[v4,0/2] Improve VM CPUfreq and task placement behavior

Message

David Dai Nov. 11, 2023, 1:49 a.m. UTC

Hi,

This patch series is a continuation of the talk Saravana gave at LPC 2022
titled "CPUfreq/sched and VM guest workload problems" [1][2][3]. The gist
of the talk is that workloads running in a guest VM get terrible task
placement and CPUfreq behavior when compared to running the same workload
in the host. Effectively, no EAS(Energy Aware Scheduling) for threads
inside VMs. This would make power and performance terrible just by running
the workload in a VM even if we assume there is zero virtualization
overhead.

With this series, a workload running in a VM gets the same task placement
and CPUfreq behavior as it would when running in the host.

The idea is to improve VM CPUfreq/sched behavior by:
- Having guest kernel do accurate load tracking by taking host CPU
  arch/type and frequency into account.
- Sharing vCPU frequency requirements with the host so that the
  host can do proper frequency scaling and task placement on the host side.

Based on feedback from RFC v1 proposal[4], we've revised our
implementation to using MMIO reads and writes to pass information
from/to host instead of using hypercalls. In our example, the
VMM(Virtual Machine Manager) translates the frequency requests into
Uclamp_min and applies it to the vCPU thread as a hint to the host
kernel.

To achieve the results below, configure the host to:
- Affine vCPUs to specific clusters.
- Set vCPU capacity to match the host CPU they are running on.

To make it easy for folks to try this out with CrosVM, we have put up
userspace patches[5][6]. With those patches, you can configure CrosVM
correctly by adding the options "--host-cpu-topology" and "--virt-cpufreq".

Results:
========

Here are some side-by-side comparisons of RFC v1 proposal vs the current
patch series and are labelled as follows.

- (RFC v1) UtilHyp = hypercall + util_guest
- (current) UClampMMIO = MMIO + UClamp_min

Use cases running a minimal system inside a VM on a Pixel 6:
============================================================

FIO
Higher is better
+-------------------+----------+---------+--------+------------+--------+
| Usecase(avg MB/s) | Baseline | UtilHyp | %delta | UClampMMIO | %delta |
+-------------------+----------+---------+--------+------------+--------+
| Seq Write         |     13.3 |    16.4 |   +23% |       13.4 |    +1% |
+-------------------+----------+---------+--------+------------+--------+
| Rand Write        |     11.2 |    12.9 |   +15% |       11.2 |     0% |
+-------------------+----------+---------+--------+------------+--------+
| Seq Read          |      100 |     168 |   +68% |        136 |   +36% |
+-------------------+----------+---------+--------+------------+--------+
| Rand Read         |     20.5 |    35.6 |   +74% |       29.5 |   +44% |
+-------------------+----------+---------+--------+------------+--------+

CPU-based ML Inference Benchmark
Lower is better
+----------------+----------+------------+--------+------------+--------+
| Test Case (ms) | Baseline | UtilHyp    | %delta | UClampMMIO | %delta |
+----------------+----------+------------+--------+------------+--------+
| Cached Sample  |          |            |        |            |        |
| Inference      |     3.40 |       2.37 |   -30% |       2.97 |   -13% |
+----------------+----------+------------+--------+------------+--------+
| Small Sample   |          |            |        |            |        |
| Inference      |     9.87 |       6.78 |   -31% |       7.92 |   -20% |
+----------------+----------+------------+--------+------------+--------+
| Large Sample   |          |            |        |            |        |
| Inference      |    33.35 |      26.74 |   -20% |      31.48 |    -6% |
+----------------+----------+------------+--------+------------+--------+

Use cases running Android inside a VM on a Chromebook:
======================================================

