From: Scott Carey <scott@richrelevance.com>
To: Scott Marlowe <scott.marlowe@gmail.com>, Ron <rjpeace@earthlink.net>
CC: Joe Uhl <joeuhl@gmail.com>, "pgsql-performance@postgresql.org"
	<pgsql-performance@postgresql.org>
Date: Tue, 24 Mar 2009 17:22:24 -0700
Subject: Re: High CPU Utilization
Thread-Topic: [PERFORM] High CPU Utilization
Thread-Index: Acms1qAGVEXU/srXQCmOanGKTaQxDgACSUCl
Message-ID: <C5EEC750.3B27%scott@richrelevance.com>
In-Reply-To: <dcc563d10903241616l2c16e803q7f327e4d7aaf5c43@mail.gmail.com>
Accept-Language: en-US
Content-Language: en
acceptlanguage: en-US
Content-Type: text/plain; charset="iso-8859-1"
Content-Transfer-Encoding: quoted-printable
MIME-Version: 1.0


On 3/24/09 4:16 PM, "Scott Marlowe" <scott.marlowe@gmail.com> wrote:

> On Tue, Mar 24, 2009 at 4:58 PM, Ron <rjpeace@earthlink.net> wrote:
>> At 02:47 PM 3/24/2009, Joe Uhl wrote:
>>=20
>>> Turns out we may have an opportunity to purchase a new database server
>>> with this increased load. =A0Seems that the best route, based on feedba=
ck to
>>> this thread, is to go whitebox, get quad opterons, and get a very good =
disk
>>> controller.
>>>=20
>>> Can anyone recommend a whitebox vendor?
>>=20
>> I'll 2nd the Aberdeen recommendation. =A0I'll add Pogolinux to that list=
 as
>> well.
>>=20
>>=20
>>> Is there a current controller anyone on this list has experience with t=
hat
>>> they could recommend?
>>=20
>> The 2 best performing RAID controller vendors at this time are AMCC (AKA
>> 3Ware) and Areca.
>> In general, the 8+ port Areca's with their BB cache maxed outperform eve=
ry
>> other controller available.

I personally have had rather bad performance experiences with 3Ware
9550/9650 SATA cards.  I have no experience with the AMCC SAS stuff though.
Adaptec demolished the 9650 on arrays larger than 4 drives, and Areca will
do better at the very high end.

However, if CPU is the issue for this particular case, then the RAID
controller details are less significant.

I don't know how much data you have, but don't forget the option of SSDs, o=
r
a mix of hard drives and SSDs for different data.  Ideally, you would want
the OS to just extend its pagecache onto a SSD, but only OpenSolaris can do
that right now and it is rather new (needs to be persistent across reboots)=
.

http://blogs.sun.com/brendan/entry/test
http://blogs.sun.com/brendan/entry/l2arc_screenshots
=20

>>=20
>>=20
>>> This will be a bigger purchase so will be doing research and benchmarki=
ng
>>> but any general pointers to a vendor/controller greatly appreciated.
>>=20
>> Be =3Dvery=3D careful to thoroughly bench both the AMD and Intel CPU opt=
ions.
>> =A0It is far from clear which is the better purchase.
>=20
> My anecdotal experience has been that the Opterons stay afloat longer
> as load increases, but I haven't had machines with similar enough
> hardware to really test that.
>=20

One may want to note that Intel's next generation servers are due out withi=
n
45 days from what I can sense ('Q2' traditionally means ~April 1 for Intel
when on time).  These should be a rather significant bump for a database as
they adopt the AMD / Alpha style memory-controller-on-CPU architecture and
add a lot of cache.  Other relevant improvements:  increased performance on
compare-and-swap operations, the return of hyper threading, and ridiculous
memory bandwidth per CPU (3 DDR3 memory channels per CPU).

>> I'd be very interested to see the results of your research and benchmark=
s
>> posted here on pgsql-performance.
>=20
> Me too.  I'm gonna spend some time this summer benchmarking and tuning
> the database servers that I pretty much had to burn in and put in
> production this year due to time pressures.
>=20
> --
> Sent via pgsql-performance mailing list (pgsql-performance@postgresql.org=
)
> To make changes to your subscription:
> http://www.postgresql.org/mailpref/pgsql-performance
>=20