On Wed, Mar 01, 2023 at 05:32:14PM +0100, Antonio Paolillo wrote:
Dear all,
I want to provide some support to Hernan regarding performance claims.
I used lock_torture to evaluate the different proposed patches on two different server machines:
- a Huawei TaiShan 200 (Model 2280) rack server that has 128 GB of RAM and 2x Kunpeng 920-4826 processors, a HiSilicon chip with 48 ARMv8.2 64-bit cores totaling 96 cores (no SMT) [1, 2], denoted as taishan200-96c;
- a GIGABYTE R182-Z91-00 rack server that has 128 GB of RAM and 2x EPYC 7352 processors, an AMD chip with 24 x86_64 cores, totaling 48 cores (96 CPUs when counting hyperthreading) [3, 4], denoted as gigabyte-96c.
I ran the evaluation on a Ubuntu 22.04 distro, with custom kernels based on v6.2-rc6 (6d796c50f84ca79f1722bb131799e5a5710c4700). The different kernels are combination of patches:
- (0) Stock kernel;
- (1) With relaxed set owner barrier (as discussed in [5] and questioned by Peter, the barrier seems not to be needed);
- (2) With READ_ONCE(), as originally proposed in this thread;
- (3) With atomic_long_or() as proposed by Peter;
- (4) With relaxed set owner barrier and READ_ONCE();
- (5) With relaxed set owner barrier and atomic_long_or().
I ran lock_torture several times, exploring the following parameter space:
- torture_type="rtmutex_lock",
- nwriters_stress=[1, 2, 3, 4, 8, 16, 32, 64, 95],
- stat_interval=4,
- stutter=0,
- shuffle_interval=0.
For each value of "nwriters_stress", I ran the configuration 5 times.
By feeding the lock_torture kthread pids to "taskset -p", I overruled the scheduling such that the distribution of kthreads to CPUs is fixed. I also disabled "irq balance" and "numa balance" daemons, fixed the frequency to 1.5GHz using the "userspace" cpufreq governor and isolated all the cores used (using isolcpus=1-95 at boot-time) to avoid any source of interference.
As a warm-up phase, I ignored the first reported results and only considered the latest 60 seconds of execution (after all kthreads migrated to their final CPU). The reported throughput is computed by dividing the reported number of operations by the duration of the measurement for each dot (60 seconds), so higher is better.
Here follows the results on taishan200-96c (the 'rel' column is the mean relative to the mean of the stock kernel, in percent, and each mean is the average over 5 independent runs):
Kernel: k0-stock-6.2.0-rc6 k1-rmacq k2-readonce k3-alongor k4-rmacq+readonce k5-rmacq+alongor Statistic (kops/s): mean std mean std rel mean std rel mean std rel mean std rel mean std rel nwriters_stress: 1 899.91 24.95 880.10 29.62 -2% 871.57 44.27 -3% 888.65 37.90 -1% 898.63 29.82 -0% 889.83 25.64 -1% 2 359.30 25.92 416.83 32.77 +16% 360.65 28.32 +0% 404.79 42.64 +13% 380.65 21.29 +6% 404.37 23.27 +13% 3 314.97 24.32 308.41 9.68 -2% 315.00 9.97 +0% 313.86 13.47 -0% 313.47 4.01 -0% 322.77 20.82 +2% 4 328.02 15.09 330.65 29.33 +1% 314.83 24.28 -4% 305.71 12.72 -7% 322.95 10.39 -2% 343.32 13.73 +5% 8 292.16 22.03 288.85 10.50 -1% 288.28 18.84 -1% 285.42 24.58 -2% 310.23 26.08 +6% 285.67 20.03 -2% 16 297.03 26.89 281.89 29.22 -5% 265.19 33.73 -11% 279.02 22.43 -6% 284.40 36.21 -4% 285.21 36.33 -4% 32 187.36 28.59 175.71 19.77 -6% 186.44 48.15 -0% 206.59 14.11 +10% 174.08 24.30 -7% 185.80 45.12 -1% 64 148.13 48.65 172.48 34.29 +16% 154.59 47.05 +4% 164.22 29.81 +11% 142.13 47.40 -4% 136.39 29.95 -8% 95 174.35 57.89 148.59 38.03 -15% 156.85 43.64 -10% 132.92 32.35 -24% 126.44 28.24 -27% 146.82 60.04 -16%
And the results on gigabyte-96c:
Kernel: k0-stock-6.2.0-rc6 k1-rmacq k2-readonce k3-alongor k4-rmacq+readonce k5-rmacq+alongor Statistic (kops/s): mean std mean std rel mean std rel mean std rel mean std rel mean std rel nwriters_stress: 1 713.72 25.68 707.32 17.73 -1% 718.81 12.63 +1% 712.80 13.57 -0% 709.17 14.10 -1% 730.33 9.14 +2% 2 376.25 8.19 400.09 16.24 +6% 396.71 26.09 +5% 412.61 17.80 +10% 396.48 7.02 +5% 409.90 14.61 +9% 3 415.07 16.83 410.19 19.82 -1% 423.39 9.68 +2% 417.28 10.23 +1% 424.94 17.48 +2% 422.92 11.75 +2% 4 286.77 26.63 285.13 6.80 -1% 297.33 23.62 +4% 296.49 16.60 +3% 303.99 30.38 +6% 296.93 9.90 +4% 8 296.56 20.45 308.97 12.53 +4% 305.49 19.91 +3% 294.24 17.24 -1% 294.71 24.03 -1% 294.09 25.20 -1% 16 257.34 33.94 266.03 29.60 +3% 270.72 35.22 +5% 252.28 50.16 -2% 263.83 45.84 +3% 247.42 41.01 -4% 32 278.78 51.45 215.35 68.40 -23% 259.77 87.44 -7% 217.26 79.67 -22% 201.23 70.46 -28% 282.47 116.65 +1% 64 75.82 64.87 194.52 137.19 +157% 35.57 12.14 -53% 74.24 72.04 -2% 71.29 45.55 -6% 77.93 43.57 +3% 95 60.37 68.13 198.38 116.93 +229% 43.12 17.60 -29% 58.80 36.47 -3% 57.78 63.00 -4% 61.33 71.18 +2%
We can safely conclude that the patches do not significatively affect the throughput of the lock_torture benchmark for rtmutex_lock. The values for nwriters_stress>=64 can safely be ignored as they are too spread.
Just so you know, locktorture is intended to be a stress test rather than a performance benchmark. Hugo Guiroux's dissertation gives a much better locking performance methodology:
https://hugoguiroux.github.io/assets/these.pdf
Thanx, Paul
Please notice that I pushed a landing page [6] with results in HTML that may be more convenient to browse together with interactive charts.
Cheers,
Antonio
[1] https://e.huawei.com/uk/products/servers/taishan-server/taishan-2280-v2 [2] https://en.wikichip.org/wiki/hisilicon/kunpeng/920-4826 [3] https://www.gigabyte.com/Rack-Server/R182-Z91-rev-100 [4] https://www.amd.com/en/products/cpu/amd-epyc-7352 [5] https://lkml.org/lkml/2023/1/22/160 [6] https://antonio.paolillo.be/public/rtlocks-locktorture-patches.html