On Tue, Mar 13, 2012 at 3:48 PM, Robert Haas <robertmhaas(at)gmail(dot)com> wrote:
> On Mon, Mar 12, 2012 at 4:45 PM, Jeff Janes <jeff(dot)janes(at)gmail(dot)com> wrote:
>> Rerunning all 4 benchmarks (both 16MB and 32MB wal_buffers on both
>> machines) with fsync=off (as well as synchronous_commit=off still)
>> might help clarify things.
>
> I reran the 32-client benchmark on the IBM machine with fsync=off and got this:
>
> 32MB: tps = 26809.442903 (including connections establishing)
> 16MB: tps = 26651.320145 (including connections establishing)
>
> That's a speedup of nearly a factor of two, so clearly fsync-related
> stalls are a big problem here, even with wal_buffers cranked up
> through the ceiling.
And here's a tps plot with wal_buffers = 16MB, fsync = off. The
performance still bounces up and down, so there's obviously some other
factor contributing to latency spikes, but equally clearly, needing to
wait for fsyncs makes it a lot worse. I bet if we could understand
why that happens, we could improve things here a good deal.
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company
