------+    +------+             +--+    tag    |           |         |
> |  |      |    |      |             |  |           |           |         |
> |  +------+    +------+             |  +-----------+           |         |
> |                                   |                          +---------+
> +-----------------------------------+
>
> This patch adds the sched_domain for clusters. On kunpeng 920, without
> this patch, domain0 of cpu0 would be MC for cpu0-cpu23 with
> min_interval=24, max_interval=48; with this patch, MC becomes domain1,
> a new domain0 "CL" including cpu0-cpu3 is added with min_interval=4 and
> max_interval=8.
> This will affect load balance. For example, without this patch, while cpu0
> becomes idle, it will pull a task from cpu1-cpu15. With this patch, cpu0
> will try to pull a task from cpu1-cpu3 first. This will have much less
> overhead of task migration.
>
> On the other hand, while doing WAKE_AFFINE, this patch will try to find
> a core in the target cluster before scanning the llc domain.
> This means it will proactively use a core which has better affinity with
> target core at first.

Which is at the opposite of what we are usually trying to do in the
fast wakeup path: trying to minimize resource sharing by finding an
idle core with all smt idle as an example

>
> Not much benchmark has been done yet. but here is a rough hackbench
> result.
> we run the below command with different -g parameter to increase system load
> by changing g from 1 to 4, for each one of 1-4, we run the benchmark ten times
> and record the data to get the average time:
>
> First, we run hackbench in only one NUMA node(cpu0-cpu23):
> $ numactl -N 0 hackbench -p -T -l 100000 -g $1

What is your ref tree ? v5.10-rcX or tip/sched/core ?

>
> g=1 (seen cpu utilization around 50% for each core)
> Running in threaded mode with 1 groups using 40 file descriptors
> Each sender will pass 100000 messages of 100 bytes
> w/o: 7.689 7.485 7.485 7.458 7.524 7.539 7.738 7.693 7.568 7.674=7.5853
> w/ : 7.516 7.941 7.374 7.963 7.881 7.910 7.420 7.556 7.695 7.441=7.6697
> performance improvement w/ patch: -1.01%
>
> g=2 (seen cpu utilization around 70% for each core)
> Running in threaded mode with 2 groups using 40 file descriptors
> Each sender will pass 100000 messages of 100 bytes
> w/o: 10.127 10.119 10.070 10.196 10.057 10.111 10.045 10.164 10.162 9.955=10.1006
> w/ : 9.694 9.654 9.612 9.649 9.686 9.734 9.607 9.842 9.690 9.710=9.6878
> performance improvement w/ patch: 4.08%
>
> g=3 (seen cpu utilization around 90% for each core)
> Running in threaded mode with 3 groups using 40 file descriptors
> Each sender will pass 100000 messages of 100 bytes
> w/o: 15.885 15.254 15.932 15.647 16.120 15.878 15.857 15.759 15.674 15.721=15.7727
> w/ : 14.974 14.657 13.969 14.985 14.728 15.665 15.191 14.995 14.946 14.895=14.9005
> performance improvement w/ patch: 5.53%
>
> g=4
> Running in threaded mode with 4 groups using 40 file descriptors
> Each sender will pass 100000 messages of 100 bytes
> w/o: 20.014 21.025 21.119 21.235 19.767 20.971 20.962 20.914 21.090 21.090=20.8187
> w/ : 20.331 20.608 20.338 20.445 20.456 20.146 20.693 20.797 21.381 20.452=20.5647
> performance improvement w/ patch: 1.22%
>
> After that, we run the same hackbench in both NUMA nodes(cpu0-cpu47):
> g=1
> w/o: 7.351 7.416 7.486 7.358 7.516 7.403 7.413 7.411 7.421 7.454=7.4229
> w/ : 7.609 7.596 7.647 7.571 7.687 7.571 7.520 7.513 7.530 7.681=7.5925
> performance improvement by patch: -2.2%
>
> g=2
> w/o: 9.046 9.190 9.053 8.950 9.101 8.930 9.143 8.928 8.905 9.034=9.028
> w/ : 8.247 8.057 8.258 8.310 8.083 8.201 8.044 8.158 8.382 8.173=8.1913
> performance improvement by patch: 9.3%
>
> g=3
> w/o: 11.664 11.767 11.277 11.619 12.557 12.760 11.664 12.165 12.235 11.849=11.9557
> w/ : 9.387 9.461 9.650 9.613 9.591 9.454 9.496 9.716 9.327 9.722=9.5417
> performance improvement by patch: 20.2%
>
> g=4
> w/o: 17.347 17.299 17.655 18.775 16.707 18.879 17.255 18.356 16.859 18.515=17.7647
> w/ : 10.416 10.496 10.601 10.318 10.459 10.617 10.510 10.642 10.467 10.401=10.4927
> performance improvement by patch: 40.9%
>
> g=5
> w/o: 27.805 26.633 24.138 28.086 24.405 27.922 30.043 28.458 31.073 25.819=27.4382
> w/ : 13.817 13.976 14.166 13.688 14.132 14.095 14.003 13.997 13.954 13.907=13.9735
> performance improvement by patch: 49.1%
>
> It seems the patch can bring a huge increase on hackbench especially when
> we bind hackbench to all of cpu0-cpu47, comparing to 5.53% while running
> on single NUMA node(cpu0-cpu23)

Interesting that this patch mainly impacts the numa case

>
> Signed-off-by: Barry Song <[e-mail]>
> ---
>  arch/arm64/Kconfig       |  7 +++++++
>  arch/arm64/kernel/smp.c  | 17 +++++++++++++++++
>  include/linux/topology.h |  7 +++++++
>  kernel/sched/fair.c      | 35 +++++++++++++++++++++++++++++++++++
>  4 files changed, 66 insertions(+)
>
> diff --git a/arch/arm64/Kconfig b/arch/arm64/Kconfig
> index 6d23283..3583c26 100644
> --- a/arch/arm64/Kconfig
> +++ b/arch/arm64/Kconfig
> @@ -938,6 +938,13 @@ config SCHED_MC
>           making when dealing with multi-core CPU chips at a cost of slightly
>           increased overhead in some places. If unsure say N here.
>
> +config SCHED_CLUSTER
> +       bool "Cluster scheduler support"
> +       help
> +         Cluster scheduler support improves the CPU scheduler's decision
> +         making when dealing with machines that have clusters(sharing internal
> +         bus or sharing LLC cache tag). If unsure say N here.
> +
>  config SCHED_SMT
>         bool "SMT scheduler support"
>         help
> diff --git a/arch/arm64/kernel/smp.c b/arch/arm64/kernel/smp.c
> index 355ee9e..5c8f026 100644
> --- a/arch/arm64/kernel/smp.c
> +++ b/arch/arm64/kernel/smp.c
> @@ -32,6 +32,7 @@
>  #include <linux/irq_work.h>
>  #include <linux/kexec.h>
>  #include <linux/kvm_host.h>
> +#include <linux/sched/topology.h>
>
>  #include <asm/alternative.h>
>  #include <asm/atomic.h>
> @@ -726,6 +727,20 @@ void __init smp_init_cpus(void)
>         }
>  }
>
> +static struct sched_domain_topology_level arm64_topology[] = {
> +#ifdef CONFIG_SCHED_SMT
> +        { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
> +#endif
> +#ifdef CONFIG_SCHED_CLUSTER
> +       { cpu_clustergroup_mask, cpu_core_flags, SD_INIT_NAME(CL) },
> +#endif
> +#ifdef CONFIG_SCHED_MC
> +        { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
> +#endif
> +       { cpu_cpu_mask, SD_INIT_NAME(DIE) },
> +        { NULL, },
> +};
> +
>  void __init smp_prepare_cpus(unsigned int max_cpus)
>  {
>         const struct cpu_operations *ops;
> @@ -735,6 +750,8 @@ void __init smp_prepare_cpus(unsigned int max_cpus)
>
>         init_cpu_topology();
>
> +       set_sched_topology(arm64_topology);
> +
>         this_cpu = smp_processor_id();
>         store_cpu_topology(this_cpu);
>         numa_store_cpu_info(this_cpu);
> diff --git a/include/linux/topology.h b/include/linux/topology.h
> index 5f66648..2c823c0 100644
> --- a/include/linux/topology.h
> +++ b/include/linux/topology.h
> @@ -211,6 +211,13 @@ static inline const struct cpumask *cpu_smt_mask(int cpu)
>  }
>  #endif
>
> +#ifdef CONFIG_SCHED_CLUSTER
> +static inline const struct cpumask *cpu_cluster_mask(int cpu)
> +{
> +       return topology_cluster_cpumask(cpu);
> +}
> +#endif
> +
>  static inline const struct cpumask *cpu_cpu_mask(int cpu)
>  {
>         return cpumask_of_node(cpu_to_node(cpu));
> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> index 1a68a05..ae8ec910 100644
> --- a/kernel/sched/fair.c
> +++ b/kernel/sched/fair.c
> @@ -6106,6 +6106,37 @@ static inline int select_idle_smt(struct task_struct *p, int target)
>
>  #endif /* CONFIG_SCHED_SMT */
>
> +#ifdef CONFIG_SCHED_CLUSTER
> +/*
> + * Scan the local CLUSTER mask for idle CPUs.
> + */
> +static int select_idle_cluster(struct task_struct *p, int target)
> +{
> +       int cpu;
> +
> +       /* right now, no hardware with both cluster and smt to run */
> +       if (sched_smt_active())

don't use smt static key but a dedicated one if needed

> +               return -1;
> +
> +       for_each_cpu_wrap(cpu, cpu_cluster_mask(target), target) {
> +               if (!cpumask_test_cpu(cpu, p->cpus_ptr))
> +                       continue;
> +               if (available_idle_cpu(cpu))
> +                       return cpu;
> +       }
> +
> +       return -1;
> +}
> +
> +#else /* CONFIG_SCHED_CLUSTER */
> +
> +static inline int select_idle_cluster(struct task_struct *p, int target)
> +{
> +       return -1;
> +}
> +
> +#endif /* CONFIG_SCHED_CLUSTER */
> +
>  /*
>   * Scan the LLC domain for idle CPUs; this is dynamically regulated by
>   * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
> @@ -6270,6 +6301,10 @@ static int select_idle_sibling(struct task_struct *p, int prev, int target)
>         if ((unsigned)i < nr_cpumask_bits)
>                 return i;
>
> +       i = select_idle_cluster(p, target);
> +       if ((unsigned)i < nr_cpumask_bits)
> +               return i;

This is yet another loop in the fast wake up path.

I'm curious to know which part of this patch really gives the perf improvement ?
-Is it the new sched domain level with a shorter interval that is then
used by Load balance to better spread task in the cluster and between
clusters ?
-Or this new loop in the wake up path which tries to keep threads in
the same cluster ? which is at the opposite of the rest of the
scheduler which tries to spread

Also could the sched_feat(SIS_PROP) impacts significantly your
topology because it  breaks before looking for all cores in the LLC ?
And this new loop extends the number of tested core ?

> +
>         i = select_idle_cpu(p, sd, target);
>         if ((unsigned)i < nr_cpumask_bits)
>                 return i;
> --
> 2.7.4
>

Song Bao Hua (Barry Song) Dec. 2, 2020, 9:20 a.m. UTC | #3

> -----Original Message-----
> From: Vincent Guittot [mailto:[e-mail]]
> Sent: Wednesday, December 2, 2020 9:27 PM
> To: Song Bao Hua (Barry Song) <[e-mail]>
> Cc: Valentin Schneider <[e-mail]>; Catalin Marinas
> <[e-mail]>; Will Deacon <[e-mail]>; Rafael J. Wysocki
> <[e-mail]>; Cc: Len Brown <[e-mail]>;
> [e-mail]; Jonathan Cameron <[e-mail]>;
> Ingo Molnar <[e-mail]>; Peter Zijlstra <[e-mail]>; Juri
> Lelli <[e-mail]>; Dietmar Eggemann <[e-mail]>;
> Steven Rostedt <[e-mail]>; Ben Segall <[e-mail]>; Mel
> Gorman <[e-mail]>; Mark Rutland <[e-mail]>; LAK
> <[e-mail]>; linux-kernel
> <[e-mail]>; ACPI Devel Maling List
> <[e-mail]>; Linuxarm <[e-mail]>; xuwei (O)
> <[e-mail]>; Zengtao (B) <[e-mail]>
> Subject: Re: [RFC PATCH v2 2/2] scheduler: add scheduler level for clusters
> 
> On Tue, 1 Dec 2020 at 04:04, Barry Song <[e-mail]> wrote:
> >
> > ARM64 server chip Kunpeng 920 has 6 clusters in each NUMA node, and each
> > cluster has 4 cpus. All clusters share L3 cache data, but each cluster
> > has local L3 tag. On the other hand, each clusters will share some
> > internal system bus. This means cache coherence overhead inside one cluster
> > is much less than the overhead across clusters.
> >
> > +-----------------------------------+                          +---------+
> > |  +------+    +------+            +---------------------------+         |
> > |  | CPU0 |    | cpu1 |             |    +-----------+         |         |
> > |  +------+    +------+             |    |           |         |         |
> > |                                   +----+    L3     |         |         |
> > |  +------+    +------+   cluster   |    |    tag    |         |         |
> > |  | CPU2 |    | CPU3 |             |    |           |         |         |
> > |  +------+    +------+             |    +-----------+         |         |
> > |                                   |                          |         |
> > +-----------------------------------+                          |         |
> > +-----------------------------------+                          |         |
> > |  +------+    +------+             +--------------------------+         |
> > |  |      |    |      |             |    +-----------+         |         |
> > |  +------+    +------+             |    |           |         |         |
> > |                                   |    |    L3     |         |         |
> > |  +------+    +------+             +----+    tag    |         |         |
> > |  |      |    |      |             |    |           |         |         |
> > |  +------+    +------+             |    +-----------+         |         |
> > |                                   |                          |         |
> > +-----------------------------------+                          |   L3    |
> >                                                                |   data  |
> > +-----------------------------------+                          |         |
> > |  +------+    +------+             |    +-----------+         |         |
> > |  |      |    |      |             |    |           |         |         |
> > |  +------+    +------+             +----+    L3     |         |         |
> > |                                   |    |    tag    |         |         |
> > |  +------+    +------+             |    |           |         |         |
> > |  |      |    |      |            ++    +-----------+         |         |
> > |  +------+    +------+            |---------------------------+         |
> > +-----------------------------------|                          |         |
> > +-----------------------------------|                          |         |
> > |  +------+    +------+            +---------------------------+         |
> > |  |      |    |      |             |    +-----------+         |         |
> > |  +------+    +------+             |    |           |         |         |
> > |                                   +----+    L3     |         |         |
> > |  +------+    +------+             |    |    tag    |         |         |
> > |  |      |    |      |             |    |           |         |         |
> > |  +------+    +------+             |    +-----------+         |         |
> > |                                   |                          |         |
> > +-----------------------------------+                          |         |
> > +-----------------------------------+                          |         |
> > |  +------+    +------+             +--------------------------+         |
> > |  |      |    |      |             |   +-----------+          |         |
> > |  +------+    +------+             |   |           |          |         |
> > |                                   |   |    L3     |          |         |
> > |  +------+    +------+             +---+    tag    |          |         |
> > |  |      |    |      |             |   |           |          |         |
> > |  +------+    +------+             |   +-----------+          |         |
> > |                                   |                          |         |
> > +-----------------------------------+                          |         |
> > +-----------------------------------+                         ++         |
> > |  +------+    +------+             +--------------------------+         |
> > |  |      |    |      |             |  +-----------+           |         |
> > |  +------+    +------+             |  |           |           |         |
> > |                                   |  |    L3     |           |         |
> > |  +------+    +------+             +--+    tag    |           |         |
> > |  |      |    |      |             |  |           |           |         |
> > |  +------+    +------+             |  +-----------+           |         |
> > |                                   |                          +---------+
> > +-----------------------------------+
> >
> > This patch adds the sched_domain for clusters. On kunpeng 920, without
> > this patch, domain0 of cpu0 would be MC for cpu0-cpu23 with
> > min_interval=24, max_interval=48; with this patch, MC becomes domain1,
> > a new domain0 "CL" including cpu0-cpu3 is added with min_interval=4 and
> > max_interval=8.
> > This will affect load balance. For example, without this patch, while cpu0
> > becomes idle, it will pull a task from cpu1-cpu15. With this patch, cpu0
> > will try to pull a task from cpu1-cpu3 first. This will have much less
> > overhead of task migration.
> >
> > On the other hand, while doing WAKE_AFFINE, this patch will try to find
> > a core in the target cluster before scanning the llc domain.
> > This means it will proactively use a core which has better affinity with
> > target core at first.
> 
> Which is at the opposite of what we are usually trying to do in the
> fast wakeup path: trying to minimize resource sharing by finding an
> idle core with all smt idle as an example

In wake_affine case, I guess we are actually want some kind of
resource sharing such as LLC to get waker and wakee get closer
to each other. find_idlest_cpu() is really opposite.

So the real question is that LLC is always the right choice of
idle sibling? 

In this case, 6 clusters are in same LLC, but hardware has different
behavior for inside single cluster and across multiple clusters.


> 
> >
> > Not much benchmark has been done yet. but here is a rough hackbench
> > result.
> > we run the below command with different -g parameter to increase system load
> > by changing g from 1 to 4, for each one of 1-4, we run the benchmark ten times
> > and record the data to get the average time:
> >
> > First, we run hackbench in only one NUMA node(cpu0-cpu23):
> > $ numactl -N 0 hackbench -p -T -l 100000 -g $1
> 
> What is your ref tree ? v5.10-rcX or tip/sched/core ?

Actually I was using 5.9 release.  That must be weird. 
But the reason is that disk driver is getting hang
in my hardware in 5.10-rcx.

