[Pw_forum] How to choose number of pools
akohlmey at cmm.chem.upenn.edu
Thu Apr 30 16:33:07 CEST 2009
On Thu, 2009-04-30 at 07:18 -0700, Ihsan Erikat wrote:
please check out previous lengthy discussions on the subject
in the mailing list archives _again_. this topic has been discussed
repeatedly and people have done benchmarks to demonstrate and
understand how the various parallelization options help under
specific environments. particularly the number of cpus per node,
cores per cpu, main memory per cpu/core, cache memory per cpu/core
in combination with the interconnect (tcp/ip, infiniband/myrinet/...)
can lead to different results. in some cases it may even better
to not use all cpu cores in order maximize cache usage, in others
it may be important to use the more communication demanding
parallelization scheme to reduce the per process memory use.
... and this is only the beginning. compiler vendor and flags
as well as BLAS/LAPACK and communication libraries can have
and impact, not to mention the setup and configuration of machines.
(e.g. modern linux distributions come with SELinux enabled which
effectively trashes the cache by randomizing the stack frame location
and lots of useless auto-configuration demons running that interrupt
the kernel needlessly and thus create a lot of OS jitter which
affects latency sensitive applications. running a "lean mean computing
machine" requires some effort on many levels...).
do tests and figure it out yourself. what works best depends
very much on your input set size, network, hard- and software.
> Hello all
> I am running pwscf on 10 CPU and the k points in my file is 3x6x1.
> When I use the command mpirun -np xx pw.x -npool x , how can I
> choose the effective number of processors and pools . I read the
> manual and the forum archive, but I don`t know how to choose these
> numbers. I`ll appreciate any help.
> