diff options
| author |  Michael Smith <mikesmiffy128@gmail.com>
| 2026-10-11 00:47:00 +0100 |
|---|---|---|
| committer | Michael Smith <mikesmiffy128@gmail.com>
| 2026-10-11 00:47:00 +0100 |
| commit | 2986be1dd50be033cd8573afd544fc164bd12bc7 (patch) | |
| tree | ab55956fc2ab57e2a30d50ed4985b37aa379dbc2 /src/chunklets/fastspin.c | |
| parent | 5f0bf6dbf6291fbbbcf6db00eccde479f99a5608 (diff) | |
| download | sst-2986be1dd50be033cd8573afd544fc164bd12bc7.tar.gz sst-2986be1dd50be033cd8573afd544fc164bd12bc7.zip | |
Spin way less before futexing in fastspin
According to some reading I did recently on hybrid locks, a value of
around 40 is optimal, or at least was at one point. It might have
changed in more recent years and I have by no means made any effort to
tune the value, but it seems clear that 1000 is way too high.
Diffstat (limited to 'src/chunklets/fastspin.c')
| -rw-r--r-- | src/chunklets/fastspin.c | 4 |
1 files changed, 2 insertions, 2 deletions
diff --git a/src/chunklets/fastspin.c b/src/chunklets/fastspin.c index d349395..c61d781 100644 --- a/src/chunklets/fastspin.c +++ b/src/chunklets/fastspin.c @@ -236,7 +236,7 @@ int fastspin_wait(volatile int *p_) { #else if (x > 0) return x; if (!x) { - for (int c = 1000; c; --c) { + for (int c = 40; c; --c) { x = atomic_load_explicit(p, memory_order_relaxed); RELAX(); if (x > 0) return x; @@ -269,7 +269,7 @@ top: x = 0; return; } if (x) { - for (int c = 1000; c; --c) { + for (int c = 40; c; --c) { x = atomic_load_explicit(p, memory_order_relaxed); RELAX(); // note: top sets x to 0 unnecessarily but clang actually does |