| Message ID | 20250402010611.3204674-1-yu.c.chen@intel.com (mailing list archive) |
|---|---|
| State | New |
| Headers | show |
| Series | sched/numa: Add statistics of numa balance task migration and swap | expand |
Hello Chen. On Wed, Apr 02, 2025 at 09:06:11AM +0800, Chen Yu <yu.c.chen@intel.com> wrote: > On system with NUMA balancing enabled, it is found that tracking > the task activities due to NUMA balancing is helpful. ... > The following two new fields: > > numa_task_migrated > numa_task_swapped > > will be displayed in both > /sys/fs/cgroup/{GROUP}/memory.stat and /proc/{PID}/sched Why is the field /proc/$pid/sched not enough? Also, you may want to update Documentation/admin-guide/cgroup-v2.rst too. Thanks, Michal
Hi Chen Yu, On 02/04/25 06:36, Chen Yu wrote: > On system with NUMA balancing enabled, it is found that tracking > the task activities due to NUMA balancing is helpful. NUMA balancing > has two mechanisms for task migration: one is to migrate the task to > an idle CPU in its preferred node, the other is to swap tasks on > different nodes if they are on each other's preferred node. > > The kernel already has NUMA page migration statistics in > /sys/fs/cgroup/mytest/memory.stat and /proc/{PID}/sched. > but does not have statistics for task migration/swap. > Add the task migration and swap count accordingly. > > The following two new fields: > > numa_task_migrated > numa_task_swapped > > will be displayed in both > /sys/fs/cgroup/{GROUP}/memory.stat and /proc/{PID}/sched I applied this patch, but I still don't see the two new fields in /proc/{PID}/sched. Am I missing any additional steps? Thanks, Madadi Vineeth Reddy > > Previous RFC version can be found here: > https://lore.kernel.org/lkml/1847c5ef828ad4835a35e3a54b88d2e13bce0eea.1740483690.git.yu.c.chen@intel.com/ > > Signed-off-by: Chen Yu <yu.c.chen@intel.com> > --- > RFC->v1: Rename the nr_numa_task_migrated to > numa_task_migrated, and nr_numa_task_swapped > numa_task_swapped in /proc/{PID}/sched, > so both cgroup's memory.stat and task's > sched have the same field name. > --- > include/linux/sched.h | 4 ++++ > include/linux/vm_event_item.h | 2 ++ > kernel/sched/core.c | 10 ++++++++-- > kernel/sched/debug.c | 4 ++++ > mm/memcontrol.c | 2 ++ > mm/vmstat.c | 2 ++ > 6 files changed, 22 insertions(+), 2 deletions(-)
On 4/2/2025 7:03 PM, Madadi Vineeth Reddy wrote: > Hi Chen Yu, > > On 02/04/25 06:36, Chen Yu wrote: >> On system with NUMA balancing enabled, it is found that tracking >> the task activities due to NUMA balancing is helpful. NUMA balancing >> has two mechanisms for task migration: one is to migrate the task to >> an idle CPU in its preferred node, the other is to swap tasks on >> different nodes if they are on each other's preferred node. >> >> The kernel already has NUMA page migration statistics in >> /sys/fs/cgroup/mytest/memory.stat and /proc/{PID}/sched. >> but does not have statistics for task migration/swap. >> Add the task migration and swap count accordingly. >> >> The following two new fields: >> >> numa_task_migrated >> numa_task_swapped >> >> will be displayed in both >> /sys/fs/cgroup/{GROUP}/memory.stat and /proc/{PID}/sched > > I applied this patch, but I still don't see the two new fields > in /proc/{PID}/sched. > > Am I missing any additional steps? You also need to enable schedstats: echo 1 > /proc/sys/kernel/sched_schedstats After that it should be visible: $ cat /proc/4030/sched sched-messaging (4030, #threads: 641) ------------------------------------------------------------------- se.exec_start : 283818.948537 ... nr_forced_migrations : 0 numa_task_migrated : 0 numa_task_swapped : 0 nr_wakeups : 0 ...
