On 08/31/2015 12:54 PM, YUriy Zhuravlev wrote:
> Hello hackers
>
> Recently, we were given access to the test server is IBM, 9119-MHE with 8 CPUs
> * 8 cores * 8 threads. We decided to take advantage of this and to find
> bottlenecks for read scalability (pgbench -S).
>
> All detail you can read here:
> http://www.postgrespro.ru/blog/pgsql/2015/08/30/p8scaling
>
> Performance 9.4 stopped growing after 100 clients, and 9.5 / 9.6 stopped after
> 150 (at 4 NUMA nodes). After research using pref we saw that inhibits
> ProcArrayLock in GetSnaphotData. But inserting the stub instead of
> GetSnapshotData not significantly increased scalability. Trying to find the
> bottleneck with gdb, we found another place. We have noticed s_lock in
> PinBuffer and UnpinBuffer. For the test we rewrited PinBuffer and UnpinBuffer
> by atomic operations and we liked the result. Degradation of performance
> almost completely disappeared, and went scaling up to 400 clients (4 NUMA
> nodes with 256 "CPUs").
>
> To scale Postgres for large NUMA machine must be ported to the atomic
> operations bufmgr. During our tests, we no found errors in our patch, but most
> likely it is not true and this patch only for test.
>
> Who has any thoughts?
Well, I could test the patch on a x86 machine with 4 sockets (64 cores),
but I wonder whether it makes sense at this point, as the patch really
is not correct (judging by what Andres says).
Also, what pgbench scale was used for the testing?
regards
--
Tomas Vondra http://www.2ndQuadrant.com
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services
