On Nov 14, 2007, at 9:19 PM, Jeff Frost wrote:
> On an 8xRAID10 volume with the smaller Areca controller we were
> seeing around 450 seeks/sec.
On our 6 disk raid10 on a 3ware 9550sx I'm able to get about 120 seek
+ reads/sec per process, with an aggregate up to about 500 or so.
The disks are rather pooey 7.5k sata2 disks. I'd been having perf
issues and I'd been wondering why my IO stats were low.. turns out it
was going as fast as the disks or controller could go. I even went
so far as to write a small tool to sort-of simulate a PG index scan
to remove all that from the question. It proved my theory - seq
performance was murdering us.
This information led me to spend a pile of money on an MSA70 (HP) and
a pile of 15k SAS disks.
While significantly more expensive, the perf gains were astounding.
I have 8 disks in a raid6 (iirc, I had comprable numbers for R10, but
the space/cost/performance wasn't worth it). I'm able to get about
350-400tps, per process, with an aggregate somewhere in the 1000s. (I
drove it up to 2000 before stopping during testing)
Wether the problem is the controller or the disks, I don't know. I
just know what my numbers tell me. (And the day we went live on the
MSA a large number of our perf issues went away. Although, now that
the IO was plenty sufficient the CPU became the bottleneck! Its
always something!) The sata array performs remarkably well for a
sequential read though. Given our workload, we need the random perf
much more than seq, but I can see the opposite being true in a
btw, the tool I wrote is here http://pgfoundry.org/projects/pgiosim/
Jeff Trout <jeff(at)jefftrout(dot)com>
In response to
pgsql-performance by date
|Next:||From: Erik Jones||Date: 2007-11-15 16:02:36|
|Subject: Re: Curious about dead rows. |
|Previous:||From: Jean-David Beyer||Date: 2007-11-15 03:26:02|
|Subject: Re: Curious about dead rows.|