Hello Chenyu, On 4/2/2025 6:36 AM, Chen Yu wrote: > On system with NUMA balancing enabled, it is found that tracking > the task activities due to NUMA balancing is helpful. NUMA balancing > has two mechanisms for task migration: one is to migrate the task to > an idle CPU in its preferred node, the other is to swap tasks on > different nodes if they are on each other's preferred node. > > The kernel already has NUMA page migration statistics in > /sys/fs/cgroup/mytest/memory.stat and /proc/{PID}/sched. > but does not have statistics for task migration/swap. > Add the task migration and swap count accordingly. > > The following two new fields: > > numa_task_migrated > numa_task_swapped > > will be displayed in both > /sys/fs/cgroup/{GROUP}/memory.stat and /proc/{PID}/sched Running sched-messaging with schedstats enabled, I could see both "numa_task_migrated" and "numa_task_swapped" being populated for the sched-messaging threads: $ for i in $(ls /proc/4030/task/); do grep "numa_task_migrated" /proc/$i/sched; done | tr -s ' ' | cut -d ' ' -f3 | sort | uniq -c 400 0 231 1 10 2 $ for i in $(ls /proc/4030/task/); do grep "numa_task_swapped" /proc/$i/sched; done | tr -s ' ' | cut -d ' ' -f3 | sort | uniq -c 389 0 193 1 47 2 11 3 1 4 > > Previous RFC version can be found here: > https://lore.kernel.org/lkml/1847c5ef828ad4835a35e3a54b88d2e13bce0eea.1740483690.git.yu.c.chen@intel.com/ > > Signed-off-by: Chen Yu <yu.c.chen@intel.com> Feel free to add: Tested-by: K Prateek Nayak <kprateek.nayak@amd.com>
Hello Michal, On 4/2/2025 6:54 PM, Michal Koutný wrote: > Hello Chen. > > On Wed, Apr 02, 2025 at 09:06:11AM +0800, Chen Yu <yu.c.chen@intel.com> wrote: >> On system with NUMA balancing enabled, it is found that tracking >> the task activities due to NUMA balancing is helpful. > ... >> The following two new fields: >> >> numa_task_migrated >> numa_task_swapped >> >> will be displayed in both >> /sys/fs/cgroup/{GROUP}/memory.stat and /proc/{PID}/sched > > Why is the field /proc/$pid/sched not enough? The /proc/$pid/sched accounting is only done when schedstats are enabled. memcg users might want to track it separately without relying on schedstats which also enables a bunch of other scheduler related stats collection adding more overheads. > > Also, you may want to update Documentation/admin-guide/cgroup-v2.rst > too. > > Thanks, > Michal
On 02/04/25 22:53, K Prateek Nayak wrote: > On 4/2/2025 7:03 PM, Madadi Vineeth Reddy wrote: >> Hi Chen Yu, >> >> On 02/04/25 06:36, Chen Yu wrote: >>> On system with NUMA balancing enabled, it is found that tracking >>> the task activities due to NUMA balancing is helpful. NUMA balancing >>> has two mechanisms for task migration: one is to migrate the task to >>> an idle CPU in its preferred node, the other is to swap tasks on >>> different nodes if they are on each other's preferred node. >>> >>> The kernel already has NUMA page migration statistics in >>> /sys/fs/cgroup/mytest/memory.stat and /proc/{PID}/sched. >>> but does not have statistics for task migration/swap. >>> Add the task migration and swap count accordingly. >>> >>> The following two new fields: >>> >>> numa_task_migrated >>> numa_task_swapped >>> >>> will be displayed in both >>> /sys/fs/cgroup/{GROUP}/memory.stat and /proc/{PID}/sched >> >> I applied this patch, but