> 
> >
> > g=1 (seen cpu utilization around 50% for each core)
> > Running in threaded mode with 1 groups using 40 file descriptors
> > Each sender will pass 100000 messages of 100 bytes
> > w/o: 7.689 7.485 7.485 7.458 7.524 7.539 7.738 7.693 7.568 7.674=7.5853
> > w/ : 7.516 7.941 7.374 7.963 7.881 7.910 7.420 7.556 7.695 7.441=7.6697
> > performance improvement w/ patch: -1.01%
> >
> > g=2 (seen cpu utilization around 70% for each core)
> > Running in threaded mode with 2 groups using 40 file descriptors
> > Each sender will pass 100000 messages of 100 bytes
> > w/o: 10.127 10.119 10.070 10.196 10.057 10.111 10.045 10.164 10.162
> 9.955=10.1006
> > w/ : 9.694 9.654 9.612 9.649 9.686 9.734 9.607 9.842 9.690 9.710=9.6878
> > performance improvement w/ patch: 4.08%
> >
> > g=3 (seen cpu utilization around 90% for each core)
> > Running in threaded mode with 3 groups using 40 file descriptors
> > Each sender will pass 100000 messages of 100 bytes
> > w/o: 15.885 15.254 15.932 15.647 16.120 15.878 15.857 15.759 15.674
> 15.721=15.7727
> > w/ : 14.974 14.657 13.969 14.985 14.728 15.665 15.191 14.995 14.946
> 14.895=14.9005
> > performance improvement w/ patch: 5.53%
> >
> > g=4
> > Running in threaded mode with 4 groups using 40 file descriptors
> > Each sender will pass 100000 messages of 100 bytes
> > w/o: 20.014 21.025 21.119 21.235 19.767 20.971 20.962 20.914 21.090
> 21.090=20.8187
> > w/ : 20.331 20.608 20.338 20.445 20.456 20.146 20.693 20.797 21.381
> 20.452=20.5647
> > performance improvement w/ patch: 1.22%
> >
> > After that, we run the same hackbench in both NUMA nodes(cpu0-cpu47):
> > g=1
> > w/o: 7.351 7.416 7.486 7.358 7.516 7.403 7.413 7.411 7.421 7.454=7.4229
> > w/ : 7.609 7.596 7.647 7.571 7.687 7.571 7.520 7.513 7.530 7.681=7.5925
> > performance improvement by patch: -2.2%
> >
> > g=2
> > w/o: 9.046 9.190 9.053 8.950 9.101 8.930 9.143 8.928 8.905 9.034=9.028
> > w/ : 8.247 8.057 8.258 8.310 8.083 8.201 8.044 8.158 8.382 8.173=8.1913
> > performance improvement by patch: 9.3%
> >
> > g=3
> > w/o: 11.664 11.767 11.277 11.619 12.557 12.760 11.664 12.165 12.235
> 11.849=11.9557
> > w/ : 9.387 9.461 9.650 9.613 9.591 9.454 9.496 9.716 9.327 9.722=9.5417
> > performance improvement by patch: 20.2%
> >
> > g=4
> > w/o: 17.347 17.299 17.655 18.775 16.707 18.879 17.255 18.356 16.859
> 18.515=17.7647
> > w/ : 10.416 10.496 10.601 10.318 10.459 10.617 10.510 10.642 10.467
> 10.401=10.4927
> > performance improvement by patch: 40.9%
> >
> > g=5
> > w/o: 27.805 26.633 24.138 28.086 24.405 27.922 30.043 28.458 31.073
> 25.819=27.4382
> > w/ : 13.817 13.976 14.166 13.688 14.132 14.095 14.003 13.997 13.954
> 13.907=13.9735
> > performance improvement by patch: 49.1%
> >
> > It seems the patch can bring a huge increase on hackbench especially when
> > we bind hackbench to all of cpu0-cpu47, comparing to 5.53% while running
> > on single NUMA node(cpu0-cpu23)
> 
> Interesting that this patch mainly impacts the numa case
> 
> >
> > Signed-off-by: Barry Song <[e-mail]>
> > ---
> >  arch/arm64/Kconfig       |  7 +++++++
> >  arch/arm64/kernel/smp.c  | 17 +++++++++++++++++
> >  include/linux/topology.h |  7 +++++++
> >  kernel/sched/fair.c      | 35 +++++++++++++++++++++++++++++++++++
> >  4 files changed, 66 insertions(+)
> >
> > diff --git a/arch/arm64/Kconfig b/arch/arm64/Kconfig
> > index 6d23283..3583c26 100644
> > --- a/arch/arm64/Kconfig
> > +++ b/arch/arm64/Kconfig
> > @@ -938,6 +938,13 @@ config SCHED_MC
> >           making when dealing with multi-core CPU chips at a cost of slightly
> >           increased overhead in some places. If unsure say N here.
> >
> > +config SCHED_CLUSTER
> > +       bool "Cluster scheduler support"
> > +       help
> > +         Cluster scheduler support improves the CPU scheduler's decision
> > +         making when dealing with machines that have clusters(sharing internal
> > +         bus or sharing LLC cache tag). If unsure say N here.
> > +
> >  config SCHED_SMT
> >         bool "SMT scheduler support"
> >         help
> > diff --git a/arch/arm64/kernel/smp.c b/arch/arm64/kernel/smp.c
> > index 355ee9e..5c8f026 100644
> > --- a/arch/arm64/kernel/smp.c
> > +++ b/arch/arm64/kernel/smp.c
> > @@ -32,6 +32,7 @@
> >  #include <linux/irq_work.h>
> >  #include <linux/kexec.h>
> >  #include <linux/kvm_host.h>
> > +#include <linux/sched/topology.h>
> >
> >  #include <asm/alternative.h>
> >  #include <asm/atomic.h>
> > @@ -726,6 +727,20 @@ void __init smp_init_cpus(void)
> >         }
> >  }
> >
> > +static struct sched_domain_topology_level arm64_topology[] = {
> > +#ifdef CONFIG_SCHED_SMT
> > +        { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
> > +#endif
> > +#ifdef CONFIG_SCHED_CLUSTER
> > +       { cpu_clustergroup_mask, cpu_core_flags, SD_INIT_NAME(CL) },
> > +#endif
> > +#ifdef CONFIG_SCHED_MC
> > +        { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
> > +#endif
> > +       { cpu_cpu_mask, SD_INIT_NAME(DIE) },
> > +        { NULL, },
> > +};
> > +
> >  void __init smp_prepare_cpus(unsigned int max_cpus)
> >  {
> >         const struct cpu_operations *ops;
> > @@ -735,6 +750,8 @@ void __init smp_prepare_cpus(unsigned int max_cpus)
> >
> >         init_cpu_topology();
> >
> > +       set_sched_topology(arm64_topology);
> > +
> >         this_cpu = smp_processor_id();
> >         store_cpu_topology(this_cpu);
> >         numa_store_cpu_info(this_cpu);
> > diff --git a/include/linux/topology.h b/include/linux/topology.h
> > index 5f66648..2c823c0 100644
> > --- a/include/linux/topology.h
> > +++ b/include/linux/topology.h
> > @@ -211,6 +211,13 @@ static inline const struct cpumask *cpu_smt_mask(int
> cpu)
> >  }
> >  #endif
> >
> > +#ifdef CONFIG_SCHED_CLUSTER
> > +static inline const struct cpumask *cpu_cluster_mask(int cpu)
> > +{
> > +       return topology_cluster_cpumask(cpu);
> > +}
> > +#endif
> > +
> >  static inline const struct cpumask *cpu_cpu_mask(int cpu)
> >  {
> >         return cpumask_of_node(cpu_to_node(cpu));
> > diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> > index 1a68a05..ae8ec910 100644
> > --- a/kernel/sched/fair.c
> > +++ b/kernel/sched/fair.c
> > @@ -6106,6 +6106,37 @@ static inline int select_idle_smt(struct task_struct
> *p, int target)
> >
> >  #endif /* CONFIG_SCHED_SMT */
> >
> > +#ifdef CONFIG_SCHED_CLUSTER
> > +/*
> > + * Scan the local CLUSTER mask for idle CPUs.
> > + */
> > +static int select_idle_cluster(struct task_struct *p, int target)
> > +{
> > +       int cpu;
> > +
> > +       /* right now, no hardware with both cluster and smt to run */
> > +       if (sched_smt_active())
> 
> don't use smt static key but a dedicated one if needed

Sure. 

> 
> > +               return -1;
> > +
> > +       for_each_cpu_wrap(cpu, cpu_cluster_mask(target), target) {
> > +               if (!cpumask_test_cpu(cpu, p->cpus_ptr))
> > +                       continue;
> > +               if (available_idle_cpu(cpu))
> > +                       return cpu;
> > +       }
> > +
> > +       return -1;
> > +}
> > +
> > +#else /* CONFIG_SCHED_CLUSTER */
> > +
> > +static inline int select_idle_cluster(struct task_struct *p, int target)
> > +{
> > +       return -1;
> > +}
> > +
> > +#endif /* CONFIG_SCHED_CLUSTER */
> > +
> >  /*
> >   * Scan the LLC domain for idle CPUs; this is dynamically regulated by
> >   * comparing the average scan cost (tracked in sd->avg_scan_cost) against
> the
> > @@ -6270,6 +6301,10 @@ static int select_idle_sibling(struct task_struct *p,
> int prev, int target)
> >         if ((unsigned)i < nr_cpumask_bits)
> >                 return i;
> >
> > +       i = select_idle_cluster(p, target);
> > +       if ((unsigned)i < nr_cpumask_bits)
> > +               return i;
> 
> This is yet another loop in the fast wake up path.
> 
> I'm curious to know which part of this patch really gives the perf improvement ?
> -Is it the new sched domain level with a shorter interval that is then
> used by Load balance to better spread task in the cluster and between
> clusters ?
> -Or this new loop in the wake up path which tries to keep threads in
> the same cluster ? which is at the opposite of the rest of the
> scheduler which tries to spread

If I don't scan cluster first for wake_affine, I almost don't see large
hackbench change by the new sche_domain.
For example:
g=4 in hackbench on cpu0-cpu47(two numa)
w/o patch: 17.7647 (average time in 10 times of hackbench)
w/ the full patch: 10.4927 
w/ patch but drop select_idle_cluster(): 15.0931

So I don't think the hackbench increase mainly comes from the shorter
interval of load balance of the new cluster domain.

What does really matter is select_idle_cluster() according to my tests.

> 
> Also could the sched_feat(SIS_PROP) impacts significantly your
> topology because it  breaks before looking for all cores in the LLC ?
> And this new loop extends the number of tested core ?

In this case, cluster must belong to LLC. Cluster is the child of LLC.

Maybe the code Valentin suggested in his reply is a good way to
keep the code align with the existing select_idle_cpu():

static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, int target)
 {
 	struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
-	struct sched_domain *this_sd;
+	struct sched_domain *this_sd, *child = NULL;
 	u64 avg_cost, avg_idle;
 	u64 time;
 	int this = smp_processor_id();
@@ -6150,14 +6150,22 @@ static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, int t
 
 	time = cpu_clock(this);
 
-	cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
+	do {
+		/* XXX: sd should start as SMT's parent */
+		cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
+		if (child)
+			cpumask_andnot(cpus, cpus, sched_domain_span(child));
+
+		for_each_cpu_wrap(cpu, cpus, target) {
+			if (!--nr)
+				return -1;
+			if (available_idle_cpu(cpu) || sched_idle_cpu(cpu))
+				break;
+		}
 
-	for_each_cpu_wrap(cpu, cpus, target) {
-		if (!--nr)
-			return -1;
-		if (available_idle_cpu(cpu) || sched_idle_cpu(cpu))
-			break;
-	}
+		child = sd;
+		sd = sd->parent;
+	} while (sd && sd->flags & SD_SHARE_PKG_RESOURCES);

> 
> > +
> >         i = select_idle_cpu(p, sd, target);
> >         if ((unsigned)i < nr_cpumask_bits)
> >                 return i;
> > --
> > 2.7.4
> >

Thanks
Barry

Vincent Guittot Dec. 2, 2020, 10:16 a.m. UTC | #4

On Wed, 2 Dec 2020 at 10:20, Song Bao Hua (Barry Song)
<[e-mail]> wrote:
>
>
>
> > -----Original Message-----
> > From: Vincent Guittot [mailto:[e-mail]]
> > Sent: Wednesday, December 2, 2020 9:27 PM
> > To: Song Bao Hua (Barry Song) <[e-mail]>
> > Cc: Valentin Schneider <[e-mail]>; Catalin Marinas
> > <[e-mail]>; Will Deacon <[e-mail]>; Rafael J. Wysocki
> > <[e-mail]>; Cc: Len Brown <[e-mail]>;
> > [e-mail]; Jonathan Cameron <[e-mail]>;
> > Ingo Molnar <[e-mail]>; Peter Zijlstra <[e-mail]>; Juri
> > Lelli <[e-mail]>; Dietmar Eggemann <[e-mail]>;
> > Steven Rostedt <[e-mail]>; Ben Segall <[e-mail]>; Mel
> > Gorman <[e-mail]>; Mark Rutland <[e-mail]>; LAK
> > <[e-mail]>; linux-kernel
> > <[e-mail]>; ACPI Devel Maling List
> > <[e-mail]>; Linuxarm <[e-mail]>; xuwei (O)
> > <[e-mail]>; Zengtao (B) <[e-mail]>
> > Subject: Re: [RFC PATCH v2 2/2] scheduler: add scheduler level for clusters
> >
> > On Tue, 1 Dec 2020 at 04:04, Barry Song <[e-mail]> wrote:
> > >
> > > ARM64 server chip Kunpeng 920 has 6 clusters in each NUMA node, and each
> > > cluster has 4 cpus. All clusters share L3 cache data, but each cluster
> > > has local L3 tag. On the other hand, each clusters will share some
> > > internal system bus. This means cache coherence overhead inside one cluster
> > > is much less than the overhead across clusters.
> > >
> > > +-----------------------------------+                          +---------+
> > > |  +------+    +------+            +---------------------------+         |
> > > |  | CPU0 |    | cpu1 |             |    +-----------+         |         |
> > > |  +------+    +------+             |    |           |         |         |
> > > |                                   +----+    L3     |         |         |
> > > |  +------+    +------+   cluster   |    |    tag    |         |         |
> > > |  | CPU2 |    | CPU3 |             |    |           |         |         |
> > > |  +------+    +------+             |    +-----------+         |         |
> > > |                                   |                          |         |
> > > +-----------------------------------+                          |         |
> > > +-----------------------------------+                          |         |
> > > |  +------+    +------+             +--------------------------+         |
> > > |  |      |    |      |             |    +-----------+         |         |
> > > |  +------+    +------+             |    |           |         |         |
> > > |                                   |    |    L3     |         |         |
> > > |  +------+    +------+             +----+    tag    |         |         |
> > > |  |      |    |      |             |    |           |         |         |
> > > |  +------+    +------+             |    +-----------+         |         |
> > > |                                   |                          |         |
> > > +-----------------------------------+                          |   L3    |
> > >                                                                |   data  |
> > > +-----------------------------------+                          |         |
> > > |  +------+    +------+             |    +-----------+         |         |
> > > |  |      |    |      |             |    |           |         |         |
> > > |  +------+    +------+             +----+    L3     |         |         |
> > > |                                   |    |    tag    |         |         |
> > > |  +------+    +------+             |    |           |         |         |
> > > |  |      |    |      |            ++    +-----------+         |         |
> > > |  +------+    +------+            |---------------------------+         |
> > > +-----------------------------------|                          |         |
> > > +-----------------------------------|                          |         |
> > > |  +------+    +------+            +---------------------------+         |
> > > |  |      |    |      |             |    +-----------+         |         |
> > > |  +------+    +------+             |    |           |         |         |
> > > |                                   +----+    L3     |         |         |
> > > |  +------+    +------+             |    |    tag    |         |         |
> > > |  |      |    |      |             |    |           |         |         |
> > > |  +------+    +------+             |    +-----------+         |         |
> > > |                                   |                          |         |
> > > +-----------------------------------+                          |         |
> > > +-----------------------------------+                          |         |
> > > |  +------+    +------+             +--------------------------+         |
> > > |  |      |    |      |             |   +-----------+          |         |
> > > |  +------+    +------+             |   |           |          |         |
> > > |                                   |   |    L3     |          |         |
> > > |  +------+    +------+             +---+    tag    |          |         |
> > > |  |      |    |      |             |   |           |          |         |
> > > |  +------+    +------+             |   +-----------+          |         |
> > > |                                   |                          |         |
> > > +-----------------------------------+                          |         |
> > > +-----------------------------------+                         ++         |
> > > |  +------+    +------+             +--------------------------+         |
> > > |  |      |    |      |             |  +-----------+           |         |
> > > |  +------+    +------+             |  |           |           |         |
> > > |                                   |  |    L3     |           |         |
> > > |  +------+    +------+             +--+    tag    |           |         |
> > > |  |      |    |      |             |  |           |           |         |
> > > |  +------+    +------+             |  +-----------+           |         |
> > > |                                   |                          +---------+
> > > +-----------------------------------+
> > >
> > > This patch adds the sched_domain for clusters. On kunpeng 920, without
> > > this patch, domain0 of cpu0 would be MC for cpu0-cpu23 with
> > > min_interval=24, max_interval=48; with this patch, MC becomes domain1,
> > > a new domain0 "CL" including cpu0-cpu3 is added with min_interval=4 and
> > > max_interval=8.
> > > This will affect load balance. For example, without this patch, while cpu0
> > > becomes idle, it will pull a task from cpu1-cpu15. With this patch, cpu0
> > > will try to pull a task from cpu1-cpu3 first. This will have much less
> > > overhead of task migration.
> > >
> > > On the other hand, while doing WAKE_AFFINE, this patch will try to find
> > > a core in the target cluster before scanning the llc domain.
> > > This means it will proactively use a core which has better affinity with
> > > target core at first.
> >
> > Which is at the opposite of what we are usually trying to do in the
> > fast wakeup path: trying to minimize resource sharing by finding an
> > idle core with all smt idle as an example
>
> In wake_affine case, I guess we are actually want some kind of
> resource sharing such as LLC to get waker and wakee get closer

In wake_affine, we don't want to move outside the LLC  but then in the
LLC we tries to minimize resource sharing like looking for a core
fully idle for SMT

> to each other. find_idlest_cpu() is really opposite.
>
> So the real question is that LLC is always the right choice of
> idle sibling?

That's the eternal question: spread or gather

>
> In this case, 6 clusters are in same LLC, but hardware has different
> behavior for inside single cluster and across multiple clusters.
>
>
> >
> > >
> > > Not much benchmark has been done yet. but here is a rough hackbench
> > > result.
> > > we run the below command with different -g parameter to increase system load
> > > by changing g from 1 to 4, for each one of 1-4, we run the benchmark ten times
> > > and record the data to get the average time:
> > >
> > > First, we run hackbench in only one NUMA node(cpu0-cpu23):
> > > $ numactl -N 0 hackbench -p -T -l 100000 -g $1
> >
> > What is your ref tree ? v5.10-rcX or tip/sched/core ?
>
> Actually I was using 5.9 release.  That must be weird.
> But the reason is that disk driver is getting hang
> in my hardware in 5.10-rcx.

