Re: "Mysterious" issues with newly installed 8.3 - Mailing list pgsql-performance

From Greg Smith
Subject Re: "Mysterious" issues with newly installed 8.3
Date
Msg-id Pine.GSO.4.64.0810100345010.26609@westnet.com
Whole thread Raw
In response to Re: "Mysterious" issues with newly installed 8.3  ("Scott Carey" <scott@richrelevance.com>)
Responses Re: "Mysterious" issues with newly installed 8.3  (Carlos Moreno <morenopg@mochima.com>)
List pgsql-performance
On Thu, 9 Oct 2008, Scott Carey wrote:

> Sure, the processor should be faster, but Athlon64s / Opterons are very
> sensitive to the RAM used and its performance and tuning. So, you should
> find some basic CPU benchmarks and RAM benchmarks -- you'll want to
> measure latency as well as bandwidth. Athlon64 and Opteron both
> typically have two memory busses per processor, and it is possible to
> populate the memory banks in such a way that the system has half the
> bandwidth.

This is really something to watch out for.  One quick thing first though:
what frequency does the CPU on the new server show when you look at
/proc/cpuinfo?  If you see "cpu MHz:  1000.00" you probably are throttling
the CPU down hard with power management which was cause the slowness you
describe.  In that case I'd suggest editing /etc/sysconfig/cpuspeed and
changing "GOVERNER=performance".

Back to memory.  What I do with any new, untrusted system is boot with a
memtest86+ CD is take a look at the memory speed information it shows,
with the most important number being the uncached RAM speed.  If you're
not running in dual-channel mode and at the maximum frequency the RAM
supports, that can run seriously slow things down.  You probably can't
take down the production server for comparison.  I can tell you that on my
little Athlon X2@2.4GHz server, I've seen the memtest86+ reported raw
memory speed run anywhere from 2093MB/s (with crummy DDR2 667 that doesn't
match the CPU bus frequency very well) to 3367MB/s (using good DDR2 800).
You should see even better from your Opteron system.

Another really handy way to gauge memory speed on Linux, if there are
similar kernels installed on each system like your case, is to use "hdparm
-T".  That cached read figure is highly correlated with overall memory
performance.  The nice part about that is you can probably get a useful
comparison result from the old server if you run that a bunch of times
even with other activity (just take the highest number you ever see),
whereas memtest86+ requires some downtime.  Those numbers are lower than
the I'd expect around 2500MB/s out of your new server here (that's what I
got when I just tested an Opteron 2220 system @2.8GHz using the RHEL5
hdparm -T).

--
* Greg Smith gsmith@gregsmith.com http://www.gregsmith.com Baltimore, MD

pgsql-performance by date:

Previous
From: Greg Smith
Date:
Subject: Re: "Mysterious" issues with newly installed 8.3
Next
From: "Віталій Тимчишин"
Date:
Subject: Re: Delete performance again