AMBER Archive (2006)

Subject: Re: AMBER: Opteron vs. Xeon performance

From: ian gould (i.gould_at_imperial.ac.uk)
Date: Fri May 12 2006 - 02:35:12 CDT

Hi Everyone

Thought I'd throw my two pennyworth in on this issue. Cann't give you em64t
numbers but can furnish some opteron numbers which may be of use.

Machines Used

Iwill 8501
8 x 870 (2.0Ghz) = 16 cores
32 Gb Ram
1Tb sata raid

HP proliant
4 x 885 (2.6 Ghz) = 8 cores
16Gb Ram
250Gb sata raid

Both running RH AS 4

PMEMD 9 has been compiled with ifort 9.0 and MKL 8.0 and LAM
I have yet to get a successful compilation of MPICH2 to work

JAC numbers are ps per day

No. processors Iwill ( 2.0GHZ) HP (2.6Ghz)
2 417.39 552.30
4 752.88 987.95
8 1331.69 1757.83
16 1893.49

Factor IX ps per day

No. processors Iwill ( 2.0GHZ) HP (2.6Ghz)
2 190.50 249.22
4 324.57 421.94
8 547.21 721.37
16 810.00

This will hopefully give a flavour of what's possible. I will in the future
do the complete set of benchmarks as for the other big tin on the web site.
One comment I'd like to pass on is that whilst I have not got pmemd to
compile with the pathscale compiler/amd pathscale maths libraries at the
moment, busy with other things, I can tell you that this combination gives
my fastest performance of gaussian03 on my opteron tin. I'm getting anywhere
from 1.1 to 1.4 times slower that an altix 350 with 1.5 Ghz's compared with
my 2.0ghz iwill and on the HP it beats the SGI on all jobs!!, if you check
SGI's report on opteron vs altix they get > 2x slower performance with
faster processor in their opteron vs. altix 350 1.5ghz. So it may be that it
will give comparable/if not better performance on pmemd than intel/mkl.

BTW Ross/Bob could you shoot me the cellulose benchmark and I'll give that a
whirl.

Cheers
Ian

I've seen things you people wouldn't believe. Attack ships on fire off the
shoulder of Orion. I watched C-beams glitter in the dark near the Tannhauser
gate. All those moments will be lost in time, like tears in rain. Time to
die.
Roy Batty ... Bladerunner 

-- 
Dr Ian R Gould
Senior Lecturer Biological and Biophysical Chemistry
Imperial College London
Exhibition Road
London
SW7 2AY
E-mail i.gould_at_imperial.ac.uk
http://dumb.ch.ic.ac.uk/~igould
Tel +44 (0)207 594 5809

On 12/5/06 03:02, "Robert Duke" <rduke_at_email.unc.edu> wrote:

