Problem statement

The relevant part is the use of atomic_fetch_add and atomic_fetch_sub in rwlock_* functions. In short, clang-8 and earlier versions generate lock addl and lock subl for those atomics. clang-9 generates lock incl and lock decl, and also lock xadd if adding 2 or other numbers. The results show that lock addl and lock subl can be more efficient in some cases. The rwlock implementation can be wrong. Any suggestions are welcome. I have not identified the actual cause. Right now I just personally hope to get the old compilation results with clang-9.

In this microbenchmark each thread repeatedly acquires and releases a randomly selected lock from a set. The number of threads and the number of locks are changed in the benchmark. Each thread is pinned to a unique core, assuming #threads <= #cores. The program reads the process' affinity settings to determine how many cores can be used.

Change the compiler with CC=x such as CC=clang-9. The default is clang. If having multiple sockets, use numactl -N 0 ./rwlocktest.out to let it run on one socket.

Results

Tested on both Broadwell (16-core) and Skylake (14-core) servers, with Archlinux (kernel 5.3.x) and Ubuntu 18.04, respectively.

Broadwell has HT turned off. Skylake has HT turned on (not on purpose) so I used the first 14 cores out of the 28 on a socket. Skylake results are omitted (similar to Broadwell's).

Results on Broadwell are listed below. Server specs: Archlinux, linux 5.3.6, Xeon E5-2697A v4, mitigations=off.

We observed that with more locks (128+), the binary produced by clang-9 yields 15-20 percent lower throughput. However, with a _mm_pause() added in between the lock() and unlock(), the peak throughput (180+) can be reached by both of them. But the throughput under high-contention (32 locks) is much lower (roughly 40 vs 60). I have not exhausted all the possibilities by adding pauses, reordering the statements, and writing inline assembly to force doing lock addl with clang-9. Some are slightly better, but none has produced results as good as clang-8's.

results with clang 8/9 (formatted for comparison)

