diff options
| author | Garrett D'Amore <garrett@damore.org> | 2017-08-10 00:10:50 -0700 |
|---|---|---|
| committer | Garrett D'Amore <garrett@damore.org> | 2017-08-10 00:10:50 -0700 |
| commit | ac5f0ef7cf501693a9db2fcbd95b7cde419cbb2a (patch) | |
| tree | 49f479185a08e8f4b2538b3fb69ab57319a4ba60 /src/core/aio.h | |
| parent | 9feb54e9c7ab116ba566086a76604338f86e3bc3 (diff) | |
| download | nng-ac5f0ef7cf501693a9db2fcbd95b7cde419cbb2a.tar.gz nng-ac5f0ef7cf501693a9db2fcbd95b7cde419cbb2a.tar.bz2 nng-ac5f0ef7cf501693a9db2fcbd95b7cde419cbb2a.zip | |
Thundering herd kills performance.
A little benchmarking showed that we were encountering far too many
wakeups, leading to severe performance degradation; we had a bunch
of threads all sleeping on the same condition variable (taskqs)
and this woke them all up, resulting in heavy mutex contention.
Since we only need one of the threads to wake, and we don't care which
one, let's just wake only one. This reduced RTT latency from about
240 us down to about 30 s. (1/8 of the former cost.)
There's still a bunch of tuning to do; performance remains worse than
we would like.
Diffstat (limited to 'src/core/aio.h')
| -rw-r--r-- | src/core/aio.h | 3 |
1 files changed, 2 insertions, 1 deletions
diff --git a/src/core/aio.h b/src/core/aio.h index 0f41c01f..d48442eb 100644 --- a/src/core/aio.h +++ b/src/core/aio.h @@ -34,7 +34,8 @@ struct nni_aio { unsigned a_pend : 1; // completion routine pending unsigned a_active : 1; // aio was started unsigned a_expiring : 1; // expiration callback in progress - unsigned a_pad : 27; // ensure 32-bit alignment + unsigned a_waiting : 1; // a thread is waiting for this to finish + unsigned a_pad : 26; // ensure 32-bit alignment nni_task a_task; // Read/write operations. |