In fact there are several changes in v5.10 and tip/sched/core that
could help your topology

>
> >
> > >
> > > g=1 (seen cpu utilization around 50% for each core)
> > > Running in threaded mode with 1 groups using 40 file descriptors
> > > Each sender will pass 100000 messages of 100 bytes
> > > w/o: 7.689 7.485 7.485 7.458 7.524 7.539 7.738 7.693 7.568 7.674=7.5853
> > > w/ : 7.516 7.941 7.374 7.963 7.881 7.910 7.420 7.556 7.695 7.441=7.6697
> > > performance improvement w/ patch: -1.01%
> > >
> > > g=2 (seen cpu utilization around 70% for each core)
> > > Running in threaded mode with 2 groups using 40 file descriptors
> > > Each sender will pass 100000 messages of 100 bytes
> > > w/o: 10.127 10.119 10.070 10.196 10.057 10.111 10.045 10.164 10.162
> > 9.955=10.1006
> > > w/ : 9.694 9.654 9.612 9.649 9.686 9.734 9.607 9.842 9.690 9.710=9.6878
> > > performance improvement w/ patch: 4.08%
> > >
> > > g=3 (seen cpu utilization around 90% for each core)
> > > Running in threaded mode with 3 groups using 40 file descriptors
> > > Each sender will pass 100000 messages of 100 bytes
> > > w/o: 15.885 15.254 15.932 15.647 16.120 15.878 15.857 15.759 15.674
> > 15.721=15.7727
> > > w/ : 14.974 14.657 13.969 14.985 14.728 15.665 15.191 14.995 14.946
> > 14.895=14.9005
> > > performance improvement w/ patch: 5.53%
> > >
> > > g=4
> > > Running in threaded mode with 4 groups using 40 file descriptors
> > > Each sender will pass 100000 messages of 100 bytes
> > > w/o: 20.014 21.025 21.119 21.235 19.767 20.971 20.962 20.914 21.090
> > 21.090=20.8187
> > > w/ : 20.331 20.608 20.338 20.445 20.456 20.146 20.693 20.797 21.381
> > 20.452=20.5647
> > > performance improvement w/ patch: 1.22%
> > >
> > > After that, we run the same hackbench in both NUMA nodes(cpu0-cpu47):
> > > g=1
> > > w/o: 7.351 7.416 7.486 7.358 7.516 7.403 7.413 7.411 7.421 7.454=7.4229
> > > w/ : 7.609 7.596 7.647 7.571 7.687 7.571 7.520 7.513 7.530 7.681=7.5925
> > > performance improvement by patch: -2.2%
> > >
> > > g=2
> > > w/o: 9.046 9.190 9.053 8.950 9.101 8.930 9.143 8.928 8.905 9.034=9.028
> > > w/ : 8.247 8.057 8.258 8.310 8.083 8.201 8.044 8.158 8.382 8.173=8.1913
> > > performance improvement by patch: 9.3%
> > >
> > > g=3
> > > w/o: 11.664 11.767 11.277 11.619 12.557 12.760 11.664 12.165 12.235
> > 11.849=11.9557
> > > w/ : 9.387 9.461 9.650 9.613 9.591 9.454 9.496 9.716 9.327 9.722=9.5417
> > > performance improvement by patch: 20.2%
> > >
> > > g=4
> > > w/o: 17.347 17.299 17.655 18.775 16.707 18.879 17.255 18.356 16.859
> > 18.515=17.7647
> > > w/ : 10.416 10.496 10.601 10.318 10.459 10.617 10.510 10.642 10.467
> > 10.401=10.4927
> > > performance improvement by patch: 40.9%
> > >
> > > g=5
> > > w/o: 27.805 26.633 24.138 28.086 24.405 27.922 30.043 28.458 31.073
> > 25.819=27.4382
> > > w/ : 13.817 13.976 14.166 13.688 14.132 14.095 14.003 13.997 13.954
> > 13.907=13.9735
> > > performance improvement by patch: 49.1%
> > >
> > > It seems the patch can bring a huge increase on hackbench especially when
> > > we bind hackbench to all of cpu0-cpu47, comparing to 5.53% while running
> > > on single NUMA node(cpu0-cpu23)
> >
> > Interesting that this patch mainly impacts the numa case
> >
> > >
> > > Signed-off-by: Barry Song <[e-mail]>
> > > ---
> > >  arch/arm64/Kconfig       |  7 +++++++
> > >  arch/arm64/kernel/smp.c  | 17 +++++++++++++++++
> > >  include/linux/topology.h |  7 +++++++
> > >  kernel/sched/fair.c      | 35 +++++++++++++++++++++++++++++++++++
> > >  4 files changed, 66 insertions(+)
> > >
> > > diff --git a/arch/arm64/Kconfig b/arch/arm64/Kconfig
> > > index 6d23283..3583c26 100644
> > > --- a/arch/arm64/Kconfig
> > > +++ b/arch/arm64/Kconfig
> > > @@ -938,6 +938,13 @@ config SCHED_MC
> > >           making when dealing with multi-core CPU chips at a cost of slightly
> > >           increased overhead in some places. If unsure say N here.
> > >
> > > +config SCHED_CLUSTER
> > > +       bool "Cluster scheduler support"
> > > +       help
> > > +         Cluster scheduler support improves the CPU scheduler's decision
> > > +         making when dealing with machines that have clusters(sharing internal
> > > +         bus or sharing LLC cache tag). If unsure say N here.
> > > +
> > >  config SCHED_SMT
> > >         bool "SMT scheduler support"
> > >         help
> > > diff --git a/arch/arm64/kernel/smp.c b/arch/arm64/kernel/smp.c
> > > index 355ee9e..5c8f026 100644
> > > --- a/arch/arm64/kernel/smp.c
> > > +++ b/arch/arm64/kernel/smp.c
> > > @@ -32,6 +32,7 @@
> > >  #include <linux/irq_work.h>
> > >  #include <linux/kexec.h>
> > >  #include <linux/kvm_host.h>
> > > +#include <linux/sched/topology.h>
> > >
> > >  #include <asm/alternative.h>
> > >  #include <asm/atomic.h>
> > > @@ -726,6 +727,20 @@ void __init smp_init_cpus(void)
> > >         }
> > >  }
> > >
> > > +static struct sched_domain_topology_level arm64_topology[] = {
> > > +#ifdef CONFIG_SCHED_SMT
> > > +        { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
> > > +#endif
> > > +#ifdef CONFIG_SCHED_CLUSTER
> > > +       { cpu_clustergroup_mask, cpu_core_flags, SD_INIT_NAME(CL) },
> > > +#endif
> > > +#ifdef CONFIG_SCHED_MC
> > > +        { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
> > > +#endif
> > > +       { cpu_cpu_mask, SD_INIT_NAME(DIE) },
> > > +        { NULL, },
> > > +};
> > > +
> > >  void __init smp_prepare_cpus(unsigned int max_cpus)
> > >  {
> > >         const struct cpu_operations *ops;
> > > @@ -735,6 +750,8 @@ void __init smp_prepare_cpus(unsigned int max_cpus)
> > >
> > >         init_cpu_topology();
> > >
> > > +       set_sched_topology(arm64_topology);
> > > +
> > >         this_cpu = smp_processor_id();
> > >         store_cpu_topology(this_cpu);
> > >         numa_store_cpu_info(this_cpu);
> > > diff --git a/include/linux/topology.h b/include/linux/topology.h
> > > index 5f66648..2c823c0 100644
> > > --- a/include/linux/topology.h
> > > +++ b/include/linux/topology.h
> > > @@ -211,6 +211,13 @@ static inline const struct cpumask *cpu_smt_mask(int
> > cpu)
> > >  }
> > >  #endif
> > >
> > > +#ifdef CONFIG_SCHED_CLUSTER
> > > +static inline const struct cpumask *cpu_cluster_mask(int cpu)
> > > +{
> > > +       return topology_cluster_cpumask(cpu);
> > > +}
> > > +#endif
> > > +
> > >  static inline const struct cpumask *cpu_cpu_mask(int cpu)
> > >  {
> > >         return cpumask_of_node(cpu_to_node(cpu));
> > > diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> > > index 1a68a05..ae8ec910 100644
> > > --- a/kernel/sched/fair.c
> > > +++ b/kernel/sched/fair.c
> > > @@ -6106,6 +6106,37 @@ static inline int select_idle_smt(struct task_struct
> > *p, int target)
> > >
> > >  #endif /* CONFIG_SCHED_SMT */
> > >
> > > +#ifdef CONFIG_SCHED_CLUSTER
> > > +/*
> > > + * Scan the local CLUSTER mask for idle CPUs.
> > > + */
> > > +static int select_idle_cluster(struct task_struct *p, int target)
> > > +{
> > > +       int cpu;
> > > +
> > > +       /* right now, no hardware with both cluster and smt to run */
> > > +       if (sched_smt_active())
> >
> > don't use smt static key but a dedicated one if needed
>
> Sure.
>
> >
> > > +               return -1;
> > > +
> > > +       for_each_cpu_wrap(cpu, cpu_cluster_mask(target), target) {
> > > +               if (!cpumask_test_cpu(cpu, p->cpus_ptr))
> > > +                       continue;
> > > +               if (available_idle_cpu(cpu))
> > > +                       return cpu;
> > > +       }
> > > +
> > > +       return -1;
> > > +}
> > > +
> > > +#else /* CONFIG_SCHED_CLUSTER */
> > > +
> > > +static inline int select_idle_cluster(struct task_struct *p, int target)
> > > +{
> > > +       return -1;
> > > +}
> > > +
> > > +#endif /* CONFIG_SCHED_CLUSTER */
> > > +
> > >  /*
> > >   * Scan the LLC domain for idle CPUs; this is dynamically regulated by
> > >   * comparing the average scan cost (tracked in sd->avg_scan_cost) against
> > the
> > > @@ -6270,6 +6301,10 @@ static int select_idle_sibling(struct task_struct *p,
> > int prev, int target)
> > >         if ((unsigned)i < nr_cpumask_bits)
> > >                 return i;
> > >
> > > +       i = select_idle_cluster(p, target);
> > > +       if ((unsigned)i < nr_cpumask_bits)
> > > +               return i;
> >
> > This is yet another loop in the fast wake up path.
> >
> > I'm curious to know which part of this patch really gives the perf improvement ?
> > -Is it the new sched domain level with a shorter interval that is then
> > used by Load balance to better spread task in the cluster and between
> > clusters ?
> > -Or this new loop in the wake up path which tries to keep threads in
> > the same cluster ? which is at the opposite of the rest of the
> > scheduler which tries to spread
>
> If I don't scan cluster first for wake_affine, I almost don't see large
> hackbench change by the new sche_domain.
> For example:
> g=4 in hackbench on cpu0-cpu47(two numa)
> w/o patch: 17.7647 (average time in 10 times of hackbench)
> w/ the full patch: 10.4927
> w/ patch but drop select_idle_cluster(): 15.0931

And for the case with one numa node ?

I'd like to understand why this patch impacts much the numa case but
not the  one numa node case.

>
> So I don't think the hackbench increase mainly comes from the shorter
> interval of load balance of the new cluster domain.
>
> What does really matter is select_idle_cluster() according to my tests.
>
> >
> > Also could the sched_feat(SIS_PROP) impacts significantly your
> > topology because it  breaks before looking for all cores in the LLC ?
> > And this new loop extends the number of tested core ?
>
> In this case, cluster must belong to LLC. Cluster is the child of LLC.

Yes . My point is: in select_idle_cpu, we don't always loop on all
CPUs especially when you have a large number of CPUs in the LLC.
Instead, the loop can break after testing only 4 CPUs in some cases of
shart idle time (like hackbench). I don't know how the CPUs are
numbered but I can easily imagine that select_idle_cluster doesn't
loop on the same CPUs as the few ones that are then tested in
select_idle_cpu when it doesn't test all CPUs of the LLC

>
> Maybe the code Valentin suggested in his reply is a good way to
> keep the code align with the existing select_idle_cpu():
>
> static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, int target)
>  {
>         struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
> -       struct sched_domain *this_sd;
> +       struct sched_domain *this_sd, *child = NULL;
>         u64 avg_cost, avg_idle;
>         u64 time;
>         int this = smp_processor_id();
> @@ -6150,14 +6150,22 @@ static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, int t
>
>         time = cpu_clock(this);
>
> -       cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
> +       do {
> +               /* XXX: sd should start as SMT's parent */
> +               cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
> +               if (child)
> +                       cpumask_andnot(cpus, cpus, sched_domain_span(child));
> +
> +               for_each_cpu_wrap(cpu, cpus, target) {
> +                       if (!--nr)
> +                               return -1;
> +                       if (available_idle_cpu(cpu) || sched_idle_cpu(cpu))
> +                               break;
> +               }
>
> -       for_each_cpu_wrap(cpu, cpus, target) {
> -               if (!--nr)
> -                       return -1;
> -               if (available_idle_cpu(cpu) || sched_idle_cpu(cpu))
> -                       break;
> -       }
> +               child = sd;
> +               sd = sd->parent;
> +       } while (sd && sd->flags & SD_SHARE_PKG_RESOURCES);
>
> >
> > > +
> > >         i = select_idle_cpu(p, sd, target);
> > >         if ((unsigned)i < nr_cpumask_bits)
> > >                 return i;
> > > --
> > > 2.7.4
> > >
>
> Thanks
> Barry
>

Song Bao Hua (Barry Song) Dec. 2, 2020, 10:45 a.m. UTC | #5

> -----Original Message-----
> From: Vincent Guittot [mailto:[e-mail]]
> Sent: Wednesday, December 2, 2020 11:17 PM
> To: Song Bao Hua (Barry Song) <[e-mail]>
> Cc: Valentin Schneider <[e-mail]>; Catalin Marinas
> <[e-mail]>; Will Deacon <[e-mail]>; Rafael J. Wysocki
> <[e-mail]>; Cc: Len Brown <[e-mail]>;
> [e-mail]; Jonathan Cameron <[e-mail]>;
> Ingo Molnar <[e-mail]>; Peter Zijlstra <[e-mail]>; Juri
> Lelli <[e-mail]>; Dietmar Eggemann <[e-mail]>;
> Steven Rostedt <[e-mail]>; Ben Segall <[e-mail]>; Mel
> Gorman <[e-mail]>; Mark Rutland <[e-mail]>; LAK
> <[e-mail]>; linux-kernel
> <[e-mail]>; ACPI Devel Maling List
> <[e-mail]>; Linuxarm <[e-mail]>; xuwei (O)
> <[e-mail]>; Zengtao (B) <[e-mail]>
> Subject: Re: [RFC PATCH v2 2/2] scheduler: add scheduler level for clusters
> 
> On Wed, 2 Dec 2020 at 10:20, Song Bao Hua (Barry Song)
> <[e-mail]> wrote:
> >
> >
> >
> > > -----Original Message-----
> > > From: Vincent Guittot [mailto:[e-mail]]
> > > Sent: Wednesday, December 2, 2020 9:27 PM
> > > To: Song Bao Hua (Barry Song) <[e-mail]>
> > > Cc: Valentin Schneider <[e-mail]>; Catalin Marinas
> > > <[e-mail]>; Will Deacon <[e-mail]>; Rafael J. Wysocki
> > > <[e-mail]>; Cc: Len Brown <[e-mail]>;
> > > [e-mail]; Jonathan Cameron
> <[e-mail]>;
> > > Ingo Molnar <[e-mail]>; Peter Zijlstra <[e-mail]>; Juri
> > > Lelli <[e-mail]>; Dietmar Eggemann
> <[e-mail]>;
> > > Steven Rostedt <[e-mail]>; Ben Segall <[e-mail]>; Mel
> > > Gorman <[e-mail]>; Mark Rutland <[e-mail]>; LAK
> > > <[e-mail]>; linux-kernel
> > > <[e-mail]>; ACPI Devel Maling List
> > > <[e-mail]>; Linuxarm <[e-mail]>; xuwei (O)
> > > <[e-mail]>; Zengtao (B) <[e-mail]>
> > > Subject: Re: [RFC PATCH v2 2/2] scheduler: add scheduler level for clusters
> > >
> > > On Tue, 1 Dec 2020 at 04:04, Barry Song <[e-mail]> wrote:
> > > >
> > > > ARM64 server chip Kunpeng 920 has 6 clusters in each NUMA node, and each
> > > > cluster has 4 cpus. All clusters share L3 cache data, but each cluster
> > > > has local L3 tag. On the other hand, each clusters will share some
> > > > internal system bus. This means cache coherence overhead inside one cluster
> > > > is much less than the overhead across clusters.
> > > >
> > > > +-----------------------------------+                          +---------+
> > > > |  +------+    +------+            +---------------------------+         |
> > > > |  | CPU0 |    | cpu1 |             |    +-----------+         |         |
> > > > |  +------+    +------+             |    |           |         |         |
> > > > |                                   +----+    L3     |         |         |
> > > > |  +------+    +------+   cluster   |    |    tag    |         |         |
> > > > |  | CPU2 |    | CPU3 |             |    |           |         |         |
> > > > |  +------+    +------+             |    +-----------+         |         |
> > > > |                                   |                          |         |
> > > > +-----------------------------------+                          |         |
> > > > +-----------------------------------+                          |         |
> > > > |  +------+    +------+             +--------------------------+         |
> > > > |  |      |    |      |             |    +-----------+         |         |
> > > > |  +------+    +------+             |    |           |         |         |
> > > > |                                   |    |    L3     |         |         |
> > > > |  +------+    +------+             +----+    tag    |         |         |
> > > > |  |      |    |      |             |    |           |         |         |
> > > > |  +------+    +------+             |    +-----------+         |         |
> > > > |                                   |                          |         |
> > > > +-----------------------------------+                          |   L3    |
> > > >                                                                |   data  |
> > > > +-----------------------------------+                          |         |
> > > > |  +------+    +------+             |    +-----------+         |         |
> > > > |  |      |    |      |             |    |           |         |         |
> > > > |  +------+    +------+             +----+    L3     |         |         |
> > > > |                                   |    |    tag    |         |         |
> > > > |  +------+    +------+             |    |           |         |         |
> > > > |  |      |    |      |            ++    +-----------+         |         |
> > > > |  +------+    +------+            |---------------------------+         |
> > > > +-----------------------------------|                          |         |
> > > > +-----------------------------------|                          |         |
> > > > |  +------+    +------+            +---------------------------+         |
> > > > |  |      |    |      |             |    +-----------+         |         |
> > > > |  +------+    +------+             |    |           |         |         |
> > > > |                                   +----+    L3     |         |         |
> > > > |  +------+    +------+             |    |    tag    |         |         |
> > > > |  |      |    |      |             |    |           |         |         |
> > > > |  +------+    +------+             |    +-----------+         |         |
> > > > |                                   |                          |         |
> > > > +-----------------------------------+                          |         |
> > > > +-----------------------------------+                          |         |
> > > > |  +------+    +------+             +--------------------------+         |
> > > > |  |      |    |      |             |   +-----------+          |         |
> > > > |  +------+    +------+             |   |           |          |         |
> > > > |                                   |   |    L3     |          |         |
> > > > |  +------+    +------+             +---+    tag    |          |         |
> > > > |  |      |    |      |             |   |           |          |         |
> > > > |  +------+    +------+             |   +-----------+          |         |
> > > > |                                   |                          |         |
> > > > +-----------------------------------+                          |         |
> > > > +-----------------------------------+                         ++         |
> > > > |  +------+    +------+             +--------------------------+         |
> > > > |  |      |    |      |             |  +-----------+           |         |
> > > > |  +------+    +------+             |  |           |           |         |
> > > > |                                   |  |    L3     |           |         |
> > > > |  +------+    +------+             +--+    tag    |           |         |
> > > > |  |      |    |      |             |  |           |           |         |
> > > > |  +------+    +------+             |  +-----------+           |         |
> > > > |                                   |                          +---------+
> > > > +-----------------------------------+
> > > >
> > > > This patch adds the sched_domain for clusters. On kunpeng 920, without
> > > > this patch, domain0 of cpu0 would be MC for cpu0-cpu23 with
> > > > min_interval=24, max_interval=48; with this patch, MC becomes domain1,
> > > > a new domain0 "CL" including cpu0-cpu3 is added with min_interval=4 and
> > > > max_interval=8.
> > > > This will affect load balance. For example, without this patch, while
> cpu0
> > > > becomes idle, it will pull a task from cpu1-cpu15. With this patch, cpu0
> > > > will try to pull a task from cpu1-cpu3 first. This will have much less
> > > > overhead of task migration.
> > > >
> > > > On the other hand, while doing WAKE_AFFINE, this patch will try to find
> > > > a core in the target cluster before scanning the llc domain.
> > > > This means it will proactively use a core which has better affinity with
> > > > target core at first.
> > >
> > > Which is at the opposite of what we are usually trying to do in the
> > > fast wakeup path: trying to minimize resource sharing by finding an
> > > idle core with all smt idle as an example
> >
> > In wake_affine case, I guess we are actually want some kind of
> > resource sharing such as LLC to get waker and wakee get closer
> 
> In wake_affine, we don't want to move outside the LLC  but then in the
> LLC we tries to minimize resource sharing like looking for a core
> fully idle for SMT
> 
> > to each other. find_idlest_cpu() is really opposite.
> >
> > So the real question is that LLC is always the right choice of
> > idle sibling?
> 
> That's the eternal question: spread or gather

Indeed.

> 
> >
> > In this case, 6 clusters are in same LLC, but hardware has different
> > behavior for inside single cluster and across multiple clusters.
> >
> >
> > >
> > > >
> > > > Not much benchmark has been done yet. but here is a rough hackbench
> > > > result.
> > > > we run the below command with different -g parameter to increase system
> load
> > > > by changing g from 1 to 4, for each one of 1-4, we run the benchmark ten
> times
> > > > and record the data to get the average time:
> > > >
> > > > First, we run hackbench in only one NUMA node(cpu0-cpu23):
> > > > $ numactl -N 0 hackbench -p -T -l 100000 -g $1
> > >
> > > What is your ref tree ? v5.10-rcX or tip/sched/core ?
> >
> > Actually I was using 5.9 release.  That must be weird.
> > But the reason is that disk driver is getting hang
> > in my hardware in 5.10-rcx.
> 
> In fact there are several changes in v5.10 and tip/sched/core that
> could help your topology

Will figure out some way to try.

