[prev in list] [next in list] [prev in thread] [next in thread] 

List:       pgsql-performance
Subject:    Re: [PERFORM] a heavy duty operation on an "unused" table kills my
From:       Matthew Wakeling <matthew () flymine ! org>
Date:       2010-01-21 12:02:16
Message-ID: alpine.DEB.2.00.1001211151240.6195 () aragorn ! flymine ! org
[Download RAW message or body]

On Wed, 20 Jan 2010, Greg Smith wrote:
>> Basically, to an extent, that's right. However, when you get 16 drives or 
>> more into a system, then it starts being an issue.
>
> I guess if I test a system with *only* 16 drives in it one day, maybe I'll 
> find out.

*Curious* What sorts of systems have you tried so far?

As the graph I just sent shows, the four schedulers are pretty-much 
identical in performance, until you start saturating it with simultaneous 
requests. CFQ levels out at a performance a little lower than the other 
three.

> Seriously though, there is some difference between a completely synthetic 
> test like you noted issues with here, and anything you can see when running 
> the database.

Granted, this test is rather synthetic. It is testing the rather unusual 
case of lots of simultaneous random small requests - more simultaneous 
requests than we advise people to run backends on a server. You'd probably 
need to get a RAID array a whole lot bigger than 16 drives to have a 
"normal workload" capable of demonstrating the performance difference, and 
even that isn't particularly major.

Would be interesting research if anyone has a 200-spindle RAID array 
hanging around somewhere.

Matthew

-- 
 A good programmer is one who looks both ways before crossing a one-way street.
 Considering the quality and quantity of one-way streets in Cambridge, it
 should be no surprise that there are so many good programmers there.

-- 
Sent via pgsql-performance mailing list (pgsql-performance@postgresql.org)
To make changes to your subscription:
http://www.postgresql.org/mailpref/pgsql-performance
[prev in list] [next in list] [prev in thread] [next in thread]