#threads:   4   #locks:  32  clang-9:  48.66 million ops/sec   clang-8:  48.60 million ops/sec
#threads:   6   #locks:  32  clang-9:  54.94 million ops/sec   clang-8:  55.55 million ops/sec
#threads:   8   #locks:  32  clang-9:  57.63 million ops/sec   clang-8:  58.30 million ops/sec
#threads:  10   #locks:  32  clang-9:  59.89 million ops/sec   clang-8:  61.12 million ops/sec
#threads:  12   #locks:  32  clang-9:  61.24 million ops/sec   clang-8:  62.14 million ops/sec
#threads:  14   #locks:  32  clang-9:  61.87 million ops/sec   clang-8:  62.75 million ops/sec
#threads:  16   #locks:  32  clang-9:  61.41 million ops/sec   clang-8:  62.77 million ops/sec
#threads:   4   #locks: 128  clang-9:  58.54 million ops/sec   clang-8:  64.85 million ops/sec
#threads:   6   #locks: 128  clang-9:  77.23 million ops/sec   clang-8:  86.35 million ops/sec
#threads:   8   #locks: 128  clang-9:  92.27 million ops/sec   clang-8: 102.66 million ops/sec
#threads:  10   #locks: 128  clang-9: 103.32 million ops/sec   clang-8: 117.78 million ops/sec
#threads:  12   #locks: 128  clang-9: 114.16 million ops/sec   clang-8: 128.92 million ops/sec
#threads:  14   #locks: 128  clang-9: 121.44 million ops/sec   clang-8: 137.28 million ops/sec
#threads:  16   #locks: 128  clang-9: 126.56 million ops/sec   clang-8: 144.77 million ops/sec
#threads:   4   #locks: 512  clang-9:  55.80 million ops/sec   clang-8:  68.21 million ops/sec
#threads:   6   #locks: 512  clang-9:  76.55 million ops/sec   clang-8:  93.21 million ops/sec
#threads:   8   #locks: 512  clang-9:  95.32 million ops/sec   clang-8: 116.38 million ops/sec
#threads:  10   #locks: 512  clang-9: 113.31 million ops/sec   clang-8: 138.59 million ops/sec
#threads:  12   #locks: 512  clang-9: 128.26 million ops/sec   clang-8: 157.17 million ops/sec
#threads:  14   #locks: 512  clang-9: 141.07 million ops/sec   clang-8: 171.94 million ops/sec
#threads:  16   #locks: 512  clang-9: 150.64 million ops/sec   clang-8: 185.40 million ops/sec
#threads:   4   #locks:  32  clang-9:  49.35 million ops/sec   clang-8:  49.83 million ops/sec
#threads:   6   #locks:  32  clang-9:  54.67 million ops/sec   clang-8:  55.24 million ops/sec
#threads:   8   #locks:  32  clang-9:  57.63 million ops/sec   clang-8:  58.63 million ops/sec
#threads:  10   #locks:  32  clang-9:  60.32 million ops/sec   clang-8:  61.11 million ops/sec
#threads:  12   #locks:  32  clang-9:  61.18 million ops/sec   clang-8:  62.23 million ops/sec
#threads:  14   #locks:  32  clang-9:  61.82 million ops/sec   clang-8:  62.03 million ops/sec
#threads:  16   #locks:  32  clang-9:  60.91 million ops/sec   clang-8:  62.77 million ops/sec
#threads:   4   #locks: 128  clang-9:  58.78 million ops/sec   clang-8:  64.83 million ops/sec
#threads:   6   #locks: 128  clang-9:  77.39 million ops/sec   clang-8:  86.15 million ops/sec
#threads:   8   #locks: 128  clang-9:  92.70 million ops/sec   clang-8: 103.39 million ops/sec
#threads:  10   #locks: 128  clang-9: 104.53 million ops/sec   clang-8: 117.19 million ops/sec
#threads:  12   #locks: 128  clang-9: 114.24 million ops/sec   clang-8: 128.67 million ops/sec
#threads:  14   #locks: 128  clang-9: 121.34 million ops/sec   clang-8: 137.87 million ops/sec
#threads:  16   #locks: 128  clang-9: 126.83 million ops/sec   clang-8: 144.52 million ops/sec
#threads:   4   #locks: 512  clang-9:  55.78 million ops/sec   clang-8:  68.18 million ops/sec
#threads:   6   #locks: 512  clang-9:  76.47 million ops/sec   clang-8:  92.97 million ops/sec
#threads:   8   #locks: 512  clang-9:  94.64 million ops/sec   clang-8: 116.29 million ops/sec
#threads:  10   #locks: 512  clang-9: 113.06 million ops/sec   clang-8: 138.39 million ops/sec
#threads:  12   #locks: 512  clang-9: 128.22 million ops/sec   clang-8: 156.90 million ops/sec
#threads:  14   #locks: 512  clang-9: 140.87 million ops/sec   clang-8: 172.90 million ops/sec
#threads:  16   #locks: 512  clang-9: 150.84 million ops/sec   clang-8: 184.95 million ops/sec`

result with _mm_pause() between lock and unlock (uncomment line #372)