> 
> >
> > >
> > > >
> > > > g=1 (seen cpu utilization around 50% for each core)
> > > > Running in threaded mode with 1 groups using 40 file descriptors
> > > > Each sender will pass 100000 messages of 100 bytes
> > > > w/o: 7.689 7.485 7.485 7.458 7.524 7.539 7.738 7.693 7.568 7.674=7.5853
> > > > w/ : 7.516 7.941 7.374 7.963 7.881 7.910 7.420 7.556 7.695 7.441=7.6697
> > > > performance improvement w/ patch: -1.01%
> > > >
> > > > g=2 (seen cpu utilization around 70% for each core)
> > > > Running in threaded mode with 2 groups using 40 file descriptors
> > > > Each sender will pass 100000 messages of 100 bytes
> > > > w/o: 10.127 10.119 10.070 10.196 10.057 10.111 10.045 10.164 10.162
> > > 9.955=10.1006
> > > > w/ : 9.694 9.654 9.612 9.649 9.686 9.734 9.607 9.842 9.690 9.710=9.6878
> > > > performance improvement w/ patch: 4.08%
> > > >
> > > > g=3 (seen cpu utilization around 90% for each core)
> > > > Running in threaded mode with 3 groups using 40 file descriptors
> > > > Each sender will pass 100000 messages of 100 bytes
> > > > w/o: 15.885 15.254 15.932 15.647 16.120 15.878 15.857 15.759 15.674
> > > 15.721=15.7727
> > > > w/ : 14.974 14.657 13.969 14.985 14.728 15.665 15.191 14.995 14.946
> > > 14.895=14.9005
> > > > performance improvement w/ patch: 5.53%
> > > >
> > > > g=4
> > > > Running in threaded mode with 4 groups using 40 file descriptors
> > > > Each sender will pass 100000 messages of 100 bytes
> > > > w/o: 20.014 21.025 21.119 21.235 19.767 20.971 20.962 20.914 21.090
> > > 21.090=20.8187
> > > > w/ : 20.331 20.608 20.338 20.445 20.456 20.146 20.693 20.797 21.381
> > > 20.452=20.5647
> > > > performance improvement w/ patch: 1.22%
> > > >
> > > > After that, we run the same hackbench in both NUMA nodes(cpu0-cpu47):
> > > > g=1
> > > > w/o: 7.351 7.416 7.486 7.358 7.516 7.403 7.413 7.411 7.421 7.454=7.4229
> > > > w/ : 7.609 7.596 7.647 7.571 7.687 7.571 7.520 7.513 7.530 7.681=7.5925
> > > > performance improvement by patch: -2.2%
> > > >
> > > > g=2
> > > > w/o: 9.046 9.190 9.053 8.950 9.101 8.930 9.143 8.928 8.905 9.034=9.028
> > > > w/ : 8.247 8.057 8.258 8.310 8.083 8.201 8.044 8.158 8.382 8.173=8.1913
> > > > performance improvement by patch: 9.3%
> > > >
> > > > g=3
> > > > w/o: 11.664 11.767 11.277 11.619 12.557 12.760 11.664 12.165 12.235
> > > 11.849=11.9557
> > > > w/ : 9.387 9.461 9.650 9.613 9.591 9.454 9.496 9.716 9.327 9.722=9.5417
> > > > performance improvement by patch: 20.2%
> > > >
> > > > g=4
> > > > w/o: 17.347 17.299 17.655 18.775 16.707 18.879 17.255 18.356 16.859
> > > 18.515=17.7647
> > > > w/ : 10.416 10.496 10.601 10.318 10.459 10.617 10.510 10.642 10.467
> > > 10.401=10.4927
> > > > performance improvement by patch: 40.9%
> > > >
> > > > g=5
> > > > w/o: 27.805 26.633 24.138 28.086 24.405 27.922 30.043 28.458 31.073
> > > 25.819=27.4382
> > > > w/ : 13.817 13.976 14.166 13.688 14.132 14.095 14.003 13.997 13.954
> > > 13.907=13.9735
> > > > performance improvement by patch: 49.1%
> > > >
> > > > It seems the patch can bring a huge increase on hackbench especially when
> > > > we bind hackbench to all of cpu0-cpu47, comparing to 5.53% while running
> > > > on single NUMA node(cpu0-cpu23)
> > >
> > > Interesting that this patch mainly impacts the numa case
> > >
> > > >
> > > > Signed-off-by: Barry Song <[e-mail]>
> > > > ---
> > > >  arch/arm64/Kconfig       |  7 +++++++
> > > >  arch/arm64/kernel/smp.c  | 17 +++++++++++++++++
> > > >  include/linux/topology.h |  7 +++++++
> > > >  kernel/sched/fair.c      | 35 +++++++++++++++++++++++++++++++++++
> > > >  4 files changed, 66 insertions(+)
> > > >
> > > > diff --git a/arch/arm64/Kconfig b/arch/arm64/Kconfig
> > > > index 6d23283..3583c26 100644
> > > > --- a/arch/arm64/Kconfig
> > > > +++ b/arch/arm64/Kconfig
> > > > @@ -938,6 +938,13 @@ config SCHED_MC
> > > >           making when dealing with multi-core CPU chips at a cost of slightly
> > > >           increased overhead in some places. If unsure say N here.
> > > >
> > > > +config SCHED_CLUSTER
> > > > +       bool "Cluster scheduler support"
> > > > +       help
> > > > +         Cluster scheduler support improves the CPU scheduler's decision
> > > > +         making when dealing with machines that have clusters(sharing
> internal
> > > > +         bus or sharing LLC cache tag). If unsure say N here.
> > > > +
> > > >  config SCHED_SMT
> > > >         bool "SMT scheduler support"
> > > >         help
> > > > diff --git a/arch/arm64/kernel/smp.c b/arch/arm64/kernel/smp.c
> > > > index 355ee9e..5c8f026 100644
> > > > --- a/arch/arm64/kernel/smp.c
> > > > +++ b/arch/arm64/kernel/smp.c
> > > > @@ -32,6 +32,7 @@
> > > >  #include <linux/irq_work.h>
> > > >  #include <linux/kexec.h>
> > > >  #include <linux/kvm_host.h>
> > > > +#include <linux/sched/topology.h>
> > > >
> > > >  #include <asm/alternative.h>
> > > >  #include <asm/atomic.h>
> > > > @@ -726,6 +727,20 @@ void __init smp_init_cpus(void)
> > > >         }
> > > >  }
> > > >
> > > > +static struct sched_domain_topology_level arm64_topology[] = {
> > > > +#ifdef CONFIG_SCHED_SMT
> > > > +        { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
> > > > +#endif
> > > > +#ifdef CONFIG_SCHED_CLUSTER
> > > > +       { cpu_clustergroup_mask, cpu_core_flags, SD_INIT_NAME(CL) },
> > > > +#endif
> > > > +#ifdef CONFIG_SCHED_MC
> > > > +        { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
> > > > +#endif
> > > > +       { cpu_cpu_mask, SD_INIT_NAME(DIE) },
> > > > +        { NULL, },
> > > > +};
> > > > +
> > > >  void __init smp_prepare_cpus(unsigned int max_cpus)
> > > >  {
> > > >         const struct cpu_operations *ops;
> > > > @@ -735,6 +750,8 @@ void __init smp_prepare_cpus(unsigned int max_cpus)
> > > >
> > > >         init_cpu_topology();
> > > >
> > > > +       set_sched_topology(arm64_topology);
> > > > +
> > > >         this_cpu = smp_processor_id();
> > > >         store_cpu_topology(this_cpu);
> > > >         numa_store_cpu_info(this_cpu);
> > > > diff --git a/include/linux/topology.h b/include/linux/topology.h
> > > > index 5f66648..2c823c0 100644
> > > > --- a/include/linux/topology.h
> > > > +++ b/include/linux/topology.h
> > > > @@ -211,6 +211,13 @@ static inline const struct cpumask *cpu_smt_mask(int
> > > cpu)
> > > >  }
> > > >  #endif
> > > >
> > > > +#ifdef CONFIG_SCHED_CLUSTER
> > > > +static inline const struct cpumask *cpu_cluster_mask(int cpu)
> > > > +{
> > > > +       return topology_cluster_cpumask(cpu);
> > > > +}
> > > > +#endif
> > > > +
> > > >  static inline const struct cpumask *cpu_cpu_mask(int cpu)
> > > >  {
> > > >         return cpumask_of_node(cpu_to_node(cpu));
> > > > diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> > > > index 1a68a05..ae8ec910 100644
> > > > --- a/kernel/sched/fair.c
> > > > +++ b/kernel/sched/fair.c
> > > > @@ -6106,6 +6106,37 @@ static inline int select_idle_smt(struct task_struct
> > > *p, int target)
> > > >
> > > >  #endif /* CONFIG_SCHED_SMT */
> > > >
> > > > +#ifdef CONFIG_SCHED_CLUSTER
> > > > +/*
> > > > + * Scan the local CLUSTER mask for idle CPUs.
> > > > + */
> > > > +static int select_idle_cluster(struct task_struct *p, int target)
> > > > +{
> > > > +       int cpu;
> > > > +
> > > > +       /* right now, no hardware with both cluster and smt to run */
> > > > +       if (sched_smt_active())
> > >
> > > don't use smt static key but a dedicated one if needed
> >
> > Sure.
> >
> > >
> > > > +               return -1;
> > > > +
> > > > +       for_each_cpu_wrap(cpu, cpu_cluster_mask(target), target) {
> > > > +               if (!cpumask_test_cpu(cpu, p->cpus_ptr))
> > > > +                       continue;
> > > > +               if (available_idle_cpu(cpu))
> > > > +                       return cpu;
> > > > +       }
> > > > +
> > > > +       return -1;
> > > > +}
> > > > +
> > > > +#else /* CONFIG_SCHED_CLUSTER */
> > > > +
> > > > +static inline int select_idle_cluster(struct task_struct *p, int target)
> > > > +{
> > > > +       return -1;
> > > > +}
> > > > +
> > > > +#endif /* CONFIG_SCHED_CLUSTER */
> > > > +
> > > >  /*
> > > >   * Scan the LLC domain for idle CPUs; this is dynamically regulated by
> > > >   * comparing the average scan cost (tracked in sd->avg_scan_cost) against
> > > the
> > > > @@ -6270,6 +6301,10 @@ static int select_idle_sibling(struct task_struct
> *p,
> > > int prev, int target)
> > > >         if ((unsigned)i < nr_cpumask_bits)
> > > >                 return i;
> > > >
> > > > +       i = select_idle_cluster(p, target);
> > > > +       if ((unsigned)i < nr_cpumask_bits)
> > > > +               return i;
> > >
> > > This is yet another loop in the fast wake up path.
> > >
> > > I'm curious to know which part of this patch really gives the perf improvement ?
> > > -Is it the new sched domain level with a shorter interval that is then
> > > used by Load balance to better spread task in the cluster and between
> > > clusters ?
> > > -Or this new loop in the wake up path which tries to keep threads in
> > > the same cluster ? which is at the opposite of the rest of the
> > > scheduler which tries to spread
> >
> > If I don't scan cluster first for wake_affine, I almost don't see large
> > hackbench change by the new sche_domain.
> > For example:
> > g=4 in hackbench on cpu0-cpu47(two numa)
> > w/o patch: 17.7647 (average time in 10 times of hackbench)
> > w/ the full patch: 10.4927
> > w/ patch but drop select_idle_cluster(): 15.0931
> 
> And for the case with one numa node ?

That would be very frustrating as it is getting worse:

g=1
Running in threaded mode with 1 groups using 40 file descriptors
Each sender will pass 100000 messages of 100 bytes
w/o: 7.689 7.485 7.485 7.458 7.524 7.539 7.738 7.693 7.568 7.674=7.5853
w/ : 7.516 7.941 7.374 7.963 7.881 7.910 7.420 7.556 7.695 7.441=7.6697
w/ but dropped select_idle_cluster:
     7.816 7.589 7.319 7.556 7.443 7.459 7.636 7.427 7.425 7.395=7.5065

g=2
Running in threaded mode with 2 groups using 40 file descriptors
Each sender will pass 100000 messages of 100 bytes
w/o: 10.127 10.119 10.070 10.196 10.057 10.111 10.045 10.164 10.162
9.955=10.1006
w/ : 9.694 9.654 9.612 9.649 9.686 9.734 9.607 9.842 9.690 9.710=9.6878
w/ but dropped select_idle_cluster:
     10.222 10.078 10.063 10.317 9.963 10.060 10.089 9.934 10.152 10.077=10.0955

g=3
Running in threaded mode with 3 groups using 40 file descriptors
Each sender will pass 100000 messages of 100 bytes
w/o: 15.885 15.254 15.932 15.647 16.120 15.878 15.857 15.759 15.674
15.721=15.7727
w/ : 14.974 14.657 13.969 14.985 14.728 15.665 15.191 14.995 14.946
14.895=14.9005
w/ but dropped select_idle_cluster(getting worse than w/o):
     16.892 16.962 17.248 17.392 17.336 17.705 17.113 17.633 17.477
17.378=17.3136

g=4
Running in threaded mode with 4 groups using 40 file descriptors
Each sender will pass 100000 messages of 100 bytes
w/o: 20.014 21.025 21.119 21.235 19.767 20.971 20.962 20.914 21.090
21.090=20.8187
w/ : 20.331 20.608 20.338 20.445 20.456 20.146 20.693 20.797 21.381
20.452=20.5647
w/ but dropped select_idle_cluster(getting worse than w/o):
     24.075 24.122 24.243 24.000 24.223 23.791 23.246 24.904 23.990
24.431=24.1025

> 
> I'd like to understand why this patch impacts much the numa case but
> not the  one numa node case.
> 
> >
> > So I don't think the hackbench increase mainly comes from the shorter
> > interval of load balance of the new cluster domain.
> >
> > What does really matter is select_idle_cluster() according to my tests.
> >
> > >
> > > Also could the sched_feat(SIS_PROP) impacts significantly your
> > > topology because it  breaks before looking for all cores in the LLC ?
> > > And this new loop extends the number of tested core ?
> >
> > In this case, cluster must belong to LLC. Cluster is the child of LLC.
> 
> Yes . My point is: in select_idle_cpu, we don't always loop on all
> CPUs especially when you have a large number of CPUs in the LLC.
> Instead, the loop can break after testing only 4 CPUs in some cases of
> shart idle time (like hackbench). I don't know how the CPUs are
> numbered but I can easily imagine that select_idle_cluster doesn't
> loop on the same CPUs as the few ones that are then tested in
> select_idle_cpu when it doesn't test all CPUs of the LLC
> 
> >
> > Maybe the code Valentin suggested in his reply is a good way to
> > keep the code align with the existing select_idle_cpu():
> >
> > static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd,
> int target)
> >  {
> >         struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
> > -       struct sched_domain *this_sd;
> > +       struct sched_domain *this_sd, *child = NULL;
> >         u64 avg_cost, avg_idle;
> >         u64 time;
> >         int this = smp_processor_id();
> > @@ -6150,14 +6150,22 @@ static int select_idle_cpu(struct task_struct *p,
> struct sched_domain *sd, int t
> >
> >         time = cpu_clock(this);
> >
> > -       cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
> > +       do {
> > +               /* XXX: sd should start as SMT's parent */
> > +               cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
> > +               if (child)
> > +                       cpumask_andnot(cpus, cpus, sched_domain_span(child));
> > +
> > +               for_each_cpu_wrap(cpu, cpus, target) {
> > +                       if (!--nr)
> > +                               return -1;
> > +                       if (available_idle_cpu(cpu) || sched_idle_cpu(cpu))
> > +                               break;
> > +               }
> >
> > -       for_each_cpu_wrap(cpu, cpus, target) {
> > -               if (!--nr)
> > -                       return -1;
> > -               if (available_idle_cpu(cpu) || sched_idle_cpu(cpu))
> > -                       break;
> > -       }
> > +               child = sd;
> > +               sd = sd->parent;
> > +       } while (sd && sd->flags & SD_SHARE_PKG_RESOURCES);
> >
> > >
> > > > +
> > > >         i = select_idle_cpu(p, sd, target);
> > > >         if ((unsigned)i < nr_cpumask_bits)
> > > >                 return i;
> > > > --
> > > > 2.7.4
> > > >
> >

