| Message ID | 20210404083354.23060-1-psampat@linux.ibm.com (mailing list archive) |
|---|---|
| Headers | show |
| Series | CPU-Idle latency selftest framework | expand |
Hi Pratik, I tried V3 on a Intel i5-10600K processor with 6 cores and 12 CPUs. The core to cpu mappings are: core 0 has cpus 0 and 6 core 1 has cpus 1 and 7 core 2 has cpus 2 and 8 core 3 has cpus 3 and 9 core 4 has cpus 4 and 10 core 5 has cpus 5 and 11 By default, it will test CPUs 0,2,4,6,10 on cores 0,2,4,0,2,4. wouldn't it make more sense to test each core once? With the source CPU always 0, I think the results from the results from the destination CPUs 0 and 6, on core 0 bias the results, at least in the deeper idle states. They don't make much difference in the shallow states. Myself, I wouldn't include them in the results. Example, where I used the -v option for all CPUs: --IPI Latency Test--- --Baseline IPI Latency measurement: CPU Busy-- SRC_CPU DEST_CPU IPI_Latency(ns) 0 0 101 0 1 790 0 2 609 0 3 595 0 4 737 0 5 759 0 6 780 0 7 741 0 8 574 0 9 681 0 10 527 0 11 552 Baseline Avg IPI latency(ns): 620 <<<< suggest 656 here ---Enabling state: 0--- SRC_CPU DEST_CPU IPI_Latency(ns) 0 0 76 0 1 471 0 2 420 0 3 462 0 4 454 0 5 468 0 6 453 0 7 473 0 8 380 0 9 483 0 10 492 0 11 454 Expected IPI latency(ns): 0 Observed Avg IPI latency(ns) - State 0: 423 <<<<< suggest 456 here ---Enabling state: 1--- SRC_CPU DEST_CPU IPI_Latency(ns) 0 0 112 0 1 866 0 2 663 0 3 851 0 4 1090 0 5 1314 0 6 1941 0 7 1458 0 8 687 0 9 802 0 10 1041 0 11 1284 Expected IPI latency(ns): 1000 Observed Avg IPI latency(ns) - State 1: 1009 <<<< suggest 1006 here ---Enabling state: 2--- SRC_CPU DEST_CPU IPI_Latency(ns) 0 0 75 0 1 16362 0 2 16785 0 3 19650 0 4 17356 0 5 17606 0 6 2217 0 7 17958 0 8 17332 0 9 16615 0 10 17382 0 11 17423 Expected IPI latency(ns): 120000 Observed Avg IPI latency(ns) - State 2: 14730 <<<< suggest 17447 here ---Enabling state: 3--- SRC_CPU DEST_CPU IPI_Latency(ns) 0 0 103 0 1 17416 0 2 17961 0 3 16651 0 4 17867 0 5 17726 0 6 2178 0 7 16620 0 8 20951 0 9 16567 0 10 17131 0 11 17563 Expected IPI latency(ns): 1034000 Observed Avg IPI latency(ns) - State 3: 14894 <<<< suggest 17645 here Hope this helps. ... Doug
Hello Doug, On 09/04/21 10:53 am, Doug Smythies wrote: > Hi Pratik, > > I tried V3 on a Intel i5-10600K processor with 6 cores and 12 CPUs. > The core to cpu mappings are: > core 0 has cpus 0 and 6 > core 1 has cpus 1 and 7 > core 2 has cpus 2 and 8 > core 3 has cpus 3 and 9 > core 4 has cpus 4 and 10 > core 5 has cpus 5 and 11 > > By default, it will test CPUs 0,2,4,6,10 on cores 0,2,4,0,2,4. > wouldn't it make more sense to test each core once? Ideally it would be better to run on all the CPUs, however on larger systems that I'm testing on with hundreds of cores and a high a thread count, the execution time increases while not particularly bringing any additional information to the table. That is why it made sense only run on one of the threads of each core to make the experiment faster while preserving accuracy. To handle various thread topologies it maybe worthwhile if we parse /sys/devices/system/cpu/cpuX/topology/thread_siblings_list for each core and use this information to run only once per physical core, rather than assuming the topology. What are your thoughts on a mechanism like this? > With the source CPU always 0, I think the results from the results > from the destination CPUs 0 and 6, on core 0 bias the results, at > least in the deeper idle states. They don't make much difference in > the shallow states. Myself, I wouldn't include them in the results. I agree, CPU0->CPU0 same core interaction is causing a bias. I could omit that observation while computing the average. In the verbose mode I'll omit all the threads of CPU0 and in the default (quick) mode just CPU0's latency can be omitted while computing