Message ID | 20230610183518.4061159-1-dedekind1@gmail.com (mailing list archive) |
---|---|
Headers | show |
Series | Sapphire Rapids C0.x idle states support | expand |
On Sat, Jun 10 2023 at 21:35, Artem Bityutskiy wrote: Something is wrong with your patch mail script. [PATCH v3 0/3] but then the actual patches are [PATCH v3 1/2] [PATCH v3 2/2] Something does not add up here. Thanks, tglx
From: Artem Bityutskiy <artem.bityutskiy@linux.intel.com> Background ---------- Idle states reduce power consumption when a CPU has no work to do. The most shallow CPU idle state is "POLL". It has lowest wake up latency, but saves little power. The next idle state on Intel platforms is "C1". It has has higher latency, but saves more power than "POLL". Sapphire Rapids Xeons add new C0.1 and C0.2 idle states which conceptually sit between "POLL" and "C1". These provide a very attractive midpoint: near-POLL wake-up latency and power consumption halfway between "POLL" and "C1". In other words, we expect all but the most latency-sensitive users to prefer these idle state over POLL. This patch-set enables C0.2 idle state support on Sapphire Rapids Xeon (later - SPR). The new idle state is added between POLL and C1. Patch-set overview ------------------ This patch-set is based on the "linux-next" branch of the "linux-pm" plus patches from Arjan van de Ven, submitted to linux-pm mailing list on Jun 5, 2023: * Cover letter: [PATCH 0/4 v2] Add support for running in VM guests to intel_idle * https://patchwork.kernel.org/project/linux-pm/patch/20230605154716.840930-2-arjan@linux.intel.com/ In other words, the base commit is 'e8195eaff86fd2ddb5f00646b5f76e40cd1164a8', then Arjan's patches should be applied, and then these patches on top. Patch #1 does not depend on Arjan's patches, but patch #2 requires the cleanups from Arjan's patch-set. Changelog --------- * v3 - Dropped patch 'x86/umwait: Increase tpause and umwait quanta' after, as suggested by Andy Lutomirski. - Followed Peter Zijlstra's suggestion and removed explicit 'umwait' deadline. Rely on the global implicit deadline instead. - Rebased on top of Arjan's patches. - C0.2 was tested in a VM by Arjan van de Ven. - Re-measured on 2S and 4S Sapphire Rapids Xeon. * v2 - Do not mix 'raw_local_irq_enable()' and 'local_irq_disable()'. I failed to directly verify it, but I believe it'll address the '.noinstr.text' warning. - Minor kerneldoc commentary fix. C0.2 vs POLL latency and power ------------------------------ I compared POLL to C0.2 using 'wult' tool (https://github.com/intel/wult), which measures idle states latency. * In "POLL" experiments, all C-states except for POLL were disabled. * In "C0.2" experiments, all C-states except for POLL and C0.2 were disabled. Here are the measurement results. The numbers are percent change from POLL to C0.2. -----------|-----------|----------|----------- Median IR | 99th % IR | AC Power | RAPL power -----------|-----------|----------|----------- 24% | 12% | -13% | -18% -----------------------|----------|----------- * IR stands for interrupt latency. The table provides the median and 99th percentile. Wult measures it as the delay between the moment a timer interrupt fires to the moment the CPU reaches the interrupt handler. * AC Power is the wall socket AC power. * RAPL power is the CPU package power, measured using the 'turbostat' tool. Hackbench measurements ---------------------- I ran the 'hackbench' benchmark using the following commands: # 4 groups, 200 threads hackbench -s 128 -l 100000000 -g4 -f 25 -P # 8 groups, 400 threads. hackbench -s 128 -l 100000000 -g8 -f 25 -P My SPR system has 224 CPUs, so the first command did not use all CPUs, the second command used all of them. However, in both cases CPU power reached TDP. I ran hackbench 5 times for every configuration and compared hackbench "score" averages. In case of 4 groups, C0.2 improved the score by about 4%, and in case of 8 groups by about 0.6%. Q&A --- 1. Can C0.2 be disabled? C0.2 can be disabled via sysfs and with the following kernel boot option: intel_idle.states_off=2 2. Why C0.2, not C0.1? I measured both C0.1 and C0.2, but did not notice a clear C0.1 advantage in terms of latency, but did notice that C0.2 saves more power. But if users want to try using C0.1 instead of C0.2, they can do this: echo 0 > /sys/devices/system/cpu/umwait_control/enable_c02 This will make sure that C0.2 requests from 'intel_idle' are automatically converted to C0.1 requests. 3. How did you verify that system enters C0.2? I used 'perf' to read the corresponding PMU counters: perf stat -e CPU_CLK_UNHALTED.C01,CPU_CLK_UNHALTED.C02,cycles -a sleep 1 4. Ho to change the global explicit 'umwait' deadline? Via '/sys/devices/system/cpu/umwait_control/max_time' Artem Bityutskiy (2): x86/mwait: Add support for idle via umwait intel_idle: add C0.2 state for Sapphire Rapids Xeon arch/x86/include/asm/mwait.h | 65 ++++++++++++++++++++++++++++++++++++ drivers/idle/intel_idle.c | 44 +++++++++++++++++++++++- 2 files changed, 108 insertions(+), 1 deletion(-)