Re: understanding postgres issues/bottlenecks

On Wed, Jan 7, 2009 at 2:02 PM, Stefano Nichele
<stefano.nichele@gmail.com> wrote:
>
> On Tue, Jan 6, 2009 at 7:45 PM, Scott Marlowe <scott.marlowe@gmail.com>
> wrote:
>>
>> I concur with Merlin you're I/O bound.
>>
>> Adding to his post, what RAID controller are you running, does it have
>> cache, does the cache have battery backup, is the cache set to write
>> back or write through?
>
>
> At the moment I don't have such information. It's a "standard" RAID
> controller coming with a DELL server. Is there any information I can have
> asking to the SO ?

You can run lshw to see what flavor controller it is.  Dell RAID
controllers are pretty much either total crap, or mediocre at best.
The latest one, the Perc 6 series are squarely in the same performance
realm as a 4 or 5 year old LSI megaraid.  The perc 5 series and before
are total performance dogs.  The really bad news is that you can't
generally plug in a real RAID controller on a Dell.  We put an Areca
168-LP PCI-x8 in one of our 1950s and it wouldn't even turn on, got a
CPU Error.

Dells are fine for web servers and such.  For database servers they're
a total loss.  The best you can do with one is to put a generic SCSI
card in it and connect to an external array with its own controller.

We have a perc6e and a perc5e in two different servers, and no matter
how we configure them, we can't get even 1/10th the performance of an
Areca controller with the same number of drives on another machine of
the same basic class as the 1950s.

>> Also, what do you get for this (need contrib module pgbench installed)
>>
>> pgbench -i -s 100
>> pgbench -c 50 -n 10000
>>
>> ? Specifically transactions per second?
>
> I'll run pgbench in the next days.

Cool.  That pgbench is a "best case scenario" benchmark.  Lots of
small transactions on a db that should fit into memory.  If you can't
pull off a decent number there (at least a few hundred tps) then can't
expect better performance from real world usage.

Oh, and that should be:

pgbench -c 50 -t 10000

not -n... not enough sleep I guess.