Thanks
Barry

Song Bao Hua (Barry Song) Dec. 2, 2020, 10:48 a.m. UTC | #6

> -----Original Message-----
> From: Song Bao Hua (Barry Song)
> Sent: Wednesday, December 2, 2020 11:41 PM
> To: 'Vincent Guittot' <[e-mail]>
> Cc: Valentin Schneider <[e-mail]>; Catalin Marinas
> <[e-mail]>; Will Deacon <[e-mail]>; Rafael J. Wysocki
> <[e-mail]>; Cc: Len Brown <[e-mail]>;
> [e-mail]; Jonathan Cameron <[e-mail]>;
> Ingo Molnar <[e-mail]>; Peter Zijlstra <[e-mail]>; Juri
> Lelli <[e-mail]>; Dietmar Eggemann <[e-mail]>;
> Steven Rostedt <[e-mail]>; Ben Segall <[e-mail]>; Mel
> Gorman <[e-mail]>; Mark Rutland <[e-mail]>; LAK
> <[e-mail]>; linux-kernel
> <[e-mail]>; ACPI Devel Maling List
> <[e-mail]>; Linuxarm <[e-mail]>; xuwei (O)
> <[e-mail]>; Zengtao (B) <[e-mail]>
> Subject: RE: [RFC PATCH v2 2/2] scheduler: add scheduler level for clusters
> 
> 
> 
> > -----Original Message-----
> > From: Vincent Guittot [mailto:[e-mail]]
> > Sent: Wednesday, December 2, 2020 11:17 PM
> > To: Song Bao Hua (Barry Song) <[e-mail]>
> > Cc: Valentin Schneider <[e-mail]>; Catalin Marinas
> > <[e-mail]>; Will Deacon <[e-mail]>; Rafael J. Wysocki
> > <[e-mail]>; Cc: Len Brown <[e-mail]>;
> > [e-mail]; Jonathan Cameron <[e-mail]>;
> > Ingo Molnar <[e-mail]>; Peter Zijlstra <[e-mail]>; Juri
> > Lelli <[e-mail]>; Dietmar Eggemann <[e-mail]>;
> > Steven Rostedt <[e-mail]>; Ben Segall <[e-mail]>; Mel
> > Gorman <[e-mail]>; Mark Rutland <[e-mail]>; LAK
> > <[e-mail]>; linux-kernel
> > <[e-mail]>; ACPI Devel Maling List
> > <[e-mail]>; Linuxarm <[e-mail]>; xuwei (O)
> > <[e-mail]>; Zengtao (B) <[e-mail]>
> > Subject: Re: [RFC PATCH v2 2/2] scheduler: add scheduler level for clusters
> >
> > On Wed, 2 Dec 2020 at 10:20, Song Bao Hua (Barry Song)
> > <[e-mail]> wrote:
> > >
> > >
> > >
> > > > -----Original Message-----
> > > > From: Vincent Guittot [mailto:[e-mail]]
> > > > Sent: Wednesday, December 2, 2020 9:27 PM
> > > > To: Song Bao Hua (Barry Song) <[e-mail]>
> > > > Cc: Valentin Schneider <[e-mail]>; Catalin Marinas
> > > > <[e-mail]>; Will Deacon <[e-mail]>; Rafael J.
> Wysocki
> > > > <[e-mail]>; Cc: Len Brown <[e-mail]>;
> > > > [e-mail]; Jonathan Cameron
> > <[e-mail]>;
> > > > Ingo Molnar <[e-mail]>; Peter Zijlstra <[e-mail]>;
> Juri
> > > > Lelli <[e-mail]>; Dietmar Eggemann
> > <[e-mail]>;
> > > > Steven Rostedt <[e-mail]>; Ben Segall <[e-mail]>;
> Mel
> > > > Gorman <[e-mail]>; Mark Rutland <[e-mail]>; LAK
> > > > <[e-mail]>; linux-kernel
> > > > <[e-mail]>; ACPI Devel Maling List
> > > > <[e-mail]>; Linuxarm <[e-mail]>; xuwei (O)
> > > > <[e-mail]>; Zengtao (B) <[e-mail]>
> > > > Subject: Re: [RFC PATCH v2 2/2] scheduler: add scheduler level for clusters
> > > >
> > > > On Tue, 1 Dec 2020 at 04:04, Barry Song <[e-mail]> wrote:
> > > > >
> > > > > ARM64 server chip Kunpeng 920 has 6 clusters in each NUMA node, and
> each
> > > > > cluster has 4 cpus. All clusters share L3 cache data, but each cluster
> > > > > has local L3 tag. On the other hand, each clusters will share some
> > > > > internal system bus. This means cache coherence overhead inside one
> cluster
> > > > > is much less than the overhead across clusters.
> > > > >
> > > > > +-----------------------------------+                          +---------+
> > > > > |  +------+    +------+            +---------------------------+         |
> > > > > |  | CPU0 |    | cpu1 |             |    +-----------+         |         |
> > > > > |  +------+    +------+             |    |           |         |         |
> > > > > |                                   +----+    L3     |         |         |
> > > > > |  +------+    +------+   cluster   |    |    tag    |         |         |
> > > > > |  | CPU2 |    | CPU3 |             |    |           |         |         |
> > > > > |  +------+    +------+             |    +-----------+         |         |
> > > > > |                                   |                          |         |
> > > > > +-----------------------------------+                          |         |
> > > > > +-----------------------------------+                          |         |
> > > > > |  +------+    +------+             +--------------------------+         |
> > > > > |  |      |    |      |             |    +-----------+         |         |
> > > > > |  +------+    +------+             |    |           |         |         |
> > > > > |                                   |    |    L3     |         |         |
> > > > > |  +------+    +------+             +----+    tag    |         |         |
> > > > > |  |      |    |      |             |    |           |         |         |
> > > > > |  +------+    +------+             |    +-----------+         |         |
> > > > > |                                   |                          |         |
> > > > > +-----------------------------------+                          |   L3    |
> > > > >                                                                |   data  |
> > > > > +-----------------------------------+                          |         |
> > > > > |  +------+    +------+             |    +-----------+         |         |
> > > > > |  |      |    |      |             |    |           |         |         |
> > > > > |  +------+    +------+             +----+    L3     |         |         |
> > > > > |                                   |    |    tag    |         |         |
> > > > > |  +------+    +------+             |    |           |         |         |
> > > > > |  |      |    |      |            ++    +-----------+         |         |
> > > > > |  +------+    +------+            |---------------------------+         |
> > > > > +-----------------------------------|                          |         |
> > > > > +-----------------------------------|                          |         |
> > > > > |  +------+    +------+            +---------------------------+         |
> > > > > |  |      |    |      |             |    +-----------+         |         |
> > > > > |  +------+    +------+             |    |           |         |         |
> > > > > |                                   +----+    L3     |         |         |
> > > > > |  +------+    +------+             |    |    tag    |         |         |
> > > > > |  |      |    |      |             |    |           |         |         |
> > > > > |  +------+    +------+             |    +-----------+         |         |
> > > > > |                                   |                          |         |
> > > > > +-----------------------------------+                          |         |
> > > > > +-----------------------------------+                          |         |
> > > > > |  +------+    +------+             +--------------------------+         |
> > > > > |  |      |    |      |             |   +-----------+          |         |
> > > > > |  +------+    +------+             |   |           |          |         |
> > > > > |                                   |   |    L3     |          |         |
> > > > > |  +------+    +------+             +---+    tag    |          |         |
> > > > > |  |      |    |      |             |   |           |          |         |
> > > > > |  +------+    +------+             |   +-----------+          |         |
> > > > > |                                   |                          |         |
> > > > > +-----------------------------------+                          |         |
> > > > > +-----------------------------------+                         ++         |
> > > > > |  +------+    +------+             +--------------------------+         |
> > > > > |  |      |    |      |             |  +-----------+           |         |
> > > > > |  +------+    +------+             |  |           |           |         |
> > > > > |                                   |  |    L3     |           |         |
> > > > > |  +------+    +------+             +--+    tag    |           |         |
> > > > > |  |      |    |      |             |  |           |           |         |
> > > > > |  +------+    +------+             |  +-----------+           |         |
> > > > > |                                   |                          +---------+
> > > > > +-----------------------------------+
> > > > >
> > > > > This patch adds the sched_domain for clusters. On kunpeng 920, without
> > > > > this patch, domain0 of cpu0 would be MC for cpu0-cpu23 with
> > > > > min_interval=24, max_interval=48; with this patch, MC becomes domain1,
> > > > > a new domain0 "CL" including cpu0-cpu3 is added with min_interval=4
> and
> > > > > max_interval=8.
> > > > > This will affect load balance. For example, without this patch, while
> > cpu0
> > > > > becomes idle, it will pull a task from cpu1-cpu15. With this patch,
> cpu0
> > > > > will try to pull a task from cpu1-cpu3 first. This will have much less
> > > > > overhead of task migration.
> > > > >
> > > > > On the other hand, while doing WAKE_AFFINE, this patch will try to find
> > > > > a core in the target cluster before scanning the llc domain.
> > > > > This means it will proactively use a core which has better affinity
> with
> > > > > target core at first.
> > > >
> > > > Which is at the opposite of what we are usually trying to do in the
> > > > fast wakeup path: trying to minimize resource sharing by finding an
> > > > idle core with all smt idle as an example
> > >
> > > In wake_affine case, I guess we are actually want some kind of
> > > resource sharing such as LLC to get waker and wakee get closer
> >
> > In wake_affine, we don't want to move outside the LLC  but then in the
> > LLC we tries to minimize resource sharing like looking for a core
> > fully idle for SMT
> >
> > > to each other. find_idlest_cpu() is really opposite.
> > >
> > > So the real question is that LLC is always the right choice of
> > > idle sibling?
> >
> > That's the eternal question: spread or gather
> 
> Indeed.
> 
> >
> > >
> > > In this case, 6 clusters are in same LLC, but hardware has different
> > > behavior for inside single cluster and across multiple clusters.
> > >
> > >
> > > >
> > > > >
> > > > > Not much benchmark has been done yet. but here is a rough hackbench
> > > > > result.
> > > > > we run the below command with different -g parameter to increase system
> > load
> > > > > by changing g from 1 to 4, for each one of 1-4, we run the benchmark
> ten
> > times
> > > > > and record the data to get the average time:
> > > > >
> > > > > First, we run hackbench in only one NUMA node(cpu0-cpu23):
> > > > > $ numactl -N 0 hackbench -p -T -l 100000 -g $1
> > > >
> > > > What is your ref tree ? v5.10-rcX or tip/sched/core ?
> > >
> > > Actually I was using 5.9 release.  That must be weird.
> > > But the reason is that disk driver is getting hang
> > > in my hardware in 5.10-rcx.
> >
> > In fact there are several changes in v5.10 and tip/sched/core that
> > could help your topology
> 
> Will figure out some way to try.
> 
> >
> > >
> > > >
> > > > >
> > > > > g=1 (seen cpu utilization around 50% for each core)
> > > > > Running in threaded mode with 1 groups using 40 file descriptors
> > > > > Each sender will pass 100000 messages of 100 bytes
> > > > > w/o: 7.689 7.485 7.485 7.458 7.524 7.539 7.738 7.693 7.568 7.674=7.5853
> > > > > w/ : 7.516 7.941 7.374 7.963 7.881 7.910 7.420 7.556 7.695 7.441=7.6697
> > > > > performance improvement w/ patch: -1.01%
> > > > >
> > > > > g=2 (seen cpu utilization around 70% for each core)
> > > > > Running in threaded mode with 2 groups using 40 file descriptors
> > > > > Each sender will pass 100000 messages of 100 bytes
> > > > > w/o: 10.127 10.119 10.070 10.196 10.057 10.111 10.045 10.164 10.162
> > > > 9.955=10.1006
> > > > > w/ : 9.694 9.654 9.612 9.649 9.686 9.734 9.607 9.842 9.690 9.710=9.6878
> > > > > performance improvement w/ patch: 4.08%
> > > > >
> > > > > g=3 (seen cpu utilization around 90% for each core)
> > > > > Running in threaded mode with 3 groups using 40 file descriptors
> > > > > Each sender will pass 100000 messages of 100 bytes
> > > > > w/o: 15.885 15.254 15.932 15.647 16.120 15.878 15.857 15.759 15.674
> > > > 15.721=15.7727
> > > > > w/ : 14.974 14.657 13.969 14.985 14.728 15.665 15.191 14.995 14.946
> > > > 14.895=14.9005
> > > > > performance improvement w/ patch: 5.53%
> > > > >
> > > > > g=4
> > > > > Running in threaded mode with 4 groups using 40 file descriptors
> > > > > Each sender will pass 100000 messages of 100 bytes
> > > > > w/o: 20.014 21.025 21.119 21.235 19.767 20.971 20.962 20.914 21.090
> > > > 21.090=20.8187
> > > > > w/ : 20.331 20.608 20.338 20.445 20.456 20.146 20.693 20.797 21.381
> > > > 20.452=20.5647
> > > > > performance improvement w/ patch: 1.22%
> > > > >
> > > > > After that, we run the same hackbench in both NUMA nodes(cpu0-cpu47):
> > > > > g=1
> > > > > w/o: 7.351 7.416 7.486 7.358 7.516 7.403 7.413 7.411 7.421 7.454=7.4229
> > > > > w/ : 7.609 7.596 7.647 7.571 7.687 7.571 7.520 7.513 7.530 7.681=7.5925
> > > > > performance improvement by patch: -2.2%
> > > > >
> > > > > g=2
> > > > > w/o: 9.046 9.190 9.053 8.950 9.101 8.930 9.143 8.928 8.905 9.034=9.028
> > > > > w/ : 8.247 8.057 8.258 8.310 8.083 8.201 8.044 8.158 8.382 8.173=8.1913
> > > > > performance improvement by patch: 9.3%
> > > > >
> > > > > g=3
> > > > > w/o: 11.664 11.767 11.277 11.619 12.557 12.760 11.664 12.165 12.235
> > > > 11.849=11.9557
> > > > > w/ : 9.387 9.461 9.650 9.613 9.591 9.454 9.496 9.716 9.327 9.722=9.5417
> > > > > performance improvement by patch: 20.2%
> > > > >
> > > > > g=4
> > > > > w/o: 17.347 17.299 17.655 18.775 16.707 18.879 17.255 18.356 16.859
> > > > 18.515=17.7647
> > > > > w/ : 10.416 10.496 10.601 10.318 10.459 10.617 10.510 10.642 10.467
> > > > 10.401=10.4927
> > > > > performance improvement by patch: 40.9%
> > > > >
> > > > > g=5
> > > > > w/o: 27.805 26.633 24.138 28.086 24.405 27.922 30.043 28.458 31.073
> > > > 25.819=27.4382
> > > > > w/ : 13.817 13.976 14.166 13.688 14.132 14.095 14.003 13.997 13.954
> > > > 13.907=13.9735
> > > > > performance improvement by patch: 49.1%
> > > > >
> > > > > It seems the patch can bring a huge increase on hackbench especially
> when
> > > > > we bind hackbench to all of cpu0-cpu47, comparing to 5.53% while running
> > > > > on single NUMA node(cpu0-cpu23)
> > > >
> > > > Interesting that this patch mainly impacts the numa case
> > > >
> > > > >
> > > > > Signed-off-by: Barry Song <[e-mail]>
> > > > > ---
> > > > >  arch/arm64/Kconfig       |  7 +++++++
> > > > >  arch/arm64/kernel/smp.c  | 17 +++++++++++++++++
> > > > >  include/linux/topology.h |  7 +++++++
> > > > >  kernel/sched/fair.c      | 35 +++++++++++++++++++++++++++++++++++
> > > > >  4 files changed, 66 insertions(+)
> > > > >
> > > > > diff --git a/arch/arm64/Kconfig b/arch/arm64/Kconfig
> > > > > index 6d23283..3583c26 100644
> > > > > --- a/arch/arm64/Kconfig
> > > > > +++ b/arch/arm64/Kconfig
> > > > > @@ -938,6 +938,13 @@ config SCHED_MC
> > > > >           making when dealing with multi-core CPU chips at a cost of slightly
> > > > >           increased overhead in some places. If unsure say N here.
> > > > >
> > > > > +config SCHED_CLUSTER
> > > > > +       bool "Cluster scheduler support"
> > > > > +       help
> > > > > +         Cluster scheduler support improves the CPU scheduler's decision
> > > > > +         making when dealing with machines that have clusters(sharing
> > internal
> > > > > +         bus or sharing LLC cache tag). If unsure say N here.
> > > > > +
> > > > >  config SCHED_SMT
> > > > >         bool "SMT scheduler support"
> > > > >         help
> > > > > diff --git a/arch/arm64/kernel/smp.c b/arch/arm64/kernel/smp.c
> > > > > index 355ee9e..5c8f026 100644
> > > > > --- a/arch/arm64/kernel/smp.c
> > > > > +++ b/arch/arm64/kernel/smp.c
> > > > > @@ -32,6 +32,7 @@
> > > > >  #include <linux/irq_work.h>
> > > > >  #include <linux/kexec.h>
> > > > >  #include <linux/kvm_host.h>
> > > > > +#include <linux/sched/topology.h>
> > > > >
> > > > >  #include <asm/alternative.h>
> > > > >  #include <asm/atomic.h>
> > > > > @@ -726,6 +727,20 @@ void __init smp_init_cpus(void)
> > > > >         }
> > > > >  }
> > > > >
> > > > > +static struct sched_domain_topology_level arm64_topology[] = {
> > > > > +#ifdef CONFIG_SCHED_SMT
> > > > > +        { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
> > > > > +#endif
> > > > > +#ifdef CONFIG_SCHED_CLUSTER
> > > > > +       { cpu_clustergroup_mask, cpu_core_flags, SD_INIT_NAME(CL) },
> > > > > +#endif
> > > > > +#ifdef CONFIG_SCHED_MC
> > > > > +        { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
> > > > > +#endif
> > > > > +       { cpu_cpu_mask, SD_INIT_NAME(DIE) },
> > > > > +        { NULL, },
> > > > > +};
> > > > > +
> > > > >  void __init smp_prepare_cpus(unsigned int max_cpus)
> > > > >  {
> > > > >         const struct cpu_operations *ops;
> > > > > @@ -735,6 +750,8 @@ void __init smp_prepare_cpus(unsigned int max_cpus)
> > > > >
> > > > >         init_cpu_topology();
> > > > >
> > > > > +       set_sched_topology(arm64_topology);
> > > > > +
> > > > >         this_cpu = smp_processor_id();
> > > > >         store_cpu_topology(this_cpu);
> > > > >         numa_store_cpu_info(this_cpu);
> > > > > diff --git a/include/linux/topology.h b/include/linux/topology.h
> > > > > index 5f66648..2c823c0 100644
> > > > > --- a/include/linux/topology.h
> > > > > +++ b/include/linux/topology.h
> > > > > @@ -211,6 +211,13 @@ static inline const struct cpumask *cpu_smt_mask(int
> > > > cpu)
> > > > >  }
> > > > >  #endif
> > > > >
> > > > > +#ifdef CONFIG_SCHED_CLUSTER
> > > > > +static inline const struct cpumask *cpu_cluster_mask(int cpu)
> > > > > +{
> > > > > +       return topology_cluster_cpumask(cpu);
> > > > > +}
> > > > > +#endif
> > > > > +
> > > > >  static inline const struct cpumask *cpu_cpu_mask(int cpu)
> > > > >  {
> > > > >         return cpumask_of_node(cpu_to_node(cpu));
> > > > > diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> > > > > index 1a68a05..ae8ec910 100644
> > > > > --- a/kernel/sched/fair.c
> > > > > +++ b/kernel/sched/fair.c
> > > > > @@ -6106,6 +6106,37 @@ static inline int select_idle_smt(struct
> task_struct
> > > > *p, int target)
> > > > >
> > > > >  #endif /* CONFIG_SCHED_SMT */
> > > > >
> > > > > +#ifdef CONFIG_SCHED_CLUSTER
> > > > > +/*
> > > > > + * Scan the local CLUSTER mask for idle CPUs.
> > > > > + */
> > > > > +static int select_idle_cluster(struct task_struct *p, int target)
> > > > > +{
> > > > > +       int cpu;
> > > > > +
> > > > > +       /* right now, no hardware with both cluster and smt to run */
> > > > > +       if (sched_smt_active())
> > > >
> > > > don't use smt static key but a dedicated one if needed
> > >
> > > Sure.
> > >
> > > >
> > > > > +               return -1;
> > > > > +
> > > > > +       for_each_cpu_wrap(cpu, cpu_cluster_mask(target), target) {
> > > > > +               if (!cpumask_test_cpu(cpu, p->cpus_ptr))
> > > > > +                       continue;
> > > > > +               if (available_idle_cpu(cpu))
> > > > > +                       return cpu;
> > > > > +       }
> > > > > +
> > > > > +       return -1;
> > > > > +}
> > > > > +
> > > > > +#else /* CONFIG_SCHED_CLUSTER */
> > > > > +
> > > > > +static inline int select_idle_cluster(struct task_struct *p, int target)
> > > > > +{
> > > > > +       return -1;
> > > > > +}
> > > > > +
> > > > > +#endif /* CONFIG_SCHED_CLUSTER */
> > > > > +
> > > > >  /*
> > > > >   * Scan the LLC domain for idle CPUs; this is dynamically regulated
> by
> > > > >   * comparing the average scan cost (tracked in sd->avg_scan_cost) against
> > > > the
> > > > > @@ -6270,6 +6301,10 @@ static int select_idle_sibling(struct task_struct
> > *p,
> > > > int prev, int target)
> > > > >         if ((unsigned)i < nr_cpumask_bits)
> > > > >                 return i;
> > > > >
> > > > > +       i = select_idle_cluster(p, target);
> > > > > +       if ((unsigned)i < nr_cpumask_bits)
> > > > > +               return i;
> > > >
> > > > This is yet another loop in the fast wake up path.
> > > >
> > > > I'm curious to know which part of this patch really gives the perf
> improvement ?
> > > > -Is it the new sched domain level with a shorter interval that is then
> > > > used by Load balance to better spread task in the cluster and between
> > > > clusters ?
> > > > -Or this new loop in the wake up path which tries to keep threads in
> > > > the same cluster ? which is at the opposite of the rest of the
> > > > scheduler which tries to spread
> > >
> > > If I don't scan cluster first for wake_affine, I almost don't see large
> > > hackbench change by the new sche_domain.
> > > For example:
> > > g=4 in hackbench on cpu0-cpu47(two numa)
> > > w/o patch: 17.7647 (average time in 10 times of hackbench)
> > > w/ the full patch: 10.4927
> > > w/ patch but drop select_idle_cluster(): 15.0931
> >
> > And for the case with one numa node ?
> 
> That would be very frustrating as it is getting worse:
> 
> g=1
> Running in threaded mode with 1 groups using 40 file descriptors
> Each sender will pass 100000 messages of 100 bytes
> w/o: 7.689 7.485 7.485 7.458 7.524 7.539 7.738 7.693 7.568 7.674=7.5853
> w/ : 7.516 7.941 7.374 7.963 7.881 7.910 7.420 7.556 7.695 7.441=7.6697
> w/ but dropped select_idle_cluster:
>      7.816 7.589 7.319 7.556 7.443 7.459 7.636 7.427 7.425 7.395=7.5065
> 
> g=2
> Running in threaded mode with 2 groups using 40 file descriptors
> Each sender will pass 100000 messages of 100 bytes
> w/o: 10.127 10.119 10.070 10.196 10.057 10.111 10.045 10.164 10.162
> 9.955=10.1006
> w/ : 9.694 9.654 9.612 9.649 9.686 9.734 9.607 9.842 9.690 9.710=9.6878
> w/ but dropped select_idle_cluster:
>      10.222 10.078 10.063 10.317 9.963 10.060 10.089 9.934 10.152 10.077=10.0955
> 
> g=3
> Running in threaded mode with 3 groups using 40 file descriptors
> Each sender will pass 100000 messages of 100 bytes
> w/o: 15.885 15.254 15.932 15.647 16.120 15.878 15.857 15.759 15.674
> 15.721=15.7727
> w/ : 14.974 14.657 13.969 14.985 14.728 15.665 15.191 14.995 14.946
> 14.895=14.9005
> w/ but dropped select_idle_cluster(getting worse than w/o):
>      16.892 16.962 17.248 17.392 17.336 17.705 17.113 17.633 17.477
> 17.378=17.3136
> 
> g=4
> Running in threaded mode with 4 groups using 40 file descriptors
> Each sender will pass 100000 messages of 100 bytes
> w/o: 20.014 21.025 21.119 21.235 19.767 20.971 20.962 20.914 21.090
> 21.090=20.8187
> w/ : 20.331 20.608 20.338 20.445 20.456 20.146 20.693 20.797 21.381
> 20.452=20.5647
> w/ but dropped select_idle_cluster(getting worse than w/o):
>      24.075 24.122 24.243 24.000 24.223 23.791 23.246 24.904 23.990
> 24.431=24.1025

Sorry. Please ignore this. I added some printk here while testing
one numa. Will update you the data in another email.

