Josh Berkus <josh@agliodbs.com> writes:
> So, now that we know what the performance bottleneck is, how do we fix it?
Josh, I see that all of those runs seem to be using wal_buffers = 8.
Have you tried materially increasing wal_buffers (say to 100 or 1000)
and/or experimenting with different wal_sync_method options since we
fixed the bufmgrlock problem? I am wondering if the real issue is
WAL buffer contention or something like that.
It would also be useful to compare these runs to runs with fsync = off,
just to see how the performance changes.
regards, tom lane