PCMark (Emulates real world usecases)
Higher is better
+-------------------+----------+---------+--------+------------+--------+
| Test Case (score) | Baseline | UtilHyp | %delta | UClampMMIO | %delta |
+-------------------+----------+---------+--------+------------+--------+
| Weighted Total    |     5970 |    7162 |   +20% |       6782 |   +14% |
+-------------------+----------+---------+--------+------------+--------+
| Web Browsing      |     5558 |    5877 |    +6% |       5729 |    +3% |
+-------------------+----------+---------+--------+------------+--------+
| Video Editing     |     4921 |    5140 |    +4% |       5079 |    +3% |
+-------------------+----------+---------+--------+------------+--------+
| Writing           |     6864 |    9111 |   +33% |       8171 |   +10% |
+-------------------+----------+---------+--------+------------+--------+
| Photo Editing     |     7983 |   11349 |   +42% |      10313 |   +29% |
+-------------------+----------+---------+--------+------------+--------+
| Data Manipulation |     5814 |    6051 |    +4% |       6051 |    +1% |
+-------------------+----------+---------+--------+------------+--------+

PCMark Performance/mAh
Higher is better
+-------------------+----------+---------+--------+------------+--------+
|                   | Baseline | UtilHyp | %delta | UClampMMIO | %delta |
+-------------------+----------+---------+--------+------------+--------+
| Score/mAh         |       85 |     102 |   +20% |         94 |    10% |
+-------------------+----------+---------+--------+------------+--------+

Roblox
Higher is better
+-------------------+----------+---------+--------+------------+--------+
|                   | Baseline | UtilHyp | %delta | UClampMMIO | %delta |
+-------------------+----------+---------+--------+------------+--------+
| FPS               |    20.88 |   25.64 |   +23% |      24.05 |   +15% |
+-------------------+----------+---------+--------+------------+--------+

Roblox Frames/mAh
Higher is better
+-------------------+----------+---------+--------+------------+--------+
|                   | Baseline | UtilHyp | %delta | UClampMMIO | %delta |
+-------------------+----------+---------+--------+------------+--------+
| Frames/mAh        |    85.29 |  102.31 |   +20% |     94.20  |    10% |
+-------------------+----------+---------+--------+------------+--------+

We've simplified our implementation based on community feedback to make
it less intrusive and to use a more generic MMIO interface for
communication with the host. The results show that the current design
still has tangible improvements over baseline. We'll continue looking
into ways to reduce the overhead of the MMIO read/writes and submit
separate and generic patches for that if we find any good optimizations.

Thanks,
David & Saravana

Cc: Saravana Kannan <[e-mail]>
Cc: Quentin Perret <[e-mail]>
Cc: Masami Hiramatsu <[e-mail]>
Cc: Will Deacon <[e-mail]>
Cc: Peter Zijlstra <[e-mail]>
Cc: Vincent Guittot <[e-mail]>
Cc: Marc Zyngier <[e-mail]>
Cc: Oliver Upton <[e-mail]>
Cc: Dietmar Eggemann <[e-mail]>
Cc: Pavan Kondeti <[e-mail]>
Cc: Gupta Pankaj <[e-mail]>
Cc: Mel Gorman <[e-mail]>

v3 -> v4:
-Fixed dt-binding formatting issues
-Added additional dt-binding descriptions for “HW interfaces”
-Changed dt-binding to “qemu,virtual-cpufreq”
-Fixed Kconfig formatting issues
-Removed frequency downscaling when requesting frequency updates
-Removed ops and cpufreq driver data
-Added check to limit freq_scale to 1024
-Added KHZ in the register offset naming
-Added comments to explain FIE and not allowing dvfs_possible_from_any_cpu

v2 -> v3:
- Dropped patches adding new hypercalls
- Dropped patch adding util_guest in sched/fair
- Cpufreq driver now populates frequency using opp bindings
- Removed transition_delay_us=1 cpufreq setting as it was configured too
  agressively and resulted in poor I/O performance
- Modified guest cpufreq driver to read/write MMIO regions instead of
  using hypercalls to communicate with the host
- Modified guest cpufreq driver to pass frequency info instead of
  utilization of the current vCPU's runqueue which now takes
  iowait_boost into account from the schedutil governor
- Updated DT bindings for a virtual CPU frequency device
Userspace changes:
- Updated CrosVM patches to emulate a virtual cpufreq device
- Updated to newer userspace binaries when collecting more recent
  benchmark data

v1 -> v2:
- No functional changes.
- Added description for EAS and removed DVFS in coverletter.
- Added a v2 tag to the subject.
- Fixed up the inconsistent "units" between tables.
- Made sure everyone is To/Cc-ed for all the patches in the series.