I still don't see the two new fields >> in /proc/{PID}/sched. >> >> Am I missing any additional steps? > > You also need to enable schedstats: > > echo 1 > /proc/sys/kernel/sched_schedstats > > After that it should be visible: Thanks, Prateek! I had missed enabling schedstats. Now that it's enabled, I can see the fields. Thanks, Madadi Vineeth Reddy > > $ cat /proc/4030/sched > sched-messaging (4030, #threads: 641) > ------------------------------------------------------------------- > se.exec_start : 283818.948537 > > ... > > nr_forced_migrations : 0 > numa_task_migrated : 0 > numa_task_swapped : 0 > nr_wakeups : 0 > > ... >
Hi Chen Yu, On 02/04/25 06:36, Chen Yu wrote: > On system with NUMA balancing enabled, it is found that tracking > the task activities due to NUMA balancing is helpful. NUMA balancing > has two mechanisms for task migration: one is to migrate the task to > an idle CPU in its preferred node, the other is to swap tasks on > different nodes if they are on each other's preferred node. > > The kernel already has NUMA page migration statistics in > /sys/fs/cgroup/mytest/memory.stat and /proc/{PID}/sched. > but does not have statistics for task migration/swap. > Add the task migration and swap count accordingly. > > The following two new fields: > > numa_task_migrated > numa_task_swapped > > will be displayed in both > /sys/fs/cgroup/{GROUP}/memory.stat and /proc/{PID}/sched I was able to see the fields and their corresponding values for schbench: numa_task_swapped : 2 numa_task_migrated : 0 numa_task_swapped : 1 numa_task_migrated : 0 numa_task_swapped : 0 numa_task_migrated : 0 numa_task_swapped : 1 Tested-by: Madadi Vineeth Reddy <vineethr@linux.ibm.com> Thanks, Madadi Vineeth Reddy > Previous RFC version can be found here: > https://lore.kernel.org/lkml/1847c5ef828ad4835a35e3a54b88d2e13bce0eea.1740483690.git.yu.c.chen@intel.com/ > > Signed-off-by: Chen Yu <yu.c.chen@intel.com> > --- > RFC->v1: Rename the nr_numa_task_migrated to > numa_task_migrated, and nr_numa_task_swapped > numa_task_swapped in /proc/{PID}/sched, > so both cgroup's memory.stat and task's > sched have the same field name. > --- > include/linux/sched.h | 4 ++++ > include/linux/vm_event_item.h | 2 ++ > kernel/sched/core.c | 10 ++++++++-- > kernel/sched/debug.c | 4 ++++ > mm/memcontrol.c | 2 ++ > mm/vmstat.c | 2 ++ > 6 files changed, 22 insertions(+), 2 deletions(-)
Hi Michal, thanks for taking a look at this, On 4/2/2025 9:24 PM, Michal Koutný wrote: > Hello Chen. > > On Wed, Apr 02, 2025 at 09:06:11AM +0800, Chen Yu <yu.c.chen@intel.com> wrote: >> On system with NUMA balancing enabled, it is found that tracking >> the task activities due to NUMA balancing is helpful. > ... >> The following two new fields: >> >> numa_task_migrated >> numa_task_swapped >> >> will be displayed in both >> /sys/fs/cgroup/{GROUP}/memory.stat and /proc/{PID}/sched > > Why is the field /proc/$pid/sched not enough? > In the context of NUMA balancing, it would be helpful to not only monitor on the activities of individual task/thread but also the resource usage and task migrations at the group level - which helps us quickly evaluate the performance and resource usage of the container - like per memcg numa_pages_migrated, numa_pte_updates introduced in commit f77f0c751478 ("mm,memcg: provide per-cgroup