[gmx-users] how to optimize performance of IBM Power 775?

[Christopher Neale]

Beyond optimizing your number of cores and npme (by hand I guess in your case), you can try hyperthreading (talk to your sysadmins if you can not simply run 2N threads, where N is the number of cores. I found on a power 6 that if you have N cores, it is optimal to run 2N-1 threads, rather than 2N threads. Nevertheless, I have found that the IBM power 5 and power 6 machines run gromacs significantly slower than, for example, a Nehalem. I don't know if it's a lack of gromacs hardware-specific optimization or simply a function of the lower power consumption of the IBM power series.

A number of years ago, I posted a few emails to this list on how to optimize compilation on the power machines (flags, etc). You should be able to find those by searching.

Good luck,
Chris.

-- original message --

   Our institute got a  IBM Power 775 cluster and it claimed to be very 
good. However, it doesn't support g_tune_pme. I use the following script 
for job submission:



#@ job_name = gromacs_job
#@ output = gromacs.out
#@ error = gromacs.err
#@ class = kdm
#@ node = 4
#@ tasks_per_node = 32
#@ wall_clock_limit = 01:00:00
#@ network.MPI = sn_all,not_shared,US,HIGH
#@ notification = never
#@ environment = COPY_ALL
#@ job_type = parallel
#@ queue
mpiexec -n 128 /opt/gromacs/4.5.5/bin/mdrun -nosum -dlb yes -v -s md.tpr

it is only 7 ns/day.

However, in another cluster with the same system, Core number and 
parameters, I can get up to 30 ns/day.

Does anybody have any advices for this issue?

thank you very much
Albert