Re: Background writer configuration - Mailing list pgsql-performance

From Steve Atkins
Subject Re: Background writer configuration
Date
Msg-id 6D84E9DE-E8AB-4C5D-8452-E141DBE9540D@blighty.com
Whole thread Raw
In response to Re: Background writer configuration  (Evgeny Gridasov <eugrid@fpm.kubsu.ru>)
Responses Re: Background writer configuration  ("Steinar H. Gunderson" <sgunderson@bigfoot.com>)
List pgsql-performance
On Mar 17, 2006, at 4:24 AM, Evgeny Gridasov wrote:

> Yesterday we recieved a new server 2xAMD64(2core x 2chips = 4 cores)
> 8GB RAM and RAID-1 (LSI megaraid)
> I've maid some tests with pgbench (scaling 1000, database size ~ 16Gb)
>
> First of all, I'd like to mention that it was strange to see that
> the server performance degraded by 1-2% when we changed kernel/
> userland to x86_64
> from default installed i386 userland/amd64 kernel. The operating
> system was Debian Linux,
> filesystem ext3.

64 bit binaries usually run marginally slower than 32 bit binaries.
AIUI the main reason is that they're marginally bigger, so fit less
well in cache, have to haul themselves over the memory channels
and so on. They're couch potato binaries. I've seen over 10% performance
loss in compute-intensive code, so a couple of percent isn't too
bad at all.

If that 64 bit addressing gets you cheap access to lots of RAM, and
your main applications can make good use of that then
that can easily outweigh the overall loss in performance

Cheers,
   Steve


pgsql-performance by date:

Previous
From: "Merlin Moncure"
Date:
Subject: Re: Help optimizing a slow index scan
Next
From: "Steinar H. Gunderson"
Date:
Subject: Re: Background writer configuration