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 DVFS behavior when compared to running the same workload in the host. Effectively, no EAS 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. We have been iterating over different options for communicating between guest and host, ways of applying the information coming from the guest/host, etc to figure out the best performance and power improvements we could get. The patch series in its current state is NOT meant for landing in the upstream kernel. We are sending this patch series to share the current progress and data we have so far. The patch series is meant to be easy to cherry-pick and test on various devices to see what performance and power benefits this might give for others. With this series, a workload running in a VM gets the same task placement and DVFS treatment as it would when running in the host. As expected, we see significant performance improvement and better performance/power ratio. If anyone else wants to try this out for your VM workloads and report findings, that'd be very much appreciated. The idea is to improve VM CPUfreq/sched behavior by: - Having guest kernel to do accurate load tracking by taking host CPU arch/type and frequency into account. - Sharing vCPU run queue utilization information with the host so that the host can do proper frequency scaling and task placement on the host side. Results: ======== As of right now, the best results have been with using hypercalls (see more below first) to communicate between host and guest and treating the vCPU run queue util similar to util_est on the host side vCPU thread. So that's what this patch series does. Let's look at the results for this series first and then look at the other options we are trying/tried out: Use cases running Android inside a VM on a Chromebook: ====================================================== PCMark (Emulates real world usecases) Higher is better +-------------------+----------+------------+--------+ | Test Case (score) | Baseline | Util_guest | %delta | +-------------------+----------+------------+--------+ | Weighted Total | 6136 | 7274 | +19% | +-------------------+----------+------------+--------+ | Web Browsing | 5558 | 6273 | +13% | +-------------------+----------+------------+--------+ | Video Editing | 4921 | 5221 | +6% | +-------------------+----------+------------+--------+ | Writing | 6864 | 8825 | +29% | +-------------------+----------+------------+--------+ | Photo Editing | 7983 | 11593 | +45% | +-------------------+----------+------------+--------+ | Data Manipulation | 5814 | 6081 | +5% | +-------------------+----------+------------+--------+ PCMark Performance/mAh Higher is better +-----------+----------+------------+--------+ | | Baseline | Util_guest | %delta | +-----------+----------+------------+--------+ | Score/mAh | 79 | 88 | +11% | +-----------+----------+------------+--------+ Roblox Higher is better +-----+----------+------------+--------+ | | Baseline | Util_guest | %delta | +-----+----------+------------+--------+ | FPS | 18.25 | 28.66 | +57% | +-----+----------+------------+--------+ Roblox FPS/mAh Higher is better +-----+----------+------------+--------+ | | Baseline | Util_guest | %delta | +-----+----------+------------+--------+ | FPS | 0.15 | 0.19 | +26% | +-----+----------+------------+--------+ Use cases running a minimal system inside a VM on a Pixel 6: ============================================================ FIO Higher is better +----------------------+----------+------------+--------+ | Test Case (avg MB/s) | Baseline | Util_guest | %delta | +----------------------+----------+------------+--------+ | Seq Write | 9.27 | 12.6 | +36% | +----------------------+----------+------------+--------+ | Rand Write | 9.34 | 11.9 | +27% | +----------------------+----------+------------+--------+ | Seq Read | 106 | 124 | +17% | +----------------------+----------+------------+--------+ | Rand Read | 33.6 | 35 | +4% | +----------------------+----------+------------+--------+ CPU-based ML Inference Benchmark Lower is better +-------------------------+----------+------------+--------+ | Test Case (ms) | Baseline | Util_guest | %delta | +-------------------------+----------+------------+--------+ | Cached Sample Inference | 2.57 | 1.75 | -32% | +-------------------------+----------+------------+--------+ | Small Sample Inference | 6.8 | 5.57 | -18% | +-------------------------+----------+------------+--------+ | Large Sample Inference | 31.2 | 26.58 | -15% | +-------------------------+----------+------------+--------+ These patches expect the host to: - Affine vCPUs to specific clusters. - Set vCPU capacity to match the host CPU they are running on. To make this easy to do/try out, we have put up patches[4][5] to do this on CrosVM. Once you pick up those patches, you can use options "--host-cpu-topology" and "--virt-cpufreq" to achieve the above. The patch series can be broken into: Patch 1: Add util_guest as an additional PELT signal for host vCPU threads Patch 2: Hypercall for guest to get current pCPU's frequency Patch 3: Send vCPU run queue util to host and apply as util_guest Patch 4: Query pCPU freq table from guest (we'll move this to DT in the future) Patch 5: Virtual cpufreq driver that uses the hypercalls to send util to host and implement frequency invariance in the guest. Alternative we have implemented and profiled: ============================================= util_guest vs uclamp_min ======================== One suggestion at LPC was to use uclamp_min to apply the util info coming from the guest. As we suspected, it doesn't perform as well because uclamp_min is not additive, whereas the actual workload on the host CPU due to the vCPU is additive to the existing workloads on the host. Uclamp_min also has the undesirable side-effect of threads forked from the vCPU thread inheriting whatever uclamp_min value the vCPU thread had and then