Dividing a 2D ABF simulations into regions (on nodes with different speeds)

From: Ajasja Ljubetič (ajasja.ljubetic_at_gmail.com)
Date: Fri Jul 29 2011 - 04:44:38 CDT

Dear all,

I have a 2D free energy surface that I would like to sample using ABF.
Naturally I would like to speed up the sampling by dividing the the 2D
colvar space into regions. But I have four GPUs with different speeds*.
Since the system is small (25k atoms) I'm always running one simulation on
one GPU. So now I have (at least) four options (rated from best to worst in
my opinion):

   1. divide the surface into four unequal parts and run the smaller parts
   on the slower GPUs. (so all parts will be sampled aprox. equally - but
   temperatures will rise in the next week so I don't know the relative speeds
   of the GPUs: )
   2. divide the surface into two equal parts and run the same part on the
   slow and fast GPU (so all parts will be sampled equally)
   3. run the whole 2D surface on all GPUs (so that the whole region will
   be sampled equally)
   4. divide the surface into four equal parts (but some parts will have
   worst sampling then the other, since some GPUs are slower)

Did I asses the situation correctly? I can't decide between option 1 and 2.
Splitting the FES into smaller parts should help diffusion explore the
regions better,
but when combining the various regions it's probably best that all parts
have similar sampling.

Thank you for your advice and experience.
Best regards,
Ajasja Ljubetic
Young researcher
Laboratory of biophysics
Department F5
Institute Jozef Stefan
Ljubljana, Slovenia

*Actually the GPUs are the same, but since it's summer some are overheating,
and if that happens I suspend NAMD on that GPU for a short while, which
effectively means that some GPUs are slower than the others.

This archive was generated by hypermail 2.1.6 : Wed Feb 29 2012 - 15:57:31 CST