Thanks
Barry

Song Bao Hua (Barry Song) Dec. 2, 2020, 8:58 p.m. UTC | #7

> 
> Sorry. Please ignore this. I added some printk here while testing
> one numa. Will update you the data in another email.

Re-tested in one NUMA node(cpu0-cpu23):

g=1
Running in threaded mode with 1 groups using 40 file descriptors
Each sender will pass 100000 messages of 100 bytes
w/o: 7.689 7.485 7.485 7.458 7.524 7.539 7.738 7.693 7.568 7.674=7.5853
w/ : 7.516 7.941 7.374 7.963 7.881 7.910 7.420 7.556 7.695 7.441=7.6697
w/ but dropped select_idle_cluster:
     7.752 7.739 7.739 7.571 7.545 7.685 7.407 7.580 7.605 7.487=7.611

g=2
Running in threaded mode with 2 groups using 40 file descriptors
Each sender will pass 100000 messages of 100 bytes
w/o: 10.127 10.119 10.070 10.196 10.057 10.111 10.045 10.164 10.162
9.955=10.1006
w/ : 9.694 9.654 9.612 9.649 9.686 9.734 9.607 9.842 9.690 9.710=9.6878
w/ but dropped select_idle_cluster:
     9.877 10.069 9.951 9.918 9.947 9.790 9.906 9.820 9.863 9.906=9.9047

g=3
Running in threaded mode with 3 groups using 40 file descriptors
Each sender will pass 100000 messages of 100 bytes
w/o: 15.885 15.254 15.932 15.647 16.120 15.878 15.857 15.759 15.674
15.721=15.7727
w/ : 14.974 14.657 13.969 14.985 14.728 15.665 15.191 14.995 14.946
14.895=14.9005
w/ but dropped select_idle_cluster:
     15.405 15.177 15.373 15.187 15.450 15.540 15.278 15.628 15.228 15.325=15.3591

g=4
Running in threaded mode with 4 groups using 40 file descriptors
Each sender will pass 100000 messages of 100 bytes
w/o: 20.014 21.025 21.119 21.235 19.767 20.971 20.962 20.914 21.090 21.090=20.8187
w/ : 20.331 20.608 20.338 20.445 20.456 20.146 20.693 20.797 21.381 20.452=20.5647
w/ but dropped select_idle_cluster:
     19.814 20.126 20.229 20.350 20.750 20.404 19.957 19.888 20.226 20.562=20.2306

Thanks
Barry

Vincent Guittot Dec. 3, 2020, 9:03 a.m. UTC | #8

On Wed, 2 Dec 2020 at 21:58, Song Bao Hua (Barry Song)
<[e-mail]> wrote:
>
> >
> > Sorry. Please ignore this. I added some printk here while testing
> > one numa. Will update you the data in another email.
>
> Re-tested in one NUMA node(cpu0-cpu23):
>
> g=1
> Running in threaded mode with 1 groups using 40 file descriptors
> Each sender will pass 100000 messages of 100 bytes
> w/o: 7.689 7.485 7.485 7.458 7.524 7.539 7.738 7.693 7.568 7.674=7.5853
> w/ : 7.516 7.941 7.374 7.963 7.881 7.910 7.420 7.556 7.695 7.441=7.6697
> w/ but dropped select_idle_cluster:
>      7.752 7.739 7.739 7.571 7.545 7.685 7.407 7.580 7.605 7.487=7.611
>
> g=2
> Running in threaded mode with 2 groups using 40 file descriptors
> Each sender will pass 100000 messages of 100 bytes
> w/o: 10.127 10.119 10.070 10.196 10.057 10.111 10.045 10.164 10.162
> 9.955=10.1006
> w/ : 9.694 9.654 9.612 9.649 9.686 9.734 9.607 9.842 9.690 9.710=9.6878
> w/ but dropped select_idle_cluster:
>      9.877 10.069 9.951 9.918 9.947 9.790 9.906 9.820 9.863 9.906=9.9047
>
> g=3
> Running in threaded mode with 3 groups using 40 file descriptors
> Each sender will pass 100000 messages of 100 bytes
> w/o: 15.885 15.254 15.932 15.647 16.120 15.878 15.857 15.759 15.674
> 15.721=15.7727
> w/ : 14.974 14.657 13.969 14.985 14.728 15.665 15.191 14.995 14.946
> 14.895=14.9005
> w/ but dropped select_idle_cluster:
>      15.405 15.177 15.373 15.187 15.450 15.540 15.278 15.628 15.228 15.325=15.3591
>
> g=4
> Running in threaded mode with 4 groups using 40 file descriptors
> Each sender will pass 100000 messages of 100 bytes
> w/o: 20.014 21.025 21.119 21.235 19.767 20.971 20.962 20.914 21.090 21.090=20.8187
> w/ : 20.331 20.608 20.338 20.445 20.456 20.146 20.693 20.797 21.381 20.452=20.5647
> w/ but dropped select_idle_cluster:
>      19.814 20.126 20.229 20.350 20.750 20.404 19.957 19.888 20.226 20.562=20.2306
>

I assume that you have run this on v5.9 as previous tests.
The results don't show any real benefit of select_idle_cluster()
inside a node whereas this is where we could expect most of the
benefit. We have to understand why we have such an impact on numa
tests only.

> Thanks
> Barry
>

Song Bao Hua (Barry Song) Dec. 3, 2020, 9:11 a.m. UTC | #9

> -----Original Message-----
> From: Vincent Guittot [mailto:[e-mail]]
> Sent: Thursday, December 3, 2020 10:04 PM
> To: Song Bao Hua (Barry Song) <[e-mail]>
> Cc: Valentin Schneider <[e-mail]>; Catalin Marinas
> <[e-mail]>; Will Deacon <[e-mail]>; Rafael J. Wysocki
> <[e-mail]>; Cc: Len Brown <[e-mail]>;
> [e-mail]; Jonathan Cameron <[e-mail]>;
> Ingo Molnar <[e-mail]>; Peter Zijlstra <[e-mail]>; Juri
> Lelli <[e-mail]>; Dietmar Eggemann <[e-mail]>;
> Steven Rostedt <[e-mail]>; Ben Segall <[e-mail]>; Mel
> Gorman <[e-mail]>; Mark Rutland <[e-mail]>; LAK
> <[e-mail]>; linux-kernel
> <[e-mail]>; ACPI Devel Maling List
> <[e-mail]>; Linuxarm <[e-mail]>; xuwei (O)
> <[e-mail]>; Zengtao (B) <[e-mail]>
> Subject: Re: [RFC PATCH v2 2/2] scheduler: add scheduler level for clusters
> 
> On Wed, 2 Dec 2020 at 21:58, Song Bao Hua (Barry Song)
> <[e-mail]> wrote:
> >
> > >
> > > Sorry. Please ignore this. I added some printk here while testing
> > > one numa. Will update you the data in another email.
> >
> > Re-tested in one NUMA node(cpu0-cpu23):
> >
> > g=1
> > Running in threaded mode with 1 groups using 40 file descriptors
> > Each sender will pass 100000 messages of 100 bytes
> > w/o: 7.689 7.485 7.485 7.458 7.524 7.539 7.738 7.693 7.568 7.674=7.5853
> > w/ : 7.516 7.941 7.374 7.963 7.881 7.910 7.420 7.556 7.695 7.441=7.6697
> > w/ but dropped select_idle_cluster:
> >      7.752 7.739 7.739 7.571 7.545 7.685 7.407 7.580 7.605 7.487=7.611
> >
> > g=2
> > Running in threaded mode with 2 groups using 40 file descriptors
> > Each sender will pass 100000 messages of 100 bytes
> > w/o: 10.127 10.119 10.070 10.196 10.057 10.111 10.045 10.164 10.162
> > 9.955=10.1006
> > w/ : 9.694 9.654 9.612 9.649 9.686 9.734 9.607 9.842 9.690 9.710=9.6878
> > w/ but dropped select_idle_cluster:
> >      9.877 10.069 9.951 9.918 9.947 9.790 9.906 9.820 9.863 9.906=9.9047
> >
> > g=3
> > Running in threaded mode with 3 groups using 40 file descriptors
> > Each sender will pass 100000 messages of 100 bytes
> > w/o: 15.885 15.254 15.932 15.647 16.120 15.878 15.857 15.759 15.674
> > 15.721=15.7727
> > w/ : 14.974 14.657 13.969 14.985 14.728 15.665 15.191 14.995 14.946
> > 14.895=14.9005
> > w/ but dropped select_idle_cluster:
> >      15.405 15.177 15.373 15.187 15.450 15.540 15.278 15.628 15.228
> 15.325=15.3591
> >
> > g=4
> > Running in threaded mode with 4 groups using 40 file descriptors
> > Each sender will pass 100000 messages of 100 bytes
> > w/o: 20.014 21.025 21.119 21.235 19.767 20.971 20.962 20.914 21.090
> 21.090=20.8187
> > w/ : 20.331 20.608 20.338 20.445 20.456 20.146 20.693 20.797 21.381
> 20.452=20.5647
> > w/ but dropped select_idle_cluster:
> >      19.814 20.126 20.229 20.350 20.750 20.404 19.957 19.888 20.226
> 20.562=20.2306
> >
> 
> I assume that you have run this on v5.9 as previous tests.

Yep

> The results don't show any real benefit of select_idle_cluster()
> inside a node whereas this is where we could expect most of the
> benefit. We have to understand why we have such an impact on numa
> tests only.

There is a 4-5.5% increase while g=2 and g=3.

Regarding the huge increase in NUMA case,  at the first beginning, I suspect
we have wrong llc domain. For example, if cpu0's llc domain span
cpu0-cpu47, then select_idle_cpu() is running in wrong range while
it should run in cpu0-cpu23.

But after printing the llc domain's span, I find it is completely right.
Cpu0's llc span: cpu0-cpu23
Cpu24's llc span: cpu24-cpu47

Maybe I need more trace data to figure out if select_idle_cpu() is running
correctly. For example, maybe I can figure out if it is always returning -1,
or it returns -1 very often?

Or do you have any idea?


> 
> > Thanks
> > Barry

Thanks
Barry

Peter Zijlstra Dec. 3, 2020, 9:28 a.m. UTC | #10

On Tue, Dec 01, 2020 at 04:04:04PM +0000, Valentin Schneider wrote:
> 
> Gating this behind this new config only leveraged by arm64 doesn't make it
> very generic. Note that powerpc also has this newish "CACHE" level which
> seems to overlap in function with your "CLUSTER" one (both are arch
> specific, though).
> 
> I think what you are after here is an SD_SHARE_PKG_RESOURCES domain walk,
> i.e. scan CPUs by increasing cache "distance". We already have it in some
> form, as we scan SMT & LLC domains; AFAICT LLC always maps to MC, except
> for said powerpc's CACHE thingie.

There's some intel chips with a smaller L2, but I don't think we ever
bothered.

There's also the extended topology stuff from Intel: SMT, Core, Module,
Tile, Die, of which we've only partially used Die I think.

Whatever we do, it might make sense to not all use different names.

Also, I think Mel said he was cooking something for
select_idle_balance().

Also, I've previously posted patches that fold all the iterations into
one, so it might make sense to revisit some of that if Mel also already
didn.t

Vincent Guittot Dec. 3, 2020, 9:39 a.m. UTC | #11

On Thu, 3 Dec 2020 at 10:11, Song Bao Hua (Barry Song)
<[e-mail]> wrote:
>
>
>
> > -----Original Message-----
> > From: Vincent Guittot [mailto:[e-mail]]
> > Sent: Thursday, December 3, 2020 10:04 PM
> > To: Song Bao Hua (Barry Song) <[e-mail]>
> > Cc: Valentin Schneider <[e-mail]>; Catalin Marinas
> > <[e-mail]>; Will Deacon <[e-mail]>; Rafael J. Wysocki
> > <[e-mail]>; Cc: Len Brown <[e-mail]>;
> > [e-mail]; Jonathan Cameron <[e-mail]>;
> > Ingo Molnar <[e-mail]>; Peter Zijlstra <[e-mail]>; Juri
> > Lelli <[e-mail]>; Dietmar Eggemann <[e-mail]>;
> > Steven Rostedt <[e-mail]>; Ben Segall <[e-mail]>; Mel
> > Gorman <[e-mail]>; Mark Rutland <[e-mail]>; LAK
> > <[e-mail]>; linux-kernel
> > <[e-mail]>; ACPI Devel Maling List
> > <[e-mail]>; Linuxarm <[e-mail]>; xuwei (O)
> > <[e-mail]>; Zengtao (B) <[e-mail]>
> > Subject: Re: [RFC PATCH v2 2/2] scheduler: add scheduler level for clusters
> >
> > On Wed, 2 Dec 2020 at 21:58, Song Bao Hua (Barry Song)
> > <[e-mail]> wrote:
> > >
> > > >
> > > > Sorry. Please ignore this. I added some printk here while testing
> > > > one numa. Will update you the data in another email.
> > >
> > > Re-tested in one NUMA node(cpu0-cpu23):
> > >
> > > g=1
> > > Running in threaded mode with 1 groups using 40 file descriptors
> > > Each sender will pass 100000 messages of 100 bytes
> > > w/o: 7.689 7.485 7.485 7.458 7.524 7.539 7.738 7.693 7.568 7.674=7.5853
> > > w/ : 7.516 7.941 7.374 7.963 7.881 7.910 7.420 7.556 7.695 7.441=7.6697
> > > w/ but dropped select_idle_cluster:
> > >      7.752 7.739 7.739 7.571 7.545 7.685 7.407 7.580 7.605 7.487=7.611
> > >
> > > g=2
> > > Running in threaded mode with 2 groups using 40 file descriptors
> > > Each sender will pass 100000 messages of 100 bytes
> > > w/o: 10.127 10.119 10.070 10.196 10.057 10.111 10.045 10.164 10.162
> > > 9.955=10.1006
> > > w/ : 9.694 9.654 9.612 9.649 9.686 9.734 9.607 9.842 9.690 9.710=9.6878
> > > w/ but dropped select_idle_cluster:
> > >      9.877 10.069 9.951 9.918 9.947 9.790 9.906 9.820 9.863 9.906=9.9047
> > >
> > > g=3
> > > Running in threaded mode with 3 groups using 40 file descriptors
> > > Each sender will pass 100000 messages of 100 bytes
> > > w/o: 15.885 15.254 15.932 15.647 16.120 15.878 15.857 15.759 15.674
> > > 15.721=15.7727
> > > w/ : 14.974 14.657 13.969 14.985 14.728 15.665 15.191 14.995 14.946
> > > 14.895=14.9005
> > > w/ but dropped select_idle_cluster:
> > >      15.405 15.177 15.373 15.187 15.450 15.540 15.278 15.628 15.228
> > 15.325=15.3591
> > >
> > > g=4
> > > Running in threaded mode with 4 groups using 40 file descriptors
> > > Each sender will pass 100000 messages of 100 bytes
> > > w/o: 20.014 21.025 21.119 21.235 19.767 20.971 20.962 20.914 21.090
> > 21.090=20.8187
> > > w/ : 20.331 20.608 20.338 20.445 20.456 20.146 20.693 20.797 21.381
> > 20.452=20.5647
> > > w/ but dropped select_idle_cluster:
> > >      19.814 20.126 20.229 20.350 20.750 20.404 19.957 19.888 20.226
> > 20.562=20.2306
> > >
> >
> > I assume that you have run this on v5.9 as previous tests.
>
> Yep
>
> > The results don't show any real benefit of select_idle_cluster()
> > inside a node whereas this is where we could expect most of the
> > benefit. We have to understand why we have such an impact on numa
> > tests only.
>
> There is a 4-5.5% increase while g=2 and g=3.

my point was with vs without select_idle_cluster() but still having a
cluster domain level
In this case, the diff is -0.8% for g=1 +2.2% for g=2, +3% for g=3 and
-1.7% for g=4

>
> Regarding the huge increase in NUMA case,  at the first beginning, I suspect
> we have wrong llc domain. For example, if cpu0's llc domain span
> cpu0-cpu47, then select_idle_cpu() is running in wrong range while
> it should run in cpu0-cpu23.
>
> But after printing the llc domain's span, I find it is completely right.
> Cpu0's llc span: cpu0-cpu23
> Cpu24's llc span: cpu24-cpu47

Have you checked that the cluster mask was also correct ?

>
> Maybe I need more trace data to figure out if select_idle_cpu() is running
> correctly. For example, maybe I can figure out if it is always returning -1,
> or it returns -1 very often?

yes, could be interesting to check how often select_idle_cpu return -1

>
> Or do you have any idea?

tracking migration across nod could help to understand too

Vincent
>
>
> >
> > > Thanks
> > > Barry
>
> Thanks
> Barry
>

Mel Gorman Dec. 3, 2020, 9:49 a.m. UTC | #12

On Thu, Dec 03, 2020 at 10:28:31AM +0100, Peter Zijlstra wrote:
> On Tue, Dec 01, 2020 at 04:04:04PM +0000, Valentin Schneider wrote:
> > 
> > Gating this behind this new config only leveraged by arm64 doesn't make it
> > very generic. Note that powerpc also has this newish "CACHE" level which
> > seems to overlap in function with your "CLUSTER" one (both are arch
> > specific, though).
> > 
> > I think what you are after here is an SD_SHARE_PKG_RESOURCES domain walk,
> > i.e. scan CPUs by increasing cache "distance". We already have it in some
> > form, as we scan SMT & LLC domains; AFAICT LLC always maps to MC, except
> > for said powerpc's CACHE thingie.
> 
> There's some intel chips with a smaller L2, but I don't think we ever
> bothered.
> 
> There's also the extended topology stuff from Intel: SMT, Core, Module,
> Tile, Die, of which we've only partially used Die I think.
> 
> Whatever we do, it might make sense to not all use different names.
> 
> Also, I think Mel said he was cooking something for
> select_idle_balance().
> 
> Also, I've previously posted patches that fold all the iterations into
> one, so it might make sense to revisit some of that if Mel also already
> didn.t

I didn't. The NUMA/lb reconcilation took most of my attention and right
now I'm looking at select_idle_sibling again in preparation for tracking
the idle cpumask in a sensible fashion. The main idea I had in mind was
special casing EPYC as it has multiple small L3 caches that may benefit
from select_idle_sibling looking slightly beyond the LLC as a search
domain but it has not even started yet.

