Well, the scaling looks okay here, but it is only out to 8 cpu, so that does
not tell you a lot. I am peripherally aware of some activity to get pmemd
and amber on sun machines, but have not had direct access to anything myself
(including ranger, about which there was another question recently, though I
think I may get onto that machine before too long). Here's the deal. First
of all, a 700 MHz cpu is pretty darn slow these days, and generally will be
beat senseless in about all regards including price by the pre-multicore
pentiums at 3+ GHz. There is another wrinkle here when one says "Sun". Sun
is selling high-end hpc systems these days with both ultrasparc chips and
amd opterons. Since opterons are fairly similar to pentiums at the machine
instruction level, they optimize similarly, and I have had extensive access
to both. So chances are we will optimize well on something like ranger
which I believe is an amd opteron-based system. The ultrasparc may or may
not be another story. Other risc chips like the sgi mips and hp alpha have
gradually been left in the dust by intel ia32-like systems (opteron,
pentium, etc), itanium 2, and powerpc. I just don't have the experience
with Sun to know where they are at, but I do know they are putting a big
push into hpc, and that they are working hard on scaling. Eventually, it
would be good if I got my hands on an ultrasparc if it is worth messing
with. The thing about optimization to a particular chip - it is touchy. So
pmemd compiles something like 12 different ways to best fit the vagaries of
different chip architectures. So choosing the wrong collection of
conditionally defined code can cost you maybe 30-50%, just off the top of my
head (I know I have some optimizations that are good for as much as 30% on
some platforms, and no help at all on others). Maybe the Sun and Fujitsu
guys will come out of the woodwork and comment on this stuff too - I have
been asked questions in the past by them, indicating they were working on
getting amber running on their hardware, and I think I have seen a fujitsu
whitepaper out there somewhere (might try googling - sorry, but I don't have
it at my fingertips).
Regards - Bob Duke
----- Original Message -----
From: "Atro Tossavainen" <atro.tossavainen+amber.helsinki.fi>
To: <amber.scripps.edu>
Sent: Monday, March 10, 2008 4:04 AM
Subject: Re: AMBER: Benchmarks on Fujitsu PrimePower 2500?
>> Does anyone have absolute benchmarks (that they are willing to share) for
>> AMBER 8 or 9 parallel molecular dynamics (PMEMD and/or sander) running on
>> a
>> Fujitsu Primepower 2500 (SPARC64) system?
>
> I have experimented with AMBER 8 on a SPARC64 platform (Primepower 200,
> single 700 MHz SPARC64 IV) and can run the same tests on another one
> (Primepower 250, single 1.1 GHz SPARC64 V). These are single-CPU
> machines only, so I can't get you any data on scaling with multiple
> CPUs. I've used the same binary I compiled for our Suns with the
> Sun Studio compilers; I haven't got anything Fujitsu-specific.
>
> As I see it, the SPARC CPUs are really quite badly suited for this kind
> of work. If the interconnect is really spectacular and scaling is good,
> it may help balance the CPU shortcomings somewhat in large configurations.
> As an initial data point, the SPARC64 IV at 700 MHz takes roughly twice
> the time that a Pentium III Xeon at 700 MHz does to do the JAC benchmark
> on sander.
>
> Are you aware of
>
> http://anusf.anu.edu.au/collaborations/amber_on_fujitsu/benchmarks/amber_benchmarks.html
>
> ?
>
> --
> Atro Tossavainen (Mr.) / The Institute of Biotechnology at
> Systems Analyst, Techno-Amish & / the University of Helsinki, Finland,
> +358-9-19158939 UNIX Dinosaur / employs me, but my opinions are my
> own.
> < URL : http : / / www . helsinki . fi / %7E atossava / > NO FILE
> ATTACHMENTS
> -----------------------------------------------------------------------
> The AMBER Mail Reflector
> To post, send mail to amber.scripps.edu
> To unsubscribe, send "unsubscribe amber" to majordomo.scripps.edu
>
-----------------------------------------------------------------------
The AMBER Mail Reflector
To post, send mail to amber.scripps.edu
To unsubscribe, send "unsubscribe amber" to majordomo.scripps.edu
Received on Wed Mar 12 2008 - 06:07:20 PDT