counters for NUMA balancing operations"). Yes, we can iterate the /proc/$pid/sched to find the accumulated NUMA stat, and the introduction of per - cgroup numa stat can help users more conveniently track the overall data of the workload. Besides, I'm considering evaluating the per - cgroup NUMA balance control[1] to help users do fine - grain control per workload. This per - cgroup NUMA balance stat could be used to evaluate the efficiency of per - cgroup NUMA balance. > Also, you may want to update Documentation/admin-guide/cgroup-v2.rst > too. Got it, will do in next version. [1] https://lore.kernel.org/lkml/b3f1f6c478127a38b9091a8341374ba160d25c5a.1740483690.git.yu.c.chen@intel.com/ thanks, Chenyu > > Thanks, > Michal
On 4/3/2025 1:35 AM, K Prateek Nayak wrote: > Hello Chenyu, > > On 4/2/2025 6:36 AM, Chen Yu wrote: >> On system with NUMA balancing enabled, it is found that tracking >> the task activities due to NUMA balancing is helpful. NUMA balancing >> has two mechanisms for task migration: one is to migrate the task to >> an idle CPU in its preferred node, the other is to swap tasks on >> different nodes if they are on each other's preferred node. >> >> The kernel already has NUMA page migration statistics in >> /sys/fs/cgroup/mytest/memory.stat and /proc/{PID}/sched. >> but does not have statistics for task migration/swap. >> Add the task migration and swap count accordingly. >> >> The following two new fields: >> >> numa_task_migrated >> numa_task_swapped >> >> will be displayed in both >> /sys/fs/cgroup/{GROUP}/memory.stat and /proc/{PID}/sched > > Running sched-messaging with schedstats enabled, I could see both > "numa_task_migrated" and "numa_task_swapped" being populated for the > sched-messaging threads: > > $ for i in $(ls /proc/4030/task/); do grep "numa_task_migrated" /proc/ > $i/sched; done | tr -s ' ' | cut -d ' ' -f3 | sort | uniq -c > 400 0 > 231 1 > 10 2 > > $ for i in $(ls /proc/4030/task/); do grep "numa_task_swapped" /proc/$i/ > sched; done | tr -s ' ' | cut -d ' ' -f3 | sort | uniq -c > 389 0 > 193 1 > 47 2 > 11 3 > 1 4 > >> >> Previous RFC version can be found here: >> https://lore.kernel.org/ >> lkml/1847c5ef828ad4835a35e3a54b88d2e13bce0eea.1740483690.git.yu.c.chen@intel.com/ >> >> Signed-off-by: Chen Yu <yu.c.chen@intel.com> > > Feel free to add: > > Tested-by: K Prateek Nayak <kprateek.nayak@amd.com> > Thanks Prateek! Thanks, Chenyu
On 4/3/2025 2:50 AM, Madadi Vineeth Reddy wrote: > Hi Chen Yu, > > On 02/04/25 06:36, Chen Yu wrote: >> On system with NUMA balancing enabled, it is found that tracking >> the task activities due to NUMA balancing is helpful. NUMA balancing >> has two mechanisms for task migration: one is to migrate the task to >> an idle CPU in its preferred node, the other is to swap tasks on >> different nodes if they are on each other's preferred node. >> >> The kernel already has NUMA page migration statistics in >> /sys/fs/cgroup/mytest/memory.stat and /proc/{PID}/sched. >> but does not have statistics for task migration/swap. >> Add the task migration and swap count accordingly. >> >> The following two new fields: >> >> numa_task_migrated >> numa_task_swapped >> >> will be displayed in both >> /sys/fs/cgroup/{GROUP}/memory.stat and /proc/{PID}/sched > > I was able to see the fields and their corresponding