Re: Shared buffers vs large files - Mailing list pgsql-general

From Glen Parker
Subject Re: Shared buffers vs large files
Date
Msg-id 028201c1c17c$d65b7cf0$0b01a8c0@johnpark.net
Whole thread Raw
In response to Shared buffers vs large files  (Francisco Reyes <lists@natserv.com>)
Responses Re: Shared buffers vs large files  (Neil Conway <nconway@klamath.dyndns.org>)
Re: Shared buffers vs large files  (Francisco Reyes <lists@natserv.com>)
List pgsql-general
> shared_buffers at 4096 (32MB if my calculations are correct),
> sort_mem = 65536             # min 32
> vacuum_mem = 16384          # min 1024
>
> The machine has 1GB of ram.
>
> I don't expect to have more than a handfull of connections at a time (from
> 1 to 10). Should I increate the shared buffers to 64MB? 128MB?

On a 1GB machine (still PG 7.1.3) I'm currently running:

shared_buffers: 48000 (about 400MB)
sort_mem: 8192

I haven't done much testing with sort_mem values, but...

This is very very VERY unscientific, but I haven't seen a shared_buffers
value that is so big that it seems to hurt performance (unless it causes
swapping obviously), and my installation is dedicated to postgres so I don't
need the memory for much of anything else.  It appears (and it makes sense)
that the performance improvement is roughly an inverse J-curve; bigger is
never really a bad thing, it just starts to make very little difference.
Any time you can save a system call and a memory copy, you're ahead.

I'd say that 4096 is VERY low for shared_mem, especially with so much
available ram - I'd bet the farm you'd see a *significant* improvement by
bumping it to 16384 at least.

Just my $.02 :-)

Glen


pgsql-general by date:

Previous
From: Manuel Sugawara
Date:
Subject: Re: terminal with horizontal scrollbar
Next
From: Alex Rice
Date:
Subject: oids vs. serial question