Vincent Guittot Dec. 3, 2020, 9:54 a.m. UTC | #13

On Thu, 3 Dec 2020 at 10:39, Vincent Guittot <[e-mail]> wrote:
>
> On Thu, 3 Dec 2020 at 10:11, Song Bao Hua (Barry Song)
> <[e-mail]> wrote:
> >
> >
> >
> > > -----Original Message-----
> > > From: Vincent Guittot [mailto:[e-mail]]
> > > Sent: Thursday, December 3, 2020 10:04 PM
> > > To: Song Bao Hua (Barry Song) <[e-mail]>
> > > Cc: Valentin Schneider <[e-mail]>; Catalin Marinas
> > > <[e-mail]>; Will Deacon <[e-mail]>; Rafael J. Wysocki
> > > <[e-mail]>; Cc: Len Brown <[e-mail]>;
> > > [e-mail]; Jonathan Cameron <[e-mail]>;
> > > Ingo Molnar <[e-mail]>; Peter Zijlstra <[e-mail]>; Juri
> > > Lelli <[e-mail]>; Dietmar Eggemann <[e-mail]>;
> > > Steven Rostedt <[e-mail]>; Ben Segall <[e-mail]>; Mel
> > > Gorman <[e-mail]>; Mark Rutland <[e-mail]>; LAK
> > > <[e-mail]>; linux-kernel
> > > <[e-mail]>; ACPI Devel Maling List
> > > <[e-mail]>; Linuxarm <[e-mail]>; xuwei (O)
> > > <[e-mail]>; Zengtao (B) <[e-mail]>
> > > Subject: Re: [RFC PATCH v2 2/2] scheduler: add scheduler level for clusters
> > >
> > > On Wed, 2 Dec 2020 at 21:58, Song Bao Hua (Barry Song)
> > > <[e-mail]> wrote:
> > > >
> > > > >
> > > > > Sorry. Please ignore this. I added some printk here while testing
> > > > > one numa. Will update you the data in another email.
> > > >
> > > > Re-tested in one NUMA node(cpu0-cpu23):
> > > >
> > > > g=1
> > > > Running in threaded mode with 1 groups using 40 file descriptors
> > > > Each sender will pass 100000 messages of 100 bytes
> > > > w/o: 7.689 7.485 7.485 7.458 7.524 7.539 7.738 7.693 7.568 7.674=7.5853
> > > > w/ : 7.516 7.941 7.374 7.963 7.881 7.910 7.420 7.556 7.695 7.441=7.6697
> > > > w/ but dropped select_idle_cluster:
> > > >      7.752 7.739 7.739 7.571 7.545 7.685 7.407 7.580 7.605 7.487=7.611
> > > >
> > > > g=2
> > > > Running in threaded mode with 2 groups using 40 file descriptors
> > > > Each sender will pass 100000 messages of 100 bytes
> > > > w/o: 10.127 10.119 10.070 10.196 10.057 10.111 10.045 10.164 10.162
> > > > 9.955=10.1006
> > > > w/ : 9.694 9.654 9.612 9.649 9.686 9.734 9.607 9.842 9.690 9.710=9.6878
> > > > w/ but dropped select_idle_cluster:
> > > >      9.877 10.069 9.951 9.918 9.947 9.790 9.906 9.820 9.863 9.906=9.9047
> > > >
> > > > g=3
> > > > Running in threaded mode with 3 groups using 40 file descriptors
> > > > Each sender will pass 100000 messages of 100 bytes
> > > > w/o: 15.885 15.254 15.932 15.647 16.120 15.878 15.857 15.759 15.674
> > > > 15.721=15.7727
> > > > w/ : 14.974 14.657 13.969 14.985 14.728 15.665 15.191 14.995 14.946
> > > > 14.895=14.9005
> > > > w/ but dropped select_idle_cluster:
> > > >      15.405 15.177 15.373 15.187 15.450 15.540 15.278 15.628 15.228
> > > 15.325=15.3591
> > > >
> > > > g=4
> > > > Running in threaded mode with 4 groups using 40 file descriptors
> > > > Each sender will pass 100000 messages of 100 bytes
> > > > w/o: 20.014 21.025 21.119 21.235 19.767 20.971 20.962 20.914 21.090
> > > 21.090=20.8187
> > > > w/ : 20.331 20.608 20.338 20.445 20.456 20.146 20.693 20.797 21.381
> > > 20.452=20.5647
> > > > w/ but dropped select_idle_cluster:
> > > >      19.814 20.126 20.229 20.350 20.750 20.404 19.957 19.888 20.226
> > > 20.562=20.2306
> > > >
> > >
> > > I assume that you have run this on v5.9 as previous tests.
> >
> > Yep
> >
> > > The results don't show any real benefit of select_idle_cluster()
> > > inside a node whereas this is where we could expect most of the
> > > benefit. We have to understand why we have such an impact on numa
> > > tests only.
> >
> > There is a 4-5.5% increase while g=2 and g=3.
>
> my point was with vs without select_idle_cluster() but still having a
> cluster domain level
> In this case, the diff is -0.8% for g=1 +2.2% for g=2, +3% for g=3 and
> -1.7% for g=4
>
> >
> > Regarding the huge increase in NUMA case,  at the first beginning, I suspect
> > we have wrong llc domain. For example, if cpu0's llc domain span
> > cpu0-cpu47, then select_idle_cpu() is running in wrong range while
> > it should run in cpu0-cpu23.
> >
> > But after printing the llc domain's span, I find it is completely right.
> > Cpu0's llc span: cpu0-cpu23
> > Cpu24's llc span: cpu24-cpu47
>
> Have you checked that the cluster mask was also correct ?
>
> >
> > Maybe I need more trace data to figure out if select_idle_cpu() is running
> > correctly. For example, maybe I can figure out if it is always returning -1,
> > or it returns -1 very often?
>
> yes, could be interesting to check how often select_idle_cpu return -1
>
> >
> > Or do you have any idea?
>
> tracking migration across nod could help to understand too

Also the v6 of https://lkml.org/lkml/2020/11/26/187 might also help you

>
> Vincent
> >
> >
> > >
> > > > Thanks
> > > > Barry
> >
> > Thanks
> > Barry
> >

Song Bao Hua (Barry Song) Dec. 3, 2020, 9:57 a.m. UTC | #14

> -----Original Message-----
> From: Peter Zijlstra [mailto:[e-mail]]
> Sent: Thursday, December 3, 2020 10:29 PM
> To: Valentin Schneider <[e-mail]>
> Cc: Song Bao Hua (Barry Song) <[e-mail]>;
> [e-mail]; [e-mail]; [e-mail]; [e-mail];
> [e-mail]; Jonathan Cameron <[e-mail]>;
> [e-mail]; [e-mail]; [e-mail];
> [e-mail]; [e-mail]; [e-mail];
> [e-mail]; [e-mail]; [e-mail];
> [e-mail]; [e-mail]; Linuxarm
> <[e-mail]>; xuwei (O) <[e-mail]>; Zengtao (B)
> <[e-mail]>
> Subject: Re: [RFC PATCH v2 2/2] scheduler: add scheduler level for clusters
> 
> On Tue, Dec 01, 2020 at 04:04:04PM +0000, Valentin Schneider wrote:
> >
> > Gating this behind this new config only leveraged by arm64 doesn't make it
> > very generic. Note that powerpc also has this newish "CACHE" level which
> > seems to overlap in function with your "CLUSTER" one (both are arch
> > specific, though).
> >
> > I think what you are after here is an SD_SHARE_PKG_RESOURCES domain walk,
> > i.e. scan CPUs by increasing cache "distance". We already have it in some
> > form, as we scan SMT & LLC domains; AFAICT LLC always maps to MC, except
> > for said powerpc's CACHE thingie.
> 
> There's some intel chips with a smaller L2, but I don't think we ever
> bothered.
> 
> There's also the extended topology stuff from Intel: SMT, Core, Module,
> Tile, Die, of which we've only partially used Die I think.
> 
> Whatever we do, it might make sense to not all use different names.

Yep. Valentin was actually recommending the same SD_SHARE_PKG_RESOURCES sd flags
by ignoring the actual names of the hardware.
But the question is where we should start, in case we have 3 domains under llc,
maybe it is not good to scan from the first level domain as it is gathering
too much.

> 
> Also, I think Mel said he was cooking something for
> select_idle_balance().
> 
> Also, I've previously posted patches that fold all the iterations into
> one, so it might make sense to revisit some of that if Mel also already
> didn.t

Would you point out the link of your previous patches?

Thanks
Barry

Peter Zijlstra Dec. 3, 2020, 10:07 a.m. UTC | #15

On Thu, Dec 03, 2020 at 09:57:09AM +0000, Song Bao Hua (Barry Song) wrote:

> Would you point out the link of your previous patches?

https://lkml.kernel.org/r/[e-mail]

Song Bao Hua (Barry Song) Dec. 7, 2020, 9:59 a.m. UTC | #16

> -----Original Message-----
> From: Vincent Guittot [mailto:[e-mail]]
> Sent: Thursday, December 3, 2020 10:39 PM
> To: Song Bao Hua (Barry Song) <[e-mail]>
> Cc: Valentin Schneider <[e-mail]>; Catalin Marinas
> <[e-mail]>; Will Deacon <[e-mail]>; Rafael J. Wysocki
> <[e-mail]>; Cc: Len Brown <[e-mail]>;
> [e-mail]; Jonathan Cameron <[e-mail]>;
> Ingo Molnar <[e-mail]>; Peter Zijlstra <[e-mail]>; Juri
> Lelli <[e-mail]>; Dietmar Eggemann <[e-mail]>;
> Steven Rostedt <[e-mail]>; Ben Segall <[e-mail]>; Mel
> Gorman <[e-mail]>; Mark Rutland <[e-mail]>; LAK
> <[e-mail]>; linux-kernel
> <[e-mail]>; ACPI Devel Maling List
> <[e-mail]>; Linuxarm <[e-mail]>; xuwei (O)
> <[e-mail]>; Zengtao (B) <[e-mail]>
> Subject: Re: [RFC PATCH v2 2/2] scheduler: add scheduler level for clusters
> 
> On Thu, 3 Dec 2020 at 10:11, Song Bao Hua (Barry Song)
> <[e-mail]> wrote:
> >
> >
> >
> > > -----Original Message-----
> > > From: Vincent Guittot [mailto:[e-mail]]
> > > Sent: Thursday, December 3, 2020 10:04 PM
> > > To: Song Bao Hua (Barry Song) <[e-mail]>
> > > Cc: Valentin Schneider <[e-mail]>; Catalin Marinas
> > > <[e-mail]>; Will Deacon <[e-mail]>; Rafael J. Wysocki
> > > <[e-mail]>; Cc: Len Brown <[e-mail]>;
> > > [e-mail]; Jonathan Cameron
> <[e-mail]>;
> > > Ingo Molnar <[e-mail]>; Peter Zijlstra <[e-mail]>; Juri
> > > Lelli <[e-mail]>; Dietmar Eggemann
> <[e-mail]>;
> > > Steven Rostedt <[e-mail]>; Ben Segall <[e-mail]>; Mel
> > > Gorman <[e-mail]>; Mark Rutland <[e-mail]>; LAK
> > > <[e-mail]>; linux-kernel
> > > <[e-mail]>; ACPI Devel Maling List
> > > <[e-mail]>; Linuxarm <[e-mail]>; xuwei (O)
> > > <[e-mail]>; Zengtao (B) <[e-mail]>
> > > Subject: Re: [RFC PATCH v2 2/2] scheduler: add scheduler level for clusters
> > >
> > > On Wed, 2 Dec 2020 at 21:58, Song Bao Hua (Barry Song)
> > > <[e-mail]> wrote:
> > > >
> > > > >
> > > > > Sorry. Please ignore this. I added some printk here while testing
> > > > > one numa. Will update you the data in another email.
> > > >
> > > > Re-tested in one NUMA node(cpu0-cpu23):
> > > >
> > > > g=1
> > > > Running in threaded mode with 1 groups using 40 file descriptors
> > > > Each sender will pass 100000 messages of 100 bytes
> > > > w/o: 7.689 7.485 7.485 7.458 7.524 7.539 7.738 7.693 7.568 7.674=7.5853
> > > > w/ : 7.516 7.941 7.374 7.963 7.881 7.910 7.420 7.556 7.695 7.441=7.6697
> > > > w/ but dropped select_idle_cluster:
> > > >      7.752 7.739 7.739 7.571 7.545 7.685 7.407 7.580 7.605 7.487=7.611
> > > >
> > > > g=2
> > > > Running in threaded mode with 2 groups using 40 file descriptors
> > > > Each sender will pass 100000 messages of 100 bytes
> > > > w/o: 10.127 10.119 10.070 10.196 10.057 10.111 10.045 10.164 10.162
> > > > 9.955=10.1006
> > > > w/ : 9.694 9.654 9.612 9.649 9.686 9.734 9.607 9.842 9.690 9.710=9.6878
> > > > w/ but dropped select_idle_cluster:
> > > >      9.877 10.069 9.951 9.918 9.947 9.790 9.906 9.820 9.863 9.906=9.9047
> > > >
> > > > g=3
> > > > Running in threaded mode with 3 groups using 40 file descriptors
> > > > Each sender will pass 100000 messages of 100 bytes
> > > > w/o: 15.885 15.254 15.932 15.647 16.120 15.878 15.857 15.759 15.674
> > > > 15.721=15.7727
> > > > w/ : 14.974 14.657 13.969 14.985 14.728 15.665 15.191 14.995 14.946
> > > > 14.895=14.9005
> > > > w/ but dropped select_idle_cluster:
> > > >      15.405 15.177 15.373 15.187 15.450 15.540 15.278 15.628 15.228
> > > 15.325=15.3591
> > > >
> > > > g=4
> > > > Running in threaded mode with 4 groups using 40 file descriptors
> > > > Each sender will pass 100000 messages of 100 bytes
> > > > w/o: 20.014 21.025 21.119 21.235 19.767 20.971 20.962 20.914 21.090
> > > 21.090=20.8187
> > > > w/ : 20.331 20.608 20.338 20.445 20.456 20.146 20.693 20.797 21.381
> > > 20.452=20.5647
> > > > w/ but dropped select_idle_cluster:
> > > >      19.814 20.126 20.229 20.350 20.750 20.404 19.957 19.888 20.226
> > > 20.562=20.2306
> > > >
> > >
> > > I assume that you have run this on v5.9 as previous tests.
> >
> > Yep
> >
> > > The results don't show any real benefit of select_idle_cluster()
> > > inside a node whereas this is where we could expect most of the
> > > benefit. We have to understand why we have such an impact on numa
> > > tests only.
> >
> > There is a 4-5.5% increase while g=2 and g=3.
> 
> my point was with vs without select_idle_cluster() but still having a
> cluster domain level
> In this case, the diff is -0.8% for g=1 +2.2% for g=2, +3% for g=3 and
> -1.7% for g=4
> 
> >
> > Regarding the huge increase in NUMA case,  at the first beginning, I suspect
> > we have wrong llc domain. For example, if cpu0's llc domain span
> > cpu0-cpu47, then select_idle_cpu() is running in wrong range while
> > it should run in cpu0-cpu23.
> >
> > But after printing the llc domain's span, I find it is completely right.
> > Cpu0's llc span: cpu0-cpu23
> > Cpu24's llc span: cpu24-cpu47
> 
> Have you checked that the cluster mask was also correct ?
> 
> >
> > Maybe I need more trace data to figure out if select_idle_cpu() is running
> > correctly. For example, maybe I can figure out if it is always returning -1,
> > or it returns -1 very often?
> 
> yes, could be interesting to check how often select_idle_cpu return -1
> 
> >
> > Or do you have any idea?
> 
> tracking migration across nod could help to understand too

I set a bootargs mem=4G to do swapping test before working on cluster
scheduler issue. but I forgot to remove the parameter.

The huge increase on across-numa case can only be reproduced while
i use this mem=4G cmdline which means numa1 has no memory.
After removing the limitation, I can't reproduce the huge increase
for two NUMAs any more.

Guess select_idle_cluster() somehow workaround an scheduler issue
for numa without memory.

> 
> Vincent
> >
> >

Thanks
Barry

Vincent Guittot Dec. 7, 2020, 3:29 p.m. UTC | #17

On Mon, 7 Dec 2020 at 10:59, Song Bao Hua (Barry Song)
<[e-mail]> wrote:
>
>
>
> > -----Original Message-----
> > From: Vincent Guittot [mailto:[e-mail]]
> > Sent: Thursday, December 3, 2020 10:39 PM
> > To: Song Bao Hua (Barry Song) <[e-mail]>
> > Cc: Valentin Schneider <[e-mail]>; Catalin Marinas
> > <[e-mail]>; Will Deacon <[e-mail]>; Rafael J. Wysocki
> > <[e-mail]>; Cc: Len Brown <[e-mail]>;
> > [e-mail]; Jonathan Cameron <[e-mail]>;
> > Ingo Molnar <[e-mail]>; Peter Zijlstra <[e-mail]>; Juri
> > Lelli <[e-mail]>; Dietmar Eggemann <[e-mail]>;
> > Steven Rostedt <[e-mail]>; Ben Segall <[e-mail]>; Mel
> > Gorman <[e-mail]>; Mark Rutland <[e-mail]>; LAK
> > <[e-mail]>; linux-kernel
> > <[e-mail]>; ACPI Devel Maling List
> > <[e-mail]>; Linuxarm <[e-mail]>; xuwei (O)
> > <[e-mail]>; Zengtao (B) <[e-mail]>
> > Subject: Re: [RFC PATCH v2 2/2] scheduler: add scheduler level for clusters
> >
> > On Thu, 3 Dec 2020 at 10:11, Song Bao Hua (Barry Song)
> > <[e-mail]> wrote:
> > >
> > >
> > >
> > > > -----Original Message-----
> > > > From: Vincent Guittot [mailto:[e-mail]]
> > > > Sent: Thursday, December 3, 2020 10:04 PM
> > > > To: Song Bao Hua (Barry Song) <[e-mail]>
> > > > Cc: Valentin Schneider <[e-mail]>; Catalin Marinas
> > > > <[e-mail]>; Will Deacon <[e-mail]>; Rafael J. Wysocki
> > > > <[e-mail]>; Cc: Len Brown <[e-mail]>;
> > > > [e-mail]; Jonathan Cameron
> > <[e-mail]>;
> > > > Ingo Molnar <[e-mail]>; Peter Zijlstra <[e-mail]>; Juri
> > > > Lelli <[e-mail]>; Dietmar Eggemann
> > <[e-mail]>;
> > > > Steven Rostedt <[e-mail]>; Ben Segall <[e-mail]>; Mel
> > > > Gorman <[e-mail]>; Mark Rutland <[e-mail]>; LAK
> > > > <[e-mail]>; linux-kernel
> > > > <[e-mail]>; ACPI Devel Maling List
> > > > <[e-mail]>; Linuxarm <[e-mail]>; xuwei (O)
> > > > <[e-mail]>; Zengtao (B) <[e-mail]>
> > > > Subject: Re: [RFC PATCH v2 2/2] scheduler: add scheduler level for clusters
> > > >
> > > > On Wed, 2 Dec 2020 at 21:58, Song Bao Hua (Barry Song)
> > > > <[e-mail]> wrote:
> > > > >
> > > > > >
> > > > > > Sorry. Please ignore this. I added some printk here while testing
> > > > > > one numa. Will update you the data in another email.
> > > > >
> > > > > Re-tested in one NUMA node(cpu0-cpu23):
> > > > >
> > > > > g=1
> > > > > Running in threaded mode with 1 groups using 40 file descriptors
> > > > > Each sender will pass 100000 messages of 100 bytes
> > > > > w/o: 7.689 7.485 7.485 7.458 7.524 7.539 7.738 7.693 7.568 7.674=7.5853
> > > > > w/ : 7.516 7.941 7.374 7.963 7.881 7.910 7.420 7.556 7.695 7.441=7.6697
> > > > > w/ but dropped select_idle_cluster:
> > > > >      7.752 7.739 7.739 7.571 7.545 7.685 7.407 7.580 7.605 7.487=7.611
> > > > >
> > > > > g=2
> > > > > Running in threaded mode with 2 groups using 40 file descriptors
> > > > > Each sender will pass 100000 messages of 100 bytes
> > > > > w/o: 10.127 10.119 10.070 10.196 10.057 10.111 10.045 10.164 10.162
> > > > > 9.955=10.1006
> > > > > w/ : 9.694 9.654 9.612 9.649 9.686 9.734 9.607 9.842 9.690 9.710=9.6878
> > > > > w/ but dropped select_idle_cluster:
> > > > >      9.877 10.069 9.951 9.918 9.947 9.790 9.906 9.820 9.863 9.906=9.9047
> > > > >
> > > > > g=3
> > > > > Running in threaded mode with 3 groups using 40 file descriptors
> > > > > Each sender will pass 100000 messages of 100 bytes
> > > > > w/o: 15.885 15.254 15.932 15.647 16.120 15.878 15.857 15.759 15.674
> > > > > 15.721=15.7727
> > > > > w/ : 14.974 14.657 13.969 14.985 14.728 15.665 15.191 14.995 14.946
> > > > > 14.895=14.9005
> > > > > w/ but dropped select_idle_cluster:
> > > > >      15.405 15.177 15.373 15.187 15.450 15.540 15.278 15.628 15.228
> > > > 15.325=15.3591
> > > > >
> > > > > g=4
> > > > > Running in threaded mode with 4 groups using 40 file descriptors
> > > > > Each sender will pass 100000 messages of 100 bytes
> > > > > w/o: 20.014 21.025 21.119 21.235 19.767 20.971 20.962 20.914 21.090
> > > > 21.090=20.8187
> > > > > w/ : 20.331 20.608 20.338 20.445 20.456 20.146 20.693 20.797 21.381
> > > > 20.452=20.5647
> > > > > w/ but dropped select_idle_cluster:
> > > > >      19.814 20.126 20.229 20.350 20.750 20.404 19.957 19.888 20.226
> > > > 20.562=20.2306
> > > > >
> > > >
> > > > I assume that you have run this on v5.9 as previous tests.
> > >
> > > Yep
> > >
> > > > The results don't show any real benefit of select_idle_cluster()
> > > > inside a node whereas this is where we could expect most of the
> > > > benefit. We have to understand why we have such an impact on numa
> > > > tests only.
> > >
> > > There is a 4-5.5% increase while g=2 and g=3.
> >
> > my point was with vs without select_idle_cluster() but still having a
> > cluster domain level
> > In this case, the diff is -0.8% for g=1 +2.2% for g=2, +3% for g=3 and
> > -1.7% for g=4
> >
> > >
> > > Regarding the huge increase in NUMA case,  at the first beginning, I suspect
> > > we have wrong llc domain. For example, if cpu0's llc domain span
> > > cpu0-cpu47, then select_idle_cpu() is running in wrong range while
> > > it should run in cpu0-cpu23.
> > >
> > > But after printing the llc domain's span, I find it is completely right.
> > > Cpu0's llc span: cpu0-cpu23
> > > Cpu24's llc span: cpu24-cpu47
> >
> > Have you checked that the cluster mask was also correct ?
> >
> > >
> > > Maybe I need more trace data to figure out if select_idle_cpu() is running
> > > correctly. For example, maybe I can figure out if it is always returning -1,
> > > or it returns -1 very often?
> >
> > yes, could be interesting to check how often select_idle_cpu return -1
> >
> > >
> > > Or do you have any idea?
> >
> > tracking migration across nod could help to understand too
>
> I set a bootargs mem=4G to do swapping test before working on cluster
> scheduler issue. but I forgot to remove the parameter.
>
> The huge increase on across-numa case can only be reproduced while
> i use this mem=4G cmdline which means numa1 has no memory.
> After removing the limitation, I can't reproduce the huge increase
> for two NUMAs any more.