values for schbench: > > numa_task_swapped : 2 > numa_task_migrated : 0 > numa_task_swapped : 1 > numa_task_migrated : 0 > numa_task_swapped : 0 > numa_task_migrated : 0 > numa_task_swapped : 1 > > Tested-by: Madadi Vineeth Reddy <vineethr@linux.ibm.com> > Yes, the sysfs schedstat has to be enabled. Thanks for your test, Madadi! thanks, Chenyu > Thanks, > Madadi Vineeth Reddy > >> Previous RFC version can be found here: >> https://lore.kernel.org/lkml/1847c5ef828ad4835a35e3a54b88d2e13bce0eea.1740483690.git.yu.c.chen@intel.com/ >> >> Signed-off-by: Chen Yu <yu.c.chen@intel.com> >> --- >> RFC->v1: Rename the nr_numa_task_migrated to >> numa_task_migrated, and nr_numa_task_swapped >> numa_task_swapped in /proc/{PID}/sched, >> so both cgroup's memory.stat and task's >> sched have the same field name. >> --- >> include/linux/sched.h | 4 ++++ >> include/linux/vm_event_item.h | 2 ++ >> kernel/sched/core.c | 10 ++++++++-- >> kernel/sched/debug.c | 4 ++++ >> mm/memcontrol.c | 2 ++ >> mm/vmstat.c | 2 ++ >> 6 files changed, 22 insertions(+), 2 deletions(-)
On Wed, Apr 02, 2025 at 11:13:03PM +0530, K Prateek Nayak <kprateek.nayak@amd.com> wrote: > The /proc/$pid/sched accounting is only done when schedstats are > enabled. memcg users might want to track it separately without relying > on schedstats which also enables a bunch of other scheduler related > stats collection adding more overheads. .oO(memory.[numa_]stat could end up with something similar since not all users are interested in all fields but all users are affected by added fields) Thanks, Michal
On Thu, Apr 03, 2025 at 10:47:44AM +0800, "Chen, Yu C" <yu.c.chen@intel.com> wrote: > In the context of NUMA balancing, it would be helpful to not only monitor on > the activities of individual task/thread but also the resource usage and > task migrations at the group level - which helps us quickly evaluate the > performance and resource usage of the container - like per memcg > numa_pages_migrated, numa_pte_updates introduced in Somehow I thought that these are the useful metrics (amount of misplaced memory) and the aggregated task stats aren't interesting (they'd be more or less proportion of total number of tasks in the group and you don't know which task it was). > Got it, will do in next version. OK. Thanks, Michal
diff --git a/include/linux/sched.h b/include/linux/sched.h index 0785268c76f8..9623e5300453 100644 --- a/include/linux/sched.h +++ b/include/linux/sched.h @@ -532,6 +532,10 @@ struct sched_statistics { u64 nr_failed_migrations_running; u64 nr_failed_migrations_hot; u64 nr_forced_migrations; +#ifdef CONFIG_NUMA_BALANCING + u64 numa_task_migrated; + u64 numa_task_swapped; +#endif u64 nr_wakeups; u64 nr_wakeups_sync; diff --git a/include/linux/vm_event_item.h b/include/linux/vm_event_item.h index f70d0958095c..aef817474781 100644 --- a/include/linux/vm_event_item.h +++ b/include/linux/vm_event_item.h @@ -64,6 +64,8 @@ enum vm_event_item { PGPGIN, PGPGOUT, PSWPIN, PSWPOUT, NUMA_HINT_FAULTS, NUMA_HINT_FAULTS_LOCAL, NUMA_PAGE_MIGRATE, + NUMA_TASK_MIGRATE, + NUMA_TASK_SWAP, #endif #ifdef CONFIG_MIGRATION PGMIGRATE_SUCCESS, PGMIGRATE_FAIL, diff --git a/kernel/sched/core.c b/kernel/sched/core.c index c86c05264719..314d5cbce2b6 100644 --- a/kernel/sched/core.c +++ b/kernel/sched/core.c @@ -3348,6 +3348,11 @@ void set_task_cpu(struct task_struct *p, unsigned int new_cpu) #ifdef CONFIG_NUMA_BALANCING static void __migrate_swap_task(struct task_struct *p, int cpu) { + __schedstat_inc(p->stats.numa_task_swapped); + + if (p->mm) + count_memcg_events_mm(p->mm, NUMA_TASK_SWAP, 1); + if (task_on_rq_queued(p)) { struct rq *src_rq, *dst_rq; struct rq_flags srf, drf; @@ -7948,8 +7953,9 @@ int migrate_task_to(struct task_struct *p, int target_cpu) if (!cpumask_test_cpu(target_cpu, p->cpus_ptr)) return -EINVAL; - /* TODO: This is not properly updating schedstats */ - + __schedstat_inc(p->stats.numa_task_migrated); + if (p->mm) + count_memcg_events_mm(p->mm, NUMA_TASK_MIGRATE, 1); trace_sched_move_numa(p, curr_cpu, target_cpu); return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg); } diff --git a/kernel/sched/debug.c b/kernel/sched/debug.c index 56ae54e0ce6a..f971c2af7912 100644 --- a/kernel/sched/debug.c +++ b/kernel/sched/debug.c @@ -1206,6 +1206,10 @@ void proc_sched_show_task(struct task_struct *p, struct pid_namespace *ns, P_SCHEDSTAT(nr_failed_migrations_running); P_SCHEDSTAT(nr_failed_migrations_hot); P_SCHEDSTAT(nr_forced_migrations); +#ifdef CONFIG_NUMA_BALANCING + P_SCHEDSTAT(numa_task_migrated); + P_SCHEDSTAT(numa_task_swapped); +#endif P_SCHEDSTAT(nr_wakeups); P_SCHEDSTAT(nr_wakeups_sync); P_SCHEDSTAT(nr_wakeups_migrate); diff --git a/mm/memcontrol.c b/mm/memcontrol.c index 4de6acb9b8ec..1656c90b2381 100644 --- a/mm/memcontrol.c +++ b/mm/memcontrol.c @@ -460,6 +460,8 @@ static const unsigned int memcg_vm_event_stat[] = { NUMA_PAGE_MIGRATE, NUMA_PTE_UPDATES, NUMA_HINT_FAULTS, + NUMA_TASK_MIGRATE, + NUMA_TASK_SWAP, #endif }; diff --git a/mm/vmstat.c b/mm/vmstat.c index 16bfe1c694dd..7de1583a63c9 100644 --- a/mm/vmstat.c +++ b/mm/vmstat.c @@ -1339,6 +1339,8 @@ const char * const vmstat_text[] = { "numa_hint_faults", "numa_hint_faults_local", "numa_pages_migrated", + "numa_task_migrated", + "numa_task_swapped", #endif #ifdef CONFIG_MIGRATION "pgmigrate_success",
On system with NUMA balancing enabled, it is found that tracking the task activities due to NUMA balancing is helpful. NUMA balancing has two mechanisms for task migration: one is to migrate the task to an idle CPU in its preferred node, the other is to swap tasks on different nodes if they are on each other's preferred node. The kernel already has NUMA page migration statistics in /sys/fs/cgroup/mytest/memory.stat and /proc/{PID}/sched. but does not have statistics for task migration/swap. Add the task migration and swap count accordingly. The following two new fields: numa_task_migrated numa_task_swapped will be displayed in both /sys/fs/cgroup/{GROUP}/memory.stat and /proc/{PID}/sched Previous RFC version can be found here: https://lore.kernel.org/lkml/1847c5ef828ad4835a35e3a54b88d2e13bce0eea.1740483690.git.yu.c.chen@intel.com/ Signed-off-by: Chen Yu <yu.c.chen@intel.com> --- RFC->v1: Rename the nr_numa_task_migrated to numa_task_migrated, and nr_numa_task_swapped numa_task_swapped in /proc/{PID}/sched, so both cgroup's memory.stat and task's sched have the same field name. --- include/linux/sched.h | 4 ++++ include/linux/vm_event_item.h | 2 ++ kernel/sched/core.c | 10 ++++++++-- kernel/sched/debug.c | 4 ++++ mm/memcontrol.c | 2 ++ mm/vmstat.c | 2 ++ 6 files changed, 22 insertions(+), 2 deletions(-)