#threads:   4   #locks:  32  clang-9:  38.18 million ops/sec   clang-8:  39.24 million ops/sec
#threads:   6   #locks:  32  clang-9:  40.24 million ops/sec   clang-8:  40.11 million ops/sec
#threads:   8   #locks:  32  clang-9:  40.58 million ops/sec   clang-8:  40.55 million ops/sec
#threads:  10   #locks:  32  clang-9:  39.76 million ops/sec   clang-8:  40.95 million ops/sec
#threads:  12   #locks:  32  clang-9:  39.70 million ops/sec   clang-8:  40.93 million ops/sec
#threads:  14   #locks:  32  clang-9:  39.65 million ops/sec   clang-8:  40.56 million ops/sec
#threads:  16   #locks:  32  clang-9:  39.36 million ops/sec   clang-8:  39.61 million ops/sec
#threads:   4   #locks: 128  clang-9:  63.01 million ops/sec   clang-8:  63.58 million ops/sec
#threads:   6   #locks: 128  clang-9:  73.84 million ops/sec   clang-8:  74.76 million ops/sec
#threads:   8   #locks: 128  clang-9:  82.95 million ops/sec   clang-8:  84.40 million ops/sec
#threads:  10   #locks: 128  clang-9:  85.74 million ops/sec   clang-8:  89.98 million ops/sec
#threads:  12   #locks: 128  clang-9:  90.47 million ops/sec   clang-8:  93.14 million ops/sec
#threads:  14   #locks: 128  clang-9:  90.55 million ops/sec   clang-8:  96.25 million ops/sec
#threads:  16   #locks: 128  clang-9:  92.44 million ops/sec   clang-8:  98.99 million ops/sec
#threads:   4   #locks: 512  clang-9:  74.36 million ops/sec   clang-8:  74.64 million ops/sec
#threads:   6   #locks: 512  clang-9: 100.40 million ops/sec   clang-8: 100.49 million ops/sec
#threads:   8   #locks: 512  clang-9: 123.69 million ops/sec   clang-8: 122.31 million ops/sec
#threads:  10   #locks: 512  clang-9: 140.79 million ops/sec   clang-8: 143.30 million ops/sec
#threads:  12   #locks: 512  clang-9: 158.99 million ops/sec   clang-8: 158.62 million ops/sec
#threads:  14   #locks: 512  clang-9: 172.64 million ops/sec   clang-8: 172.19 million ops/sec
#threads:  16   #locks: 512  clang-9: 184.51 million ops/sec   clang-8: 184.57 million ops/sec
#threads:   4   #locks:  32  clang-9:  38.36 million ops/sec   clang-8:  39.67 million ops/sec
#threads:   6   #locks:  32  clang-9:  40.14 million ops/sec   clang-8:  40.06 million ops/sec
#threads:   8   #locks:  32  clang-9:  40.62 million ops/sec   clang-8:  40.85 million ops/sec
#threads:  10   #locks:  32  clang-9:  39.99 million ops/sec   clang-8:  40.77 million ops/sec
#threads:  12   #locks:  32  clang-9:  39.58 million ops/sec   clang-8:  40.88 million ops/sec
#threads:  14   #locks:  32  clang-9:  39.69 million ops/sec   clang-8:  40.50 million ops/sec
#threads:  16   #locks:  32  clang-9:  39.17 million ops/sec   clang-8:  40.54 million ops/sec
#threads:   4   #locks: 128  clang-9:  63.34 million ops/sec   clang-8:  63.62 million ops/sec
#threads:   6   #locks: 128  clang-9:  73.35 million ops/sec   clang-8:  75.12 million ops/sec
#threads:   8   #locks: 128  clang-9:  80.88 million ops/sec   clang-8:  84.62 million ops/sec
#threads:  10   #locks: 128  clang-9:  85.64 million ops/sec   clang-8:  89.63 million ops/sec
#threads:  12   #locks: 128  clang-9:  88.48 million ops/sec   clang-8:  92.91 million ops/sec
#threads:  14   #locks: 128  clang-9:  92.44 million ops/sec   clang-8:  95.29 million ops/sec
#threads:  16   #locks: 128  clang-9:  91.18 million ops/sec   clang-8:  98.35 million ops/sec
#threads:   4   #locks: 512  clang-9:  74.33 million ops/sec   clang-8:  74.70 million ops/sec
#threads:   6   #locks: 512  clang-9: 100.29 million ops/sec   clang-8:  99.91 million ops/sec
#threads:   8   #locks: 512  clang-9: 123.39 million ops/sec   clang-8: 124.08 million ops/sec
#threads:  10   #locks: 512  clang-9: 143.11 million ops/sec   clang-8: 143.09 million ops/sec
#threads:  12   #locks: 512  clang-9: 160.06 million ops/sec   clang-8: 158.95 million ops/sec
#threads:  14   #locks: 512  clang-9: 173.70 million ops/sec   clang-8: 173.07 million ops/sec
#threads:  16   #locks: 512  clang-9: 185.78 million ops/sec   clang-8: 184.45 million ops/sec

clang versions

$ clang --version
clang version 9.0.0 (tags/RELEASE_900/final)
Target: x86_64-pc-linux-gnu
Thread model: posix
InstalledDir: /usr/bin

