On Nov 21, 2010, at 12:16 PM, Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:
> tv(at)fuzzy(dot)cz writes:
>>> Second, I modified the work_mem setting to 2GB (reloaded config) and I see
>>> a response time of 38 seconds. Results below from EXPLAIN ANALYZE:
>> How did you reload the config? Using 'kill -HUP pid'? That should work
>> fine. Have you cheched 'work_mem' after the reload?
>> Because the explain plans are exactly the same (structure, estimated
>> costs). The really interesting bit is this and it did not change at all
>> Buckets: 1024 Batches: 64 Memory Usage: 650kB
> If that didn't change, I'm prepared to bet that the OP didn't actually
> manage to change the active value of work_mem.
Yep. All this speculation about slow disks and/or COALESCE strikes me as likely totally off-base. I think the original poster needs to run "show work_mem" right before the EXPLAIN ANALYZE to make sure the new value they set actually stuck. There's no reason for the planner to have used only 650kB if work_mem is set to anything >=2MB.
In response to
pgsql-performance by date
|Next:||From: Ivan Voras||Date: 2010-11-22 00:15:43|
|Subject: Performance under contention|
|Previous:||From: Tom Lane||Date: 2010-11-21 17:16:01|
|Subject: Re: Query Performance SQL Server vs. Postgresql |