As Jason suggested, try benchmarking your system before starting doing
the real calculations. It all depend on the computer system you are using
as well as the molecular system you are trying to simulate.
Not everyone is benchmarking their system prior doing the real
calculations, but I think this should be a regular protocol if the cluster
has infiniband connection. So, create your system (prmtop/inpcrd files),
and then benchmark it for short time (let's say between 10000-50000 steps)
using 2, 4, 8, 16, 32, 64, 128 cpus. I highly doubt that the scaling will
be good over 64 but you can check it for yourself. Also, use pmemd (or
pmemd.MPI depending on which version of AMBER package you have). From my
experiences, pmemd is around 1.2-1.3 faster than sander.MPI.
I do not know how the parallelization is done in pmemd or sander.MPI, but
it seems that the molecule is divided into a specific number of parts and
assigned to each core. So, the number of residues (not including the water
and ions) you have in your system will decide what the optimum number of
cores you can use (without losing too much efficiency). As an example, in
an RNA system with 34 residues well solvated with water and ions, I
decided to use 36 cores (3X12) because the efficiency was lost over 36.
So, test it for your system and decide what the best core size is for your
system.
Good luck,
Ilyas Yildirim, Ph.D.
-----------------------------------------------------------
= Department of Chemistry - 2145 Sheridan Road =
= Northwestern University - Evanston, IL 60208 =
= Ryan Hall #4035 (Nano Building) - Ph.: (847)467-4986 =
=
http://www.pas.rochester.edu/~yildirim/ =
-----------------------------------------------------------
On Tue, 20 Dec 2011, Jason Swails wrote:
> You can look on http://ambermd.org/#benchmarks for details.
>
> Note, though, that results are strongly dependent on system configuration
> and specifications (for instance, interconnect speed is crucial to parallel
> scaling and performance). The best way to know for sure, though, is to run
> short benchmarks yourself on your system. This is actually what the core
> developers themselves do since performance is so hardware-dependent.
>
> HTH,
> Jason
>
> On Tue, Dec 20, 2011 at 9:54 AM, Catein Catherine <askamber23.hotmail.com>wrote:
>
>>
>> Dear Sir/Madam,
>> I used to run sander with 8 nodes on my supercomputer facilities. I am
>> wonder if we can effectively scale up to use 16 or even 32 nodes for amber
>> calculations. I understand it is machine dependent, I am looking for a
>> guidelines to start with.
>> I am sure users or Case's research team have done preliminary tests.
>> Where can I find this official test results?
>> Best regards & many thanks,
>> Catherine
>> _______________________________________________
>> AMBER mailing list
>> AMBER.ambermd.org
>> http://lists.ambermd.org/mailman/listinfo/amber
>>
>
>
>
> --
> Jason M. Swails
> Quantum Theory Project,
> University of Florida
> Ph.D. Candidate
> 352-392-4032
> _______________________________________________
> AMBER mailing list
> AMBER.ambermd.org
> http://lists.ambermd.org/mailman/listinfo/amber
>
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Tue Dec 20 2011 - 13:00:05 PST