[1] - https://lpc.events/event/16/contributions/1195/
[2] - https://lpc.events/event/16/contributions/1195/attachments/970/1893/LPC%202022%20-%20VM%20DVFS.pdf
[3] - https://www.youtube.com/watch?v=hIg_5bg6opU
[4] - https://lore.kernel.org/all/[e-mail]/
[5] - https://chromium-review.googlesource.com/c/crosvm/crosvm/+/4208668
[6] - https://chromium-review.googlesource.com/q/topic:%22virtcpufreq-v4%22

David Dai (2):
  dt-bindings: cpufreq: add virtual cpufreq device
  cpufreq: add virtual-cpufreq driver

 .../cpufreq/qemu,cpufreq-virtual.yaml         |  99 +++++++++
 drivers/cpufreq/Kconfig                       |  15 ++
 drivers/cpufreq/Makefile                      |   1 +
 drivers/cpufreq/virtual-cpufreq.c             | 201 ++++++++++++++++++
 include/linux/arch_topology.h                 |   1 +
 5 files changed, 317 insertions(+)
 create mode 100644 Documentation/devicetree/bindings/cpufreq/qemu,cpufreq-virtual.yaml
 create mode 100644 drivers/cpufreq/virtual-cpufreq.c

Comments

Hongyan Xia Nov. 13, 2023, 12:20 p.m. UTC | #1

Hi David,

On 11/11/2023 01:49, David Dai wrote:
> Hi,
> 
> This patch series is a continuation of the talk Saravana gave at LPC 2022
> titled "CPUfreq/sched and VM guest workload problems" [1][2][3]. The gist
> of the talk is that workloads running in a guest VM get terrible task
> placement and CPUfreq behavior when compared to running the same workload
> in the host. Effectively, no EAS(Energy Aware Scheduling) for threads
> inside VMs. This would make power and performance terrible just by running
> the workload in a VM even if we assume there is zero virtualization
> overhead.
> 
> With this series, a workload running in a VM gets the same task placement
> and CPUfreq behavior as it would when running in the host.
> 
> The idea is to improve VM CPUfreq/sched behavior by:
> - Having guest kernel do accurate load tracking by taking host CPU
>    arch/type and frequency into account.
> - Sharing vCPU frequency requirements with the host so that the
>    host can do proper frequency scaling and task placement on the host side.
> 
> Based on feedback from RFC v1 proposal[4], we've revised our
> implementation to using MMIO reads and writes to pass information
> from/to host instead of using hypercalls. In our example, the
> VMM(Virtual Machine Manager) translates the frequency requests into
> Uclamp_min and applies it to the vCPU thread as a hint to the host
> kernel.

Sorry for not noticing this series until now.

The problem you are having with uclamp is actually the same as what
I'm tackling right now. Basically my conclusion so far is that uclamp
max aggregation faces quite many problems, which can be easily solved by
sum aggregation (summing up the clamped utilization values instead of
applying the max uclamp value to the whole rq):

https://lore.kernel.org/all/[e-mail]/

What you described as util_guest sounds to me as exactly what uclamp_min
under sum aggregation does. I'm really tempted to ask you to apply my
series and see if the new uclamp_min does what you want, instead of
introducing a new util_guest signal. If you have no time for this I can
try to replicate your setup and do the experiments myself.

Also, my knowledge with KVM is limited. May I know where the vCPU fork
happens? Can't you just set the p->sched_reset_on_fork flag on fork to
not carry forward the uclamp values?

> 
> [...]
Hongyan

Marc Zyngier Nov. 13, 2023, 12:26 p.m. UTC | #2

On Mon, 13 Nov 2023 12:20:29 +0000,
Hongyan Xia <[e-mail]> wrote:
> 
> Also, my knowledge with KVM is limited. May I know where the vCPU fork
> happens? Can't you just set the p->sched_reset_on_fork flag on fork to
> not carry forward the uclamp values?

There is no "vCPU fork". The vcpu is a VMM (qemu, crosvm, kvmtool)
thread, nothing else.

	M.