Skip site navigation (1)Skip section navigation (2)
Date:      Sun, 20 Jun 1999 10:05:44 -0600 (MDT)
From:      Jonathon Doran <doranj@Colorado.EDU>
To:        freebsd-questions@freebsd.org
Subject:   Re: SMP, 4GB RAM, 4x CPU
Message-ID:  <199906201605.KAA00945@ucsu.Colorado.EDU>
In-Reply-To: <Pine.BSF.4.10.9906201414380.28322-100000@ipamzlx.physik.uni-mainz.de> from "O. Hartmann" at Jun 20, 99 02:34:47 pm

next in thread | previous in thread | raw e-mail | index | archive | help
Benchmarking is an art (a black art, some say).  I believe that the best
benchmark is your application, at least you know what you are measuring.

First, lets go over SMP a bit.  SMP schedules at the process level, so
if your Fortran program isn't parallel to start with, it won't be split
magically across processors.  This shouldn't be surprising, the system
has no idea where sync points should be, and if it did split jobs then
one processor would be accessing data the other processor was computing.

With a little work you can split most scientific programs up into multiple
processes.  (As an aside, Multiflow had a tracescheduling compiler which
was proven to obtain all the parallelism which could be detected at compile
time.  Intel bought them out after Multiflow folded).

On the 4G of RAM issue, what motherboard/chipset supports this much?  The
440GX chipsets will support 2G of RAM, and that is the max that I am aware
of.  While tons of RAM is nice to have, the paging system does a pretty
good job, and assuming local access patterns (your program isn't touching
all pages all the time), you can get good performance with less RAM.

On the other hand, cache is king.  The Alphas have shown this.  At work,
we use Alphas exclusively, since they give us the best throughput for our
task (training our speech recognizer).  The 2M Xeon's are _starting_ to
get into a decent realm of cache, however the Alphas we use have 4M and are
not the top of the line (are several years old).  Given you wish to stay
with Intel, then I certainly recommend large cache Xeons.  Not the 512k
versions, go for 2M.  Since these Xeons are around $3000 per CPU the last
time I checked, the Alphas are cheaper.  Your objections to the cost of
the Alphas don't quite make sense at this point.

> workgroups with a limited budget ...). For such enormous numbercrunching 
> purposes we considered to buy (money rules, sorry, no Alpha!) a 4x SMP
> machine with 2 or 4GB RAM to server the necessary performance and to keep 
> costs low -> so we have to rely on a free UN*X and I like FBSD because
> of its stability ... and I like BSD style systems. But that is not the question.
> We heard about tests, tests and tests again, made with the new Linux kernel (2.2x)
> and many provider offering so called "number crunching" Linux systems with
> 2 CPUs ( P III/550MHz). 

The fastest CPU isn't always the best choice for number crunching.  If your
app parallelizes well, then 4x PII450 will cost less and crunch more.  Consider
buying more than one SMP system, and running half your dataset on each box.
The non-Xeon flavor of these are running around $250 each.  As Grace Hopper
said, you don't buy a bigger ox, you buy more oxen.

> I read a lot about problems with DRAM growing up to 4GB and problems with 4 CPUs.
> I have problems with two - how big are then problems and "performance losses"
> with four CPUs ...

Big problem:  I haven't seen a motherboard supporting 4G of RAM.  I'd love
to see one.  If anyone has a pointer, please share it so I don't mislead
people in the future. 

> maketime of 90 to 100 minutes. That's funny, isn't it? Well, I tried 
> make -j8, make -j12, make -j16 and lowered it to make -j5, but always the same
> result - and be aware of the fact, that the system is not used in the time
> of making world!!!!

You've discovered what I said before, the best benchmark is your application.
Try running your Fortran app on these boxes.

> 
> how "ggod" the SMP implementation of FreeBSD 3.2 is, how "stable" and usable
> the system is for usage with 4x CPU (Xeon) and 4GB RAM. We have some offers
> of Fortran vendors, and I don't want me spending a lot of money for a Linux-
> emulation to get not the power of the native system running on a Linux box.
> Where is the FreeBSD-SMP Roadmap? What has changed in FBSD 4.0?

I've been running SMP under FreeBSD for 3+ years now.  (I reported some problems
when I moved up to 3.2, but the thought is that its heat.  This is a good
thought, especially when I opened the case and found it uncomfortable to place
my hand on any metal parts :-)

Since SMP is so important to your application, I really recommend you read
the Intel SMP spec.  Pointers off:
	http://www.freebsd.org/~fsmp/SMP/SMP.html

Jon Doran


To Unsubscribe: send mail to majordomo@FreeBSD.org
with "unsubscribe freebsd-questions" in the body of the message




Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?199906201605.KAA00945>