Ok. Make more sense

>
> Guess select_idle_cluster() somehow workaround an scheduler issue
> for numa without memory.
>
> >
> > Vincent
> > >
> > >
>
> Thanks
> Barry
>

Song Bao Hua (Barry Song) Dec. 9, 2020, 11:35 a.m. UTC | #18

> -----Original Message-----
> From: Vincent Guittot [mailto:[e-mail]]
> Sent: Tuesday, December 8, 2020 4:29 AM
> To: Song Bao Hua (Barry Song) <[e-mail]>
> Cc: Valentin Schneider <[e-mail]>; Catalin Marinas
> <[e-mail]>; Will Deacon <[e-mail]>; Rafael J. Wysocki
> <[e-mail]>; Cc: Len Brown <[e-mail]>;
> [e-mail]; Jonathan Cameron <[e-mail]>;
> Ingo Molnar <[e-mail]>; Peter Zijlstra <[e-mail]>; Juri
> Lelli <[e-mail]>; Dietmar Eggemann <[e-mail]>;
> Steven Rostedt <[e-mail]>; Ben Segall <[e-mail]>; Mel
> Gorman <[e-mail]>; Mark Rutland <[e-mail]>; LAK
> <[e-mail]>; linux-kernel
> <[e-mail]>; ACPI Devel Maling List
> <[e-mail]>; Linuxarm <[e-mail]>; xuwei (O)
> <[e-mail]>; Zengtao (B) <[e-mail]>
> Subject: Re: [RFC PATCH v2 2/2] scheduler: add scheduler level for clusters
> 
> On Mon, 7 Dec 2020 at 10:59, Song Bao Hua (Barry Song)
> <[e-mail]> wrote:
> >
> >
> >
> > > -----Original Message-----
> > > From: Vincent Guittot [mailto:[e-mail]]
> > > Sent: Thursday, December 3, 2020 10:39 PM
> > > To: Song Bao Hua (Barry Song) <[e-mail]>
> > > Cc: Valentin Schneider <[e-mail]>; Catalin Marinas
> > > <[e-mail]>; Will Deacon <[e-mail]>; Rafael J. Wysocki
> > > <[e-mail]>; Cc: Len Brown <[e-mail]>;
> > > [e-mail]; Jonathan Cameron
> <[e-mail]>;
> > > Ingo Molnar <[e-mail]>; Peter Zijlstra <[e-mail]>; Juri
> > > Lelli <[e-mail]>; Dietmar Eggemann
> <[e-mail]>;
> > > Steven Rostedt <[e-mail]>; Ben Segall <[e-mail]>; Mel
> > > Gorman <[e-mail]>; Mark Rutland <[e-mail]>; LAK
> > > <[e-mail]>; linux-kernel
> > > <[e-mail]>; ACPI Devel Maling List
> > > <[e-mail]>; Linuxarm <[e-mail]>; xuwei (O)
> > > <[e-mail]>; Zengtao (B) <[e-mail]>
> > > Subject: Re: [RFC PATCH v2 2/2] scheduler: add scheduler level for clusters
> > >
> > > On Thu, 3 Dec 2020 at 10:11, Song Bao Hua (Barry Song)
> > > <[e-mail]> wrote:
> > > >
> > > >
> > > >
> > > > > -----Original Message-----
> > > > > From: Vincent Guittot [mailto:[e-mail]]
> > > > > Sent: Thursday, December 3, 2020 10:04 PM
> > > > > To: Song Bao Hua (Barry Song) <[e-mail]>
> > > > > Cc: Valentin Schneider <[e-mail]>; Catalin Marinas
> > > > > <[e-mail]>; Will Deacon <[e-mail]>; Rafael J.
> Wysocki
> > > > > <[e-mail]>; Cc: Len Brown <[e-mail]>;
> > > > > [e-mail]; Jonathan Cameron
> > > <[e-mail]>;
> > > > > Ingo Molnar <[e-mail]>; Peter Zijlstra <[e-mail]>;
> Juri
> > > > > Lelli <[e-mail]>; Dietmar Eggemann
> > > <[e-mail]>;
> > > > > Steven Rostedt <[e-mail]>; Ben Segall <[e-mail]>;
> Mel
> > > > > Gorman <[e-mail]>; Mark Rutland <[e-mail]>; LAK
> > > > > <[e-mail]>; linux-kernel
> > > > > <[e-mail]>; ACPI Devel Maling List
> > > > > <[e-mail]>; Linuxarm <[e-mail]>; xuwei
> (O)
> > > > > <[e-mail]>; Zengtao (B) <[e-mail]>
> > > > > Subject: Re: [RFC PATCH v2 2/2] scheduler: add scheduler level for clusters
> > > > >
> > > > > On Wed, 2 Dec 2020 at 21:58, Song Bao Hua (Barry Song)
> > > > > <[e-mail]> wrote:
> > > > > >
> > > > > > >
> > > > > > > Sorry. Please ignore this. I added some printk here while testing
> > > > > > > one numa. Will update you the data in another email.
> > > > > >
> > > > > > Re-tested in one NUMA node(cpu0-cpu23):
> > > > > >
> > > > > > g=1
> > > > > > Running in threaded mode with 1 groups using 40 file descriptors
> > > > > > Each sender will pass 100000 messages of 100 bytes
> > > > > > w/o: 7.689 7.485 7.485 7.458 7.524 7.539 7.738 7.693 7.568 7.674=7.5853
> > > > > > w/ : 7.516 7.941 7.374 7.963 7.881 7.910 7.420 7.556 7.695 7.441=7.6697
> > > > > > w/ but dropped select_idle_cluster:
> > > > > >      7.752 7.739 7.739 7.571 7.545 7.685 7.407 7.580 7.605 7.487=7.611
> > > > > >
> > > > > > g=2
> > > > > > Running in threaded mode with 2 groups using 40 file descriptors
> > > > > > Each sender will pass 100000 messages of 100 bytes
> > > > > > w/o: 10.127 10.119 10.070 10.196 10.057 10.111 10.045 10.164 10.162
> > > > > > 9.955=10.1006
> > > > > > w/ : 9.694 9.654 9.612 9.649 9.686 9.734 9.607 9.842 9.690 9.710=9.6878
> > > > > > w/ but dropped select_idle_cluster:
> > > > > >      9.877 10.069 9.951 9.918 9.947 9.790 9.906 9.820 9.863 9.906=9.9047
> > > > > >
> > > > > > g=3
> > > > > > Running in threaded mode with 3 groups using 40 file descriptors
> > > > > > Each sender will pass 100000 messages of 100 bytes
> > > > > > w/o: 15.885 15.254 15.932 15.647 16.120 15.878 15.857 15.759 15.674
> > > > > > 15.721=15.7727
> > > > > > w/ : 14.974 14.657 13.969 14.985 14.728 15.665 15.191 14.995 14.946
> > > > > > 14.895=14.9005
> > > > > > w/ but dropped select_idle_cluster:
> > > > > >      15.405 15.177 15.373 15.187 15.450 15.540 15.278 15.628 15.228
> > > > > 15.325=15.3591
> > > > > >
> > > > > > g=4
> > > > > > Running in threaded mode with 4 groups using 40 file descriptors
> > > > > > Each sender will pass 100000 messages of 100 bytes
> > > > > > w/o: 20.014 21.025 21.119 21.235 19.767 20.971 20.962 20.914 21.090
> > > > > 21.090=20.8187
> > > > > > w/ : 20.331 20.608 20.338 20.445 20.456 20.146 20.693 20.797 21.381
> > > > > 20.452=20.5647
> > > > > > w/ but dropped select_idle_cluster:
> > > > > >      19.814 20.126 20.229 20.350 20.750 20.404 19.957 19.888 20.226
> > > > > 20.562=20.2306
> > > > > >
> > > > >
> > > > > I assume that you have run this on v5.9 as previous tests.
> > > >
> > > > Yep
> > > >
> > > > > The results don't show any real benefit of select_idle_cluster()
> > > > > inside a node whereas this is where we could expect most of the
> > > > > benefit. We have to understand why we have such an impact on numa
> > > > > tests only.
> > > >
> > > > There is a 4-5.5% increase while g=2 and g=3.
> > >
> > > my point was with vs without select_idle_cluster() but still having a
> > > cluster domain level
> > > In this case, the diff is -0.8% for g=1 +2.2% for g=2, +3% for g=3 and
> > > -1.7% for g=4
> > >
> > > >
> > > > Regarding the huge increase in NUMA case,  at the first beginning, I suspect
> > > > we have wrong llc domain. For example, if cpu0's llc domain span
> > > > cpu0-cpu47, then select_idle_cpu() is running in wrong range while
> > > > it should run in cpu0-cpu23.
> > > >
> > > > But after printing the llc domain's span, I find it is completely right.
> > > > Cpu0's llc span: cpu0-cpu23
> > > > Cpu24's llc span: cpu24-cpu47
> > >
> > > Have you checked that the cluster mask was also correct ?
> > >
> > > >
> > > > Maybe I need more trace data to figure out if select_idle_cpu() is running
> > > > correctly. For example, maybe I can figure out if it is always returning
> -1,
> > > > or it returns -1 very often?
> > >
> > > yes, could be interesting to check how often select_idle_cpu return -1
> > >
> > > >
> > > > Or do you have any idea?
> > >
> > > tracking migration across nod could help to understand too
> >
> > I set a bootargs mem=4G to do swapping test before working on cluster
> > scheduler issue. but I forgot to remove the parameter.
> >
> > The huge increase on across-numa case can only be reproduced while
> > i use this mem=4G cmdline which means numa1 has no memory.
> > After removing the limitation, I can't reproduce the huge increase
> > for two NUMAs any more.
> 
> Ok. Make more sense

I managed to use linux-next to test after fixing the disk hang.

But I am still quite struggling with how to leverage the cluster
topology in select_idle_cpu() to make huge improvement on benchmark.

If I disable the influence of scheduler by taskset, there is
obviously a large difference in hackbench inside cluster and
across clusters:

inside a cluster:
root@ubuntu:~# taskset -c 0,1,2,3 hackbench -p -T -l 20000 -g 1
Running in threaded mode with 1 groups using 40 file descriptors each
(== 40 tasks)
Each sender will pass 20000 messages of 100 bytes
Time: 4.285

Across clusters:
root@ubuntu:~# taskset -c 0,4,8,12 hackbench -p -T -l 20000 -g 1
Running in threaded mode with 1 groups using 40 file descriptors each
(== 40 tasks)
Each sender will pass 20000 messages of 100 bytes
Time: 5.524

But no matter how I tune the code of kernel/sched/fair.c, I
don't see this large difference by running hackbench in the
whole numa node:
for i in {1..10}
do
	numactl -N 0 hackbench -p -T -l 20000 -g $1
done

usually, the difference is under (-5%~+5%).

Then I made a major change as below:
static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, int target)
{
	...


	time = cpu_clock(this);

#if 0
	cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);

	for_each_cpu_wrap(cpu, cpus, target) {
		if (!--nr)
			return -1;
		if (available_idle_cpu(cpu) || sched_idle_cpu(cpu))
			break;
	}
#else
	if ((cpu=select_idle_cluster(p,target)) = -1)
		return -1;
#endif

	time = cpu_clock(this) - time;
	update_avg(&this_sd->avg_scan_cost, time);

	return cpu;
}

That means I don't fall back to llc if cluster has no idle
cpu.

With this, I am getting 20% major difference as I am always expecting:

g=     1      2        3        4        5        6        7      8       9        10
w/o  1.5494 2.0641 3.1640 4.2438 5.3445 6.3098 7.5086 8.4721 9.7115  10.8588
w/   1.6801 2.0280 2.7890 3.7339 4.5748 5.2998 6.1413 6.6206 7.7641  8.4782

I guess my original patch is very easy to fall back to llc as
cluster is not easy to idle. Once system is busy, the original
patch is nop as it is always falling back to llc.

> 
> >
> > Guess select_idle_cluster() somehow workaround an scheduler issue
> > for numa without memory.
> >
> > >
> > > Vincent

Thanks
Barry

Patch

diff --git a/arch/arm64/Kconfig b/arch/arm64/Kconfig
index 6d23283..3583c26 100644
--- a/arch/arm64/Kconfig
+++ b/arch/arm64/Kconfig
@@ -938,6 +938,13 @@  config SCHED_MC
 	  making when dealing with multi-core CPU chips at a cost of slightly
 	  increased overhead in some places. If unsure say N here.
 
+config SCHED_CLUSTER
+	bool "Cluster scheduler support"
+	help
+	  Cluster scheduler support improves the CPU scheduler's decision
+	  making when dealing with machines that have clusters(sharing internal
+	  bus or sharing LLC cache tag). If unsure say N here.
+
 config SCHED_SMT
 	bool "SMT scheduler support"
 	help
diff --git a/arch/arm64/kernel/smp.c b/arch/arm64/kernel/smp.c
index 355ee9e..5c8f026 100644
--- a/arch/arm64/kernel/smp.c
+++ b/arch/arm64/kernel/smp.c
@@ -32,6 +32,7 @@ 
 #include <linux/irq_work.h>
 #include <linux/kexec.h>
 #include <linux/kvm_host.h>
+#include <linux/sched/topology.h>
 
 #include <asm/alternative.h>
 #include <asm/atomic.h>
@@ -726,6 +727,20 @@  void __init smp_init_cpus(void)
 	}
 }
 
+static struct sched_domain_topology_level arm64_topology[] = {
+#ifdef CONFIG_SCHED_SMT
+        { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
+#endif
+#ifdef CONFIG_SCHED_CLUSTER
+	{ cpu_clustergroup_mask, cpu_core_flags, SD_INIT_NAME(CL) },
+#endif
+#ifdef CONFIG_SCHED_MC
+        { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
+#endif
+	{ cpu_cpu_mask, SD_INIT_NAME(DIE) },
+        { NULL, },
+};
+
 void __init smp_prepare_cpus(unsigned int max_cpus)
 {
 	const struct cpu_operations *ops;
@@ -735,6 +750,8 @@  void __init smp_prepare_cpus(unsigned int max_cpus)
 
 	init_cpu_topology();
 
+	set_sched_topology(arm64_topology);
+
 	this_cpu = smp_processor_id();
 	store_cpu_topology(this_cpu);
 	numa_store_cpu_info(this_cpu);
diff --git a/include/linux/topology.h b/include/linux/topology.h
index 5f66648..2c823c0 100644
--- a/include/linux/topology.h
+++ b/include/linux/topology.h
@@ -211,6 +211,13 @@  static inline const struct cpumask *cpu_smt_mask(int cpu)
 }
 #endif
 
+#ifdef CONFIG_SCHED_CLUSTER
+static inline const struct cpumask *cpu_cluster_mask(int cpu)
+{
+	return topology_cluster_cpumask(cpu);
+}
+#endif
+
 static inline const struct cpumask *cpu_cpu_mask(int cpu)
 {
 	return cpumask_of_node(cpu_to_node(cpu));
diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
index 1a68a05..ae8ec910 100644
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -6106,6 +6106,37 @@  static inline int select_idle_smt(struct task_struct *p, int target)
 
 #endif /* CONFIG_SCHED_SMT */
 
+#ifdef CONFIG_SCHED_CLUSTER
+/*
+ * Scan the local CLUSTER mask for idle CPUs.
+ */
+static int select_idle_cluster(struct task_struct *p, int target)
+{
+	int cpu;
+
+	/* right now, no hardware with both cluster and smt to run */
+	if (sched_smt_active())
+		return -1;
+
+	for_each_cpu_wrap(cpu, cpu_cluster_mask(target), target) {
+		if (!cpumask_test_cpu(cpu, p->cpus_ptr))
+			continue;
+		if (available_idle_cpu(cpu))
+			return cpu;
+	}
+
+	return -1;
+}
+
+#else /* CONFIG_SCHED_CLUSTER */
+
+static inline int select_idle_cluster(struct task_struct *p, int target)
+{
+	return -1;
+}
+
+#endif /* CONFIG_SCHED_CLUSTER */
+
 /*
  * Scan the LLC domain for idle CPUs; this is dynamically regulated by
  * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
@@ -6270,6 +6301,10 @@  static int select_idle_sibling(struct task_struct *p, int prev, int target)
 	if ((unsigned)i < nr_cpumask_bits)
 		return i;
 
+	i = select_idle_cluster(p, target);
+	if ((unsigned)i < nr_cpumask_bits)
+		return i;
+
 	i = select_idle_cpu(p, sd, target);
 	if ((unsigned)i < nr_cpumask_bits)
 		return i;