average. Thank you, Pratik > Example, where I used the -v option for all CPUs: > > --IPI Latency Test--- > --Baseline IPI Latency measurement: CPU Busy-- > SRC_CPU DEST_CPU IPI_Latency(ns) > 0 0 101 > 0 1 790 > 0 2 609 > 0 3 595 > 0 4 737 > 0 5 759 > 0 6 780 > 0 7 741 > 0 8 574 > 0 9 681 > 0 10 527 > 0 11 552 > Baseline Avg IPI latency(ns): 620 <<<< suggest 656 here > ---Enabling state: 0--- > SRC_CPU DEST_CPU IPI_Latency(ns) > 0 0 76 > 0 1 471 > 0 2 420 > 0 3 462 > 0 4 454 > 0 5 468 > 0 6 453 > 0 7 473 > 0 8 380 > 0 9 483 > 0 10 492 > 0 11 454 > Expected IPI latency(ns): 0 > Observed Avg IPI latency(ns) - State 0: 423 <<<<< suggest 456 here > ---Enabling state: 1--- > SRC_CPU DEST_CPU IPI_Latency(ns) > 0 0 112 > 0 1 866 > 0 2 663 > 0 3 851 > 0 4 1090 > 0 5 1314 > 0 6 1941 > 0 7 1458 > 0 8 687 > 0 9 802 > 0 10 1041 > 0 11 1284 > Expected IPI latency(ns): 1000 > Observed Avg IPI latency(ns) - State 1: 1009 <<<< suggest 1006 here > ---Enabling state: 2--- > SRC_CPU DEST_CPU IPI_Latency(ns) > 0 0 75 > 0 1 16362 > 0 2 16785 > 0 3 19650 > 0 4 17356 > 0 5 17606 > 0 6 2217 > 0 7 17958 > 0 8 17332 > 0 9 16615 > 0 10 17382 > 0 11 17423 > Expected IPI latency(ns): 120000 > Observed Avg IPI latency(ns) - State 2: 14730 <<<< suggest 17447 here > ---Enabling state: 3--- > SRC_CPU DEST_CPU IPI_Latency(ns) > 0 0 103 > 0 1 17416 > 0 2 17961 > 0 3 16651 > 0 4 17867 > 0 5 17726 > 0 6 2178 > 0 7 16620 > 0 8 20951 > 0 9 16567 > 0 10 17131 > 0 11 17563 > Expected IPI latency(ns): 1034000 > Observed Avg IPI latency(ns) - State 3: 14894 <<<< suggest 17645 here > > Hope this helps. > > ... Doug
On Fri, Apr 9, 2021 at 12:43 AM Pratik Sampat <psampat@linux.ibm.com> wrote: > On 09/04/21 10:53 am, Doug Smythies wrote: > > I tried V3 on a Intel i5-10600K processor with 6 cores and 12 CPUs. > > The core to cpu mappings are: > > core 0 has cpus 0 and 6 > > core 1 has cpus 1 and 7 > > core 2 has cpus 2 and 8 > > core 3 has cpus 3 and 9 > > core 4 has cpus 4 and 10 > > core 5 has cpus 5 and 11 > > > > By default, it will test CPUs 0,2,4,6,10 on cores 0,2,4,0,2,4. > > wouldn't it make more sense to test each core once? > > Ideally it would be better to run on all the CPUs, however on larger systems > that I'm testing on with hundreds of cores and a high a thread count, the > execution time increases while not particularly bringing any additional > information to the table. > > That is why it made sense only run on one of the threads of each core to make > the experiment faster while preserving accuracy. > > To handle various thread topologies it maybe worthwhile if we parse > /sys/devices/system/cpu/cpuX/topology/thread_siblings_list for each core and > use this information to run only once per physical core, rather than > assuming the topology. > > What are your thoughts on a mechanism like this? Yes, seems like a good solution. ... Doug
Changelog RFC v2-->v3 Based on comments by Doug Smythies, 1. Changed commit log to reflect the test must be run as super user. 2. Added a comment specifying a method to run the test bash script without recompiling. 3. Enable all the idle states after the experiments are completed so that the system is in a coherent state after the tests have run 4. Correct the return status of a CPU that cannot be off-lined. RFC v2: https://lkml.org/lkml/2021/4/1/615 --- A kernel module + userspace driver to estimate the wakeup latency caused by going into stop states. The motivation behind this program is to find significant deviations behind advertised latency and residency values. The patchset measures latencies for two kinds of events. IPIs and Timers As this is a software-only mechanism, there will additional latencies of the kernel-firmware-hardware interactions. To account for that, the program also measures a baseline latency on a 100 percent loaded CPU and the latencies achieved must be in view relative