On Sat, Mar 10, 2012 at 7:55 PM, Robert Haas <robertmhaas@gmail.com> wrote:
> I've finally been able to run some more tests of the effect of
> adjusting wal_buffers to values higher than 16MB. I ran the test on
> the 16 core (x 4 hw threads/core) IBM POWER7 machine, with my usual
> configuration settings:
>
> shared_buffers = 8GB
> maintenance_work_mem = 1GB
> synchronous_commit = off
> checkpoint_segments = 300
> checkpoint_timeout = 15min
> checkpoint_completion_target = 0.9
> wal_writer_delay = 20ms
>
> I ran three 30-minute tests at scale factor 300 with wal_buffers set
> at various values from 16MB up to 160MB, in multiples of 16MB, using
> pgbench with 32 clients and 32 threads in each case. The short
> version is that 32MB seems to be significantly better than 16MB, by
> about 1000 tps, and after that it gets murky; full results are below.
On Nate Boley's machine, the difference was ~100% increase rather than
~10%. Do you think the difference is in the CPU architecture, or the
IO subsystem?
Also, do you have the latency numbers?
Cheers,
Jeff