[Impact]
On AMD Zen systems, memcpy() calls see a heavy performance regression in Focal and Groovy, due to the way __x86_non_temporal_threshold is calculated.
Before 'glibc-2.33~455', cache values were calculated taking into consideration the number of hardware threads in the CPU. On AMD Ryzen and EPYC systems, this can be counter-productive if the number of threads is high enough for the last-level caches to "overrun" each other and cause cache line flushes. The solution is to reduce the allocated size for these non_temporal stores, removing the number of threads from the equation.
[Test Plan]
Compile the test_memcpy.c that is attached to this bug report:
$ gcc -mtune=generic -march=x86-64 -g -O3 test_memcpy.c -o test_memcpy64
This should be run before and after installing the libc packages from proposed. On Ryzen and EPYC systems a substantial improvement should be seen and on other systems, no significant change should be seen.
[Where problems could occur]
Since we're messing with the cacheinfo for x86 in general, we need to be careful not to introduce further performance regressions on memory-heavy workloads. Even though initial results might reveal improvement on AMD Ryzen and EPYC hardware, we should also validate different configurations (e.g. Intel, different buffer sizes, etc) to make sure we won't hurt performance in other non-AMD environments.
[Other Info]
This issue has been fixed by the following upstream commit:
- d3c57027470b (Reversing calculation of __x86_shared_non_temporal_threshold)
$ git describe --contains d3c57027470b
glibc-2.33~455
$ rmadison glibc -s focal,focal-updates,groovy,groovy-proposed,hirsute
glibc | 2.31-0ubuntu9 | focal | source
glibc | 2.31-0ubuntu9.2 | focal-updates | source
glibc | 2.32-0ubuntu3 | groovy | source
glibc | 2.32-0ubuntu3.2 | groovy-proposed | source
glibc | 2.33-0ubuntu5 | hirsute | source
Affected releases include Ubuntu Focal and Groovy. Bionic is not affected, and releases starting with Hirsute already ship the upstream patch to fix this regression.
glibc exports this specific variable as a tunable, so we could also tweak it with the GLIBC_TUNABLES env var:
$ hyperfine -n clean-env 'lxc exec focal env ./test_memcpy64 32' -n tunables 'lxc exec focal env GLIBC_TUNABLES=glibc.cpu.x86_non_temporal_threshold=1024*1024*3*4 ./test_memcpy64 32'
Benchmark #1: clean-env
Time (mean ± σ): 2.529 s ± 0.061 s [User: 6.0 ms, System: 4.7 ms]
Range (min … max): 2.457 s … 2.615 s 10 runs
Benchmark #2: tunables
Time (mean ± σ): 1.427 s ± 0.030 s [User: 6.5 ms, System: 3.8 ms]
Range (min … max): 1.402 s … 1.482 s 10 runs
Summary
'tunables' ran
1.77 ± 0.06 times faster than 'clean-env'
This solution is not ideal, but it offers a secondary way of fixing the performance issues. However, the speed gains for memcpy() are noticeable enough that we should strongly consider changing the defaults in the Focal LTS release, so that it performs similarly to Bionic and future Ubuntu releases starting with Hirsute.
[old test case section]
Attached to this bug is a short C program that exercises memcpy() calls in buffers of variable length. This has been obtained from a similar bug report for Red Hat, and is publicly available at [0].
This test program was compiled with gcc 10.2.0, using the following flags:
$ gcc -mtune=generic -march=x86_64 -g -03 test_memcpy.c -o test_memcpy64
Tests were performed with the following criteria:
- use 32Mb buffers ("./test_memcpy64 32")
- benchmark with the hyperfine tool [1], as it calculates relevant statistics automatically
- benchmark with at least 10 runs in the same environment, to minimize variance
- measure on AMD Zen (3700X) and on Intel Xeon (E5-2683), to ensure we don't penalize one x86 vendor in favor of the other
Below is a comparison between two Focal containers, leveraging LXD to make use of different libc versions on the same host:
$ hyperfine -n libc-2.31-0ubuntu9.2 'lxc exec focal ./test_memcpy64 32' -n libc-patched 'lxc exec focal-patched ./test_memcpy64 32'
Benchmark #1: libc-2.31-0ubuntu9.2
Time (mean ± σ): 2.723 s ± 0.013 s [User: 4.7 ms, System: 5.1 ms]
Range (min … max): 2.693 s … 2.735 s 10 runs
Benchmark #2: libc-patched
Time (mean ± σ): 1.522 s ± 0.004 s [User: 3.9 ms, System: 5.6 ms]
Range (min … max): 1.515 s … 1.528 s 10 runs
Summary
'libc-patched' ran
1.79 ± 0.01 times faster than 'libc-2.31-0ubuntu9.2'
$ head -n5 /proc/cpuinfo
processor : 0
vendor_id : AuthenticAMD
cpu family : 23
model : 113
model name : AMD Ryzen 7 3700X 8-Core Processor
[0] https://bugzilla.redhat.com/show_bug.cgi?id=1880670
[1] https://github.com/sharkdp/hyperfine/
Below is the same benchmark from the case description, validating the performance regression on Groovy. Patched glibc builds for both Focal and Groovy amd64/i386 are available at ppa:halves/
$ hyperfine -n groovy-
Benchmark #1: groovy-
Time (mean ± σ): 2.392 s ± 0.039 s [User: 4.8 ms, System: 5.4 ms]
Range (min … max): 2.366 s … 2.494 s 10 runs
Benchmark #2: groovy-patched
Time (mean ± σ): 1.381 s ± 0.010 s [User: 6.1 ms, System: 4.3 ms]
Range (min … max): 1.372 s … 1.407 s 10 runs
Summary
'groovy-patched' ran
1.73 ± 0.03 times faster than 'groovy-
$ head -n5 /proc/cpuinfo
processor : 0
vendor_id : AuthenticAMD
cpu family : 23
model : 113
model name : AMD Ryzen 7 3700X 8-Core Processor
[0] https:/