to that. To achieve this, we introduce a kernel module and expose its control knobs through the debugfs interface that the selftests can engage with. The kernel module provides the following interfaces within /sys/kernel/debug/latency_test/ for, IPI test: ipi_cpu_dest = Destination CPU for the IPI ipi_cpu_src = Origin of the IPI ipi_latency_ns = Measured latency time in ns Timeout test: timeout_cpu_src = CPU on which the timer to be queued timeout_expected_ns = Timer duration timeout_diff_ns = Difference of actual duration vs expected timer Sample output on a POWER9 system is as follows: # --IPI Latency Test--- # Baseline Average IPI latency(ns): 3114 # Observed Average IPI latency(ns) - State0: 3265 # Observed Average IPI latency(ns) - State1: 3507 # Observed Average IPI latency(ns) - State2: 3739 # Observed Average IPI latency(ns) - State3: 3807 # Observed Average IPI latency(ns) - State4: 17070 # Observed Average IPI latency(ns) - State5: 1038174 # Observed Average IPI latency(ns) - State6: 1068784 # # --Timeout Latency Test-- # Baseline Average timeout diff(ns): 1420 # Observed Average timeout diff(ns) - State0: 1640 # Observed Average timeout diff(ns) - State1: 1764 # Observed Average timeout diff(ns) - State2: 1715 # Observed Average timeout diff(ns) - State3: 1845 # Observed Average timeout diff(ns) - State4: 16581 # Observed Average timeout diff(ns) - State5: 939977 # Observed Average timeout diff(ns) - State6: 1073024 Things to keep in mind: 1. This kernel module + bash driver does not guarantee idleness on a core when the IPI and the Timer is armed. It only invokes sleep and hopes that the core is idle once the IPI/Timer is invoked onto it. Hence this program must be run on a completely idle system for best results 2. Even on a completely idle system, there maybe book-keeping tasks or jitter tasks that can run on the core we want idle. This can create outliers in the latency measurement. Thankfully, these outliers should be large enough to easily weed them out. 3. A userspace only selftest variant was also sent out as RFC based on suggestions over the previous patchset to simply the kernel complexeity. However, a userspace only approach had more noise in the latency measurement due to userspace-kernel interactions which led to run to run variance and a lesser accurate test. Another downside of the nature of a userspace program is that it takes orders of magnitude longer to complete a full system test compared to the kernel framework. RFC patch: https://lkml.org/lkml/2020/9/2/356 4. For Intel Systems, the Timer based latencies don't exactly give out the measure of idle latencies. This is because of a hardware optimization mechanism that pre-arms a CPU when a timer is set to wakeup. That doesn't make this metric useless for Intel systems, it just means that is measuring IPI/Timer responding latency rather than idle wakeup latencies. (Source: https://lkml.org/lkml/2020/9/2/610) For solution to this problem, a hardware based latency analyzer is devised by Artem Bityutskiy from Intel. https://youtu.be/Opk92aQyvt0?t=8266 https://intel.github.io/wult/ Pratik Rajesh Sampat (2): cpuidle: Extract IPI based and timer based wakeup latency from idle states selftest/cpuidle: Add support for cpuidle latency measurement drivers/cpuidle/Makefile | 1 + drivers/cpuidle/test-cpuidle_latency.c | 157 ++++++++++ lib/Kconfig.debug | 10 + tools/testing/selftests/Makefile | 1 + tools/testing/selftests/cpuidle/Makefile | 6 + tools/testing/selftests/cpuidle/cpuidle.sh | 326 +++++++++++++++++++++ tools/testing/selftests/cpuidle/settings | 2 + 7 files changed, 503 insertions(+) create mode 100644 drivers/cpuidle/test-cpuidle_latency.c create mode 100644 tools/testing/selftests/cpuidle/Makefile create mode 100755 tools/testing/selftests/cpuidle/cpuidle.sh create mode 100644 tools/testing/selftests/cpuidle/settings