$ clang --version
clang version 8.0.1 (tags/RELEASE_801/final)
Target: x86_64-pc-linux-gnu
Thread model: posix
InstalledDir: /usr/bin

Disassembly

clang-8

rwlock_trylock_read():
/home/x/work/rw/rwlocktest.c:284
  if ((atomic_fetch_add(pvar, 1) >> RWLOCK_WSHIFT) == 0) {
    2869: f0 83 04 f0 01        lock addl $0x1,(%rax,%rsi,8)
    286e: 7e e0                 jle    2850 <rwlock_worker+0x250>
rwlock_worker():
/home/x/work/rw/rwlocktest.c:373
        rwlock_unlock_read(&(info->locks[x]));
    2870: 48 8b 03              mov    (%rbx),%rax
rwlock_unlock_read():
/home/x/work/rw/rwlocktest.c:307
  atomic_fetch_sub(pvar, 1);
    2873: f0 83 2c f0 01        lock subl $0x1,(%rax,%rsi,8)

clang-9

rwlock_trylock_read():
    283f: 90                    nop
/home/x/work/rw/rwlocktest.c:284
  if ((atomic_fetch_add(pvar, 1) >> RWLOCK_WSHIFT) == 0) {
    2840: f0 ff 04 f0           lock incl (%rax,%rsi,8)
    2844: 7f 1a                 jg     2860 <rwlock_worker+0x270>
/home/x/work/rw/rwlocktest.c:287
    atomic_fetch_sub(pvar, 1);
    2846: f0 ff 0c f0           lock decl (%rax,%rsi,8)

clang-9, but with `atomic_fetch_add(pvar, 3)`

(benchmarking results omitted, not improved)

rwlock_trylock_read():
/home/x/work/rw/rwlocktest.c:284
  if ((atomic_fetch_add(pvar, 3) >> RWLOCK_WSHIFT) == 0) {
    2840: ba 03 00 00 00        mov    $0x3,%edx
    2845: f0 0f c1 14 f0        lock xadd %edx,(%rax,%rsi,8)
    284a: 85 d2                 test   %edx,%edx
    284c: 79 22                 jns    2870 <rwlock_worker+0x280>
/home/x/work/rw/rwlocktest.c:287
    atomic_fetch_sub(pvar, 3);
    284e: f0 83 2c f0 03        lock subl $0x3,(%rax,%rsi,8)

Archlinux clang archive packages

https://archive.archlinux.org/packages/c/compiler-rt/
https://archive.archlinux.org/packages/l/llvm/
https://archive.archlinux.org/packages/l/llvm-libs/
https://archive.archlinux.org/packages/c/clang/

wuxb45 / atomictest Goto Github PK

atomictest's Introduction

Problem statement

Results

results with clang 8/9 (formatted for comparison)

result with _mm_pause() between lock and unlock (uncomment line #372)

clang versions

Disassembly

clang-8

clang-9

clang-9, but with `atomic_fetch_add(pvar, 3)`

Archlinux clang archive packages

atomictest's People

Contributors

Watchers

Recommend Projects

React

Vue.js

Typescript

TensorFlow

Django

Laravel

D3

Recommend Topics

javascript

web

server

Machine learning

Visualization

Game

Recommend Org

Facebook

Microsoft

Google

Alibaba

D3

Tencent

wuxb45 / atomictest Goto Github PK

atomictest's Introduction

Problem statement

Results

results with clang 8/9 (formatted for comparison)

result with _mm_pause() between lock and unlock (uncomment line #372)

clang versions

Disassembly

clang-8

clang-9

clang-9, but with atomic_fetch_add(pvar, 3)

Archlinux clang archive packages

atomictest's People

Contributors

Watchers

Recommend Projects

Recommend Topics

Recommend Org

clang-9, but with `atomic_fetch_add(pvar, 3)`