On Wed, 2007-01-24 at 09:32 +0530, Pavan Deolasee wrote:
> On a typical desktop class 2 CPU Dell machine, we have seen pgbench
> clocking more than 1500 tps. That implies CLOG would get filled up in
> less
> than 262144/1500=174 seconds. VACUUM on accounts table takes much
> longer to trigger.
You assume that all of the top level transactions have no
subtransactions. On that test, subtransactions are in use because of the
EXCEPTION clause in the PL/pgSQL used. That should at least double the
number of Xids.
> So
> most of the 636528 reads in the next 55 minutes can be attributed to
> VACUUM.
A similar argument might also be applied to subtrans, so a similar
investigation seems worthwhile. Subtrans has space for less Xids than
clog, BTW.
OTOH, I do think that 99% of that will not cause I/O.
-- Simon Riggs EnterpriseDB http://www.enterprisedb.com