Re: Perc 3 DC - Mailing list pgsql-performance

From Scott Marlowe
Subject Re: Perc 3 DC
Date
Msg-id dcc563d10811221027t6115bc5cq1d705d3c6c89c411@mail.gmail.com
Whole thread Raw
In response to Re: Perc 3 DC  (Glyn Astill <glynastill@yahoo.co.uk>)
Responses Re: Perc 3 DC
List pgsql-performance
On Sat, Nov 22, 2008 at 9:59 AM, Glyn Astill <glynastill@yahoo.co.uk> wrote:
> --- On Sat, 22/11/08, Scott Marlowe <scott.marlowe@gmail.com> wrote:
>
>> You really have two choices.  First is to try and use it as
>> a plain
>> SCSI card, maybe with caching turned on, and do the raid in
>> software.
>> Second is to cut it into pieces and make jewelry out of it.
>
> Haha, I'm not really into jewelry, although I had thought of smacking it into a pile of dust with a lump hammer,
that'smuch more my thing. 

Well, I think the important thing here is to follow your bliss. :)

>>  Anything
>> before the Perc 6 series is seriously brain damaged, and
>> the Perc6
>> brings the dell raid array line squarly in line with a 5
>> year old LSI
>> megaraid, give or take.  And that's being generous.
>>
>
> Well this card thinks it's a 5 year old lsi megaraid. I've got a pile of perc5i megaraid paperweights on my desk at
work,so this was kinda expected really. 

Yeah we just wound up buying a RAID controller from Dell for handling
8 500Gig drives and no one checked and they sold us a Perc5, which can
only handle a 2TB array apparently.  Set it up as a RAID-10 and it's
definitely got "meh" levels of performance.  I had an old workstation
with a 4 port SATA card (no raid) running software raid and it handily
stomps this 8 disk machine into the ground.

>> > I've tried writeback and write through modes,
>> tried changing cache flush times, disabled and enabled
>> multiple PCI delayed transactions, all seem to have little
>> effect.
>>
>> Yeah, it's like trying to performance tune a yugo.
>
> Did I mention I drive a yugo?

Well, they're fine vehicles for what they do (I hear they're still
quite an icon in eastern europe) but they aren't gonna win a lot of
cash at the race track. :)

>> down.  But
>> you should certainly expect > 20 M/s on a modern RAID
>> controller
>> writing out to a 4 disk RAID10
>>
>
> Yeah the battery's on it, that and the 128Mb is really the only reason I thought I'd give it a whirl.

We had a bunch of 18xx series servers last company I was at (we went
from unix / linux to Microsoft, so ordered some 400 machines to
replace a dozen or so unix machines) and they all came with the
adaptec based Perc 3s.  We wound up with a large cardboard box full of
them and running software RAID to get decent performance and to stop
them from locking up randomly.  I had the only two running linux and
equipped with LSI based Perc3. They were stable, but the pair of intel
OEM boxes they replaced, which had 1/4 the CPU horsepower, were still
noticeably faster with their older but LSI based RAID controllers.  I
left that place as soon as I could, I just couldn't handle a large
portion of my job being to walk around a data center resetting locked
up windows boxes each morning.

I think as much as anything the busses on the dells are the problem,
resulting in pretty poor throughput, especially true of the old
serverworks chipset machines.  Those things are pretty much boat
anchors.

pgsql-performance by date:

Previous
From: PFC
Date:
Subject: Re: Hash join on int takes 8..114 seconds
Next
From: "Andrus"
Date:
Subject: Re: Hash join on int takes 8..114 seconds