Skip site navigation (1)Skip section navigation (2)
Date:      Wed, 18 Jul 2007 09:05:00 -0400
From:      "Jack Toering" <Jack.Toering@LeadingEdgeITA.com>
To:        <freebsd-performance@freebsd.org>
Subject:   RE: Which SMP CPU for FreeBSD 6.2?
Message-ID:  <032101c7c93c$3fe17770$6480010a@DELL9400>
In-Reply-To: <20070718120302.GA11968@voi.aagh.net>
References:  <029401c7c8af$e5d19c60$6480010a@DELL9400> <b41c75520707171458q510b7e11xc554ddb4452c38c9@mail.gmail.com> <02b101c7c8bf$442c2140$6480010a@DELL9400> <20070718120302.GA11968@voi.aagh.net>

next in thread | previous in thread | raw e-mail | index | archive | help
>You probably don't need to follow every move, but it might make sense to do
some actual research if performance matters to you?  You don't even say what
your performance critical server is going to be doing ;)<

At the moment I'm not too sure either. :)  The reason I say this is we have
a 246 Dual Opteron that is handling 15,000 unique visitors an hour pulling
data from around the world, generating weather maps, and pounding MySQL.  It
has no problems.  There is host a large hotel site on the same server that
is doing 900 and hour.  2/3 of the bandwidth is inbound to make the
calculations.  It took us a lot off tweaking to get it where it is today.  

>but K8's are still pretty good with many workloads thanks to their far
superior interconnects.<

Which is why we went with dual 246s at the time.  The second processor on
the Intel was a waste of money and it dual chip at the server level couldn't
hold a candle to their laptop chips and the cost more than AMD.  That
decision was a no-brainer.

>Most numbers I've seen have been more based on games and media encoding in
single socket configurations, and in 32bit mode at that;<

Yes, I'm quite skeptical about benchmarks even on web servers having learned
the hard way.  We tried 3 big names on Linux where the site fell down before
9AM.  A 4th, Rackspace, the salesman didn't return my calls after they had
an engineer check out what we were doing.  Before and after we tried 3
versions off Linux ourselves with lots of configurations.  Based on all of
the benchmarks, FreeBSD didn't make sense which is why it took us so long to
get here.  Our competitors are running 5 box server farms to do what we
accomplish with this setup and we have better response times.  We are
running SCSI raid.  There is no measurable user response difference between
a load of 17 or .7.  Due to optimizations, our current load doesn't exceed 6
or 7 with normal being more around 2.  

Then the NOCs hated us because we used the bandwidth they promised and we
were told to move along one way or another 3 times.  The best NOC we've ever
had is SAVVIS, Texas.  They could not handle it either where the server was
at first, but after I proved to them with a packet sniffer that there was a
problem, they worked hard for us to get us a place in the DC that could
handle it.  We've lived happily ever after there.  That's where the next one
is going also.  We've had all we can handle of being a fugitive and a
vagabond.

Naturally, my thoughts are drifting toward, "Hey, the new site should be
doing the 246s and the old site something new."  In the mean time I've kept
up off and on, on the issues with AMD code not being nearly as efficient as
the Intel code due to optimization for the Intel due to its far greater
numbers.  Next, the abosolute 180 degree change in the market between now
and then concerning AMD and Intel.  I've read the Tom's hardware and
AnandTech (I think) comparisions.  However, a lot of beautify theories get
beat up by ugly realities and I have the scars to prove it.  That's why I'm
here fishing.  I'm not expecting to find someone else with traffic and loads
like this, but any insights are welcome when it comes to rolling the dice
again because anything you can do to cut down the number of tries you need
to throw the dice is a huge help.

Thanks!




Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?032101c7c93c$3fe17770$6480010a>