Hi Vivek,
if you use separate PME nodes (-npme) then one group of the processors
will calculate the long-range (reciprocal space) part while the
remaining processors do the short-range (direct space) part of the
Coulomb forces. The goal is to choose the number of nodes in both
groups such that the LR calculation takes the same amount of time as
the SR calculation - otherwise one of the groups just does nothing for
part of the time step.
Also it is advisable not to have too many nodes in the LR group - a
quarter or a third is a good number as pointed out by Justin and
Berk. This you can achieve by shifting work between LR and SR part by
enlarging the grid spacing and at the same time enlarging the Coulomb
radius.
Carsten
On Nov 11, 2008, at 12:06 PM, vivek sharma wrote:
Hi Carsten,
I have also tried scaling gromacs for a number of nodes ....but was
not able to optimize it beyond 20 processor..on 20 nodes i.e. 1
processor per node..
I am not getting the point of optimizing PME for the number of
nodes, is it like we can change the parameters for PME for MDS or
using some other coulomb type for the same, please explain.
and suggest me the way to do it.
With Thanks,
Vivek
2008/11/10 Carsten Kutzner <[EMAIL PROTECTED]>
Hi,
most likely the Ethernet is the problem here. I compiled some
numbers for the DPPC
benchmark in the paper "Speeding up parallel GROMACS on high-latency
networks",
http://www3.interscience.wiley.com/journal/114205207/abstract?CRETRY=1&SRETRY=0
which are for version 3.3, but PME will behave similarly. If you did
not already use
separate PME nodes, this is worth a try, since on Ethernet the
performance will drastically
depend on the number of nodes involved in the FFT. I also have a
tool which finds the
optimal PME settings for a given number of nodes, by varying the
number of PME nodes
and the fourier grid settings. I can send it to you if you want.
Carsten
On Nov 9, 2008, at 10:30 PM, Yawar JQ wrote:
I was wondering if anyone could comment on these benchmark results
for the d.dppc benchmark?
Nodes Cutoff (ns/day) PME (ns/day)
4 1.331 0.797
8 2.564 1.497
16 4.5 1.92
32 8.308 0.575
64 13.5 0.275
128 20.093 -
192 21.6 -
It seems to scale relatively well up to 32-64 nodes without PME.
This seems slightly better than the benchmark results for Gromacs 3
on www.gromacs.org.
Can someone comment on the magnitude of the performance hit and
lack of scaling with PME is worrying me.
For the PME runs, I set rlist,rvdw,rouloumb=1.2 and the rest set to
the defaults. I can try it with some other settings, larger spacing
for the grid, but I'm not sure how much more that would help. Is
there a more standardized system I should use for testing PME
scaling?
This is with GNU compilers and parallelization with OpenMPI 1.2.
I'm not sure what we're using for the FFTW The compute nodes are
Dell m600 blades w/ 16GB of RAM and dual quad core Intel Xeon 3GHz
processors. I believe it's all ethernet interconnects.
Thanks,
YQ
_______________________________________________
gmx-users mailing list gmx-users@gromacs.org
http://www.gromacs.org/mailman/listinfo/gmx-users
Please search the archive at http://www.gromacs.org/search before
posting!
Please don't post (un)subscribe requests to the list. Use the
www interface or send it to [EMAIL PROTECTED]
Can't post? Read http://www.gromacs.org/mailing_lists/users.php
_______________________________________________
gmx-users mailing list gmx-users@gromacs.org
http://www.gromacs.org/mailman/listinfo/gmx-users
Please search the archive at http://www.gromacs.org/search before
posting!
Please don't post (un)subscribe requests to the list. Use the
www interface or send it to [EMAIL PROTECTED]
Can't post? Read http://www.gromacs.org/mailing_lists/users.php
_______________________________________________
gmx-users mailing list gmx-users@gromacs.org
http://www.gromacs.org/mailman/listinfo/gmx-users
Please search the archive at http://www.gromacs.org/search before
posting!
Please don't post (un)subscribe requests to the list. Use the
www interface or send it to [EMAIL PROTECTED]
Can't post? Read http://www.gromacs.org/mailing_lists/users.php
_______________________________________________
gmx-users mailing list gmx-users@gromacs.org
http://www.gromacs.org/mailman/listinfo/gmx-users
Please search the archive at http://www.gromacs.org/search before posting!
Please don't post (un)subscribe requests to the list. Use the
www interface or send it to [EMAIL PROTECTED]
Can't post? Read http://www.gromacs.org/mailing_lists/users.php