AMBER Archive (2007)

Subject: Re: AMBER: amber on AMD opteron-250

From: Robert Duke (rduke_at_email.unc.edu)
Date: Wed Dec 05 2007 - 14:35:04 CST

Servaas,
There is way too little info here for us to be of any help. PMEMD may not
scale on a cluster for a wide number of reasons. It may be that the run is
of a type that does not scale particularly well (though what you are seeing
here is substantially worse than I would anticipate). It could be the
hardware choice (ie., gigabit ethernet does rather poorly) or how the
hardware/system software are set up (lots of notes about that sort of thing
pointed to by links on the amber webpage amber.scripps.edu). Or is the
hardware simultaneously being used by others? Or is the interconnect that
mpich uses also being used for general net traffic, or even worse, NFS? So
those are issues you have to address. Do the mpi installation tests run?
Then, how does one of the amber benchmarks scale? (see the benchmarks tree
in the amber release). Then, how does your scaling compare to what we
publish at amber.scripps.edu for similar hardware? Processor order should
make a small difference, assuming you have multiple cpu nodes (shared memory
vs. net comm), and in general you want the "on-node" cpu's clustered
together, but I would think you would have to work to screw that up.
Getting a cluster running properly is not a job for a casual computer user
(which you may not be, but I have no way of knowing, so based on available
input I cannot rule anything out).
Regards - Bob Duke

----- Original Message -----
From: "servaas michielssens" <servaas.michielssens_at_student.kuleuven.be>
To: <amber_at_scripps.edu>
Sent: Wednesday, December 05, 2007 3:10 PM
Subject: AMBER: amber on AMD opteron-250

>I ran a 20ps simulation of a system of 20000 atoms on an AMD opteron 250
> cluster with 8 processors, I used amber8 and pmemd for the simulation. I
> found some strange results:
> proc time(min)
> 2 31
> 3 29
> 4 20
> 5 23
> 6 24
> 7 20
> 8 21
>
> 4 processors gives the optimum, it seems to be independent of how I
> adress the processors. So for 5 processors 1-2-3-4-5 or 1-2-3-4-7 gives
> the same results, always on for processors there is an optimum. Anyone
> who experienced this scaling problem?
>
> kind regards,
>
> servaas michielssens
>
> -----------------------------------------------------------------------
> 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