getting stuck with that uclamp_min value. Below are some additional benchmark results comparing the uclamp_min prototype (listed as Uclamp) using the same test environment as before (including hypercalls). As before, %delta is always comparing to baseline. PCMark Higher is better +-------------------+----------+------------+--------+--------+--------+ | Test Case (score) | Baseline | Util_guest | %delta | Uclamp | %delta | +-------------------+----------+------------+--------+--------+--------+ | Weighted Total | 6136 | 7274 | +19% | 6848 | +12% | +-------------------+----------+------------+--------+--------+--------+ | Web Browsing | 5558 | 6273 | +13% | 6050 | +9% | +-------------------+----------+------------+--------+--------+--------+ | Video Editing | 4921 | 5221 | +6% | 5091 | +3% | +-------------------+----------+------------+--------+--------+--------+ | Writing | 6864 | 8825 | +29% | 8523 | +24% | +-------------------+----------+------------+--------+--------+--------+ | Photo Editing | 7983 | 11593 | +45% | 9865 | +24% | +-------------------+----------+------------+--------+--------+--------+ | Data Manipulation | 5814 | 6081 | +5% | 5836 | 0% | +-------------------+----------+------------+--------+--------+--------+ PCMark Performance/mAh Higher is better +-----------+----------+------------+--------+--------+--------+ | | Baseline | Util_guest | %delta | Uclamp | %delta | +-----------+----------+------------+--------+--------+--------+ | Score/mAh | 79 | 88 | +11% | 83 | +7% | +-----------+----------+------------+--------+--------+--------+ Hypercalls vs MMIO: =================== We realize that hypercalls are not the recommended choice for this and we have no attachment to any communication method as long as it gives good results. We started off with hypercalls to see what is the best we could achieve if we didn't have to context switch into host side userspace. To see the impact of switching from hypercalls to MMIO, we kept util_guest and only switched from hypercall to MMIO. So in the results below: - Hypercall = hypercall + util_guest - MMIO = MMIO + util_guest As before, %delta is always comparing to baseline. PCMark Higher is better +-------------------+----------+------------+--------+-------+--------+ | Test Case (score) | Baseline | Hypercall | %delta | MMIO | %delta | +-------------------+----------+------------+--------+-------+--------+ | Weighted Total | 6136 | 7274 | +19% | 6867 | +12% | +-------------------+----------+------------+--------+-------+--------+ | Web Browsing | 5558 | 6273 | +13% | 6035 | +9% | +-------------------+----------+------------+--------+-------+--------+ | Video Editing | 4921 | 5221 | +6% | 5167 | +5% | +-------------------+----------+------------+--------+-------+--------+ | Writing | 6864 | 8825 | +29% | 8529 | +24% | +-------------------+----------+------------+--------+-------+--------+ | Photo Editing | 7983 | 11593 | +45% | 10812 | +35% | +-------------------+----------+------------+--------+-------+--------+ | Data Manipulation | 5814 | 6081 | +5% | 5327 | -8% | +-------------------+----------+------------+--------+-------+--------+ PCMark Performance/mAh Higher is better +-----------+----------+-----------+--------+------+--------+ | | Baseline | Hypercall | %delta | MMIO | %delta | +-----------+----------+-----------+--------+------+--------+ | Score/mAh | 79 | 88 | +11% | 83 | +7% | +-----------+----------+-----------+--------+------+--------+ Roblox Higher is better +-----+----------+------------+--------+-------+--------+ | | Baseline | Hypercall | %delta | MMIO | %delta | +-----+----------+------------+--------+-------+--------+ | FPS | 18.25 | 28.66 | +57% | 24.06 | +32% | +-----+----------+------------+--------+-------+--------+ Roblox Frames/mAh Higher is better +------------+----------+------------+--------+--------+--------+ | | Baseline | Hypercall | %delta | MMIO | %delta | +------------+----------+------------+--------+--------+--------+ | Frames/mAh | 91.25 | 114.64 | +26% | 103.11 | +13% | +------------+----------+------------+--------+--------+--------+ Next steps: =========== We are continuing to look into communication mechanisms other than hypercalls that are just as/more efficient and avoid switching into the VMM userspace. Any inputs in this regard are greatly appreciated. Thanks, David & Saravana [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://chromium-review.googlesource.com/c/crosvm/crosvm/+/4208668 [5] - https://chromium-review.googlesource.com/c/crosvm/crosvm/+/4288027 David Dai (6): sched/fair: Add util_guest for tasks kvm: arm64: Add support for get_cur_cpufreq service kvm: arm64: Add support for util_hint service kvm: arm64: Add support for get_freqtbl service dt-bindings: cpufreq: add bindings for virtual kvm cpufreq cpufreq: add kvm-cpufreq driver .../bindings/cpufreq/cpufreq-virtual-kvm.yaml | 39 +++ Documentation/virt/kvm/api.rst | 28 ++ .../virt/kvm/arm/get_cur_cpufreq.rst | 21 ++ Documentation/virt/kvm/arm/get_freqtbl.rst | 23 ++ Documentation/virt/kvm/arm/index.rst | 3 + Documentation/virt/kvm/arm/util_hint.rst | 22 ++ arch/arm64/include/uapi/asm/kvm.h | 3 + arch/arm64/kvm/arm.c | 3 + arch/arm64/kvm/hypercalls.c | 60 +++++ drivers/cpufreq/Kconfig | 13 + drivers/cpufreq/Makefile | 1 + drivers/cpufreq/kvm-cpufreq.c | 245 ++++++++++++++++++ include/linux/arm-smccc.h | 21 ++ include/linux/sched.h | 12 + include/uapi/linux/kvm.h | 3 + kernel/sched/core.c | 24 +- kernel/sched/fair.c | 15 +- tools/arch/arm64/include/uapi/asm/kvm.h | 3 + 18 files changed, 536 insertions(+), 3 deletions(-) create mode 100644 Documentation/devicetree/bindings/cpufreq/cpufreq-virtual-kvm.yaml create mode 100644 Documentation/virt/kvm/arm/get_cur_cpufreq.rst create mode 100644 Documentation/virt/kvm/arm/get_freqtbl.rst create mode 100644 Documentation/virt/kvm/arm/util_hint.rst create mode 100644 drivers/cpufreq/kvm-cpufreq.c -- 2.40.0.348.gf938b09366-goog