> Okay, I do have opteron vs. intel em64t numbers; for the specific opteron
> and intel chips I run on they are very close for pme simulations using pmemd
> on the benchmarks I have done (pmemd 9).  The specifics are 2.2 GHz opteron
> dual cpu/infiniband interconnect (note, NOT dual core, which may actually
> give you more bang for the buck, though my numbers on this are loose (just
> based on some rough benchmarks on dual core cray xd1, which I have not done
> a lot of analysis on)) vs 3.6 GHz em64t intel xeon, 2 MB cache/infiniband.
> On the two clusters I bench on for these types of chip, the numbers are
> pretty darn close for pme.  For GB, I would probably tend to prefer the xeon
> because I am a little bit chicken about trying to use the Intel MKL with
> Opterons, and it really helps GB performance - off the top of my head at
> least 20% (this is a very crude recollection from GB which I hacked into
> pmemd but have not yet put a lot of effort into optimizing - but it was
> clear that the MKL was worth it.   I don't know if the AMD equivalent is as
> good, and have not tried crossing manufacturers any further than using ifort
> to compile on opterons (which is maybe a little slower than pathscale unless
> you get adventurous in lying to the compiler about what the chip is (Andreas
> Svrcek-Seiler sent me some numbers on this, and has tried lots of different
> things in compiling on the opteron; I tend to try to get a few things
> working and move on to the next platform).  Okay, so the performance is
> close, and I could believe that a really well configured opteron can do a
> bit better, but I have not seen dramatic differences myself.  Don't get me
> wrong, I like the opteron very much, but the later intel chips are
> delivering too.  Now, if I were buying hardware, I would be worried about
> getting a reliable cpu + reliable motherboard + reliable interconnect.  The
> thing that will kill you is if you buy something cheap, and one of the
> components is not well balanced, or prone to heat-induced or other failure.
> This is stuff where it is handy to be able to ask someone who has a specific
> configuration what their experience has been (so ask the vendor for
> references, or buy from a manufacturer that is first tier at a bit of a
> premium).
> 
> At this point in time, I find developing performance code for either the
> intel ia32 family, opteron or ibm sp4/5 all to be roughly equivalent
> experiences, and the chips respond in a similar fashion to attempts at
> optimization.  RISC architectures, at least the ones I have available, are
> falling behind.  The itanium is a heck of a chip for brute power, but it is
> touchy to optimize and expensive.  I mostly develop on the ia32's and move
> off to the similar chips first, and then, while grumbling something akin to
> "what a pain" over how differently it optimizes, move off to the itanium,
> and then I finally make sure everything still runs on the RISC chips.  This
> gives you a clue or two as to what pmemd will run well on for the near
> future.
> 
> Regards - Bob Duke
> 
> ----- Original Message -----
> From: "Ross Walker" <ross_at_rosswalker.co.uk>
> To: <amber_at_scripps.edu>
> Sent: Thursday, May 11, 2006 7:17 PM
> Subject: RE: AMBER: Opteron vs. Xeon performance
> 
> 
>> Dear Joseph
>> 
>>> I am looking for recent data comparing the performance
>>> between Opterons
>>> and Xeons using Amber, as well as interpretation of that data.
>> 
>> This sort of thing is never a simple as it sounds. As with all
>> comparissons
>> it very much depends what you compare. E.g. periodic boundary PME
>> simulations on a top end Opteron will probably be better than a top end
>> Xeon, however, this may be reversed when one tries Generalized Born
>> implicit
>> solvent calculations. This is all a function of cache hits, code branching
>> etc. I.e. some problems do well on some chips but worse on others and vice
>> versa. This is certainly the case with IA64 chips.
>> 
>> E.g.
>> 
>> JAC Benchmark (PME) (PMEMD v9.0)
>> Pentium-D 3.2GHz = 231.79 ps/day
>> Itanium 2 1.5GHz = 395.37 ps/day
>> Ratio IA64/P4 = 1.71
>> 
>> GB_MB Benchmark (GB) (PMEMD v9.0)
>> Pentium-D 3.2GHz = 266.03 ps/day
>> Itanium 2 1.5GHz = 191.51 ps/day
>> Ratio IA64/P4 = 0.72
>> 
>> Anyway, I know that this is not dealing with Opterons which is what you
>> are
>> interested in but at present we don't have any amber9 timings on Opteron
>> chips. However, there are some timings here for dual core opterons:
>> http://amber.scripps.edu/amber8.bench1.html
>> 
>> JAC Benchmark (PME) (PMEMD v8.0)
>> Dual Core Opteron 875 (2.2GHz) = 187.0 ps/day
>> Single Core Opteron (1.4GHz) = 124.0 ps/day
>> 
>> All the best
>> Ross
>> 
>> /\
>> \/
>> |\oss Walker
>> 
>> | HPC Consultant and Staff Scientist |
>> | San Diego Supercomputer Center |
>> | Tel: +1 858 822 0854 | EMail:- ross_at_rosswalker.co.uk |
>> | http://www.rosswalker.co.uk | PGP Key available on request |
>> 
>> Note: Electronic Mail is not secure, has no guarantee of delivery, may not
>> be read every day, and should not be used for urgent or sensitive issues.
>> 
>> 
>> -----------------------------------------------------------------------
>> The AMBER Mail Reflector
>> To post, send mail to amber_at_scripps.edu
>> To unsubscribe, send "unsubscribe amber" to majordomo_at_scripps.edu
>> 
> 
> 
> -----------------------------------------------------------------------
> The AMBER Mail Reflector
> To post, send mail to amber_at_scripps.edu
> To unsubscribe, send "unsubscribe amber" to majordomo_at_scripps.edu

-----------------------------------------------------------------------
The AMBER Mail Reflector
To post, send mail to amber_at_scripps.edu
To unsubscribe, send "unsubscribe amber" to majordomo_at_scripps.edu