Hi Ilyas,
> 1 CPU X 1
> 2 CPU X 1.88
> 4 CPU X 3.49
> 8 CPU X 5.78
>
> I have to say, I am very disappointed with the results. I had the same
> question a year ago, and Ross Walker gave a really nice answer. I am
> surprised that we cannot, still, get a full performance due to the bad cpu
> architecture.
"You cannot change the laws of physics captain". The chips were c**p when
they came out and they are still c**p. Think of it as trying to force the
flow from an entire fire hydrant through a straw.
Intel's new Quick path coming out next year might help a bit.
A few things to try. 1) Use PMEMD, 2) Try using Intel's MPI, 3) Set the
processor affinity so that processes get locked to a specific core - You can
do this with Intel MPI but I forget the exact options needed.
Also with PMEMD try running with 6 or 7 mpi threads. One of the problems you
have is that you have a big clunking Linux OS running on this machine at the
same time and the memory bandwidth to the chips is so poor that swapping the
kernel in and out of various cores hits performance hard. Leaving one core
available for the OS to use can sometimes help. With sander.MPI you take a
hit on this though since it doesn't do well with non power of 2 cpu counts.
But PMEMD shouldn't be affected by it.
> Is it possible that if I use pmemd, I can increase the efficiency? In any
> case, just wanted to share my experience with you.
You will certainly increase 'efficiency' in terms of throughput per given
amount of wallclock time. You may or may not improve scaling depending on
the specifics of your simulation.
Note GB simulations tend to do better on the quad core chips than PME
simulations because the computation to memory read write ration is higher
with GB than with PME.
All the best
Ross
/\
\/
|\oss Walker
| Assistant Research Professor |
| San Diego Supercomputer Center |
| Tel: +1 858 822 0854 | EMail:- ross.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.scripps.edu
To unsubscribe, send "unsubscribe amber" (in the *body* of the email)
to majordomo.scripps.edu
Received on Sun Aug 31 2008 - 06:07:18 PDT
