Alvaro Herrera <alvherre@commandprompt.com> writes:
> Jim C. Nasby wrote:
>> Along those lines, is there anything else that would benefit from being
>> moved? pg_clog and pg_subtrans come to mind; but maybe pg_multixact and
>> pg_twophase are candidates as well?
> Hmm, I doubt moving any of the SLRU files (clog, subtrans, multixact)
> will have much of an impact.
Certainly pushing them onto the WAL spindle would be a serious misstep.
There is a good case for giving WAL its own dedicated disk --- there is
no case that I've seen for giving any of these their own disk.
> If there's too much I/O on those, a better
> solution would be to increase the number of buffers allocated to them.
> Currently we use 8 for all of them which is probably not appropiate for
> everyone.
I've just been looking at a test case provided by Rob Creager that
causes some pretty severe contention on SubtransControlLock. There
are a number of possible answers to this, but increasing the number of
pg_subtrans buffers is definitely one of them. I think it's probably
time we got rid of the assumption that all the uses of slru.c should
have the same number of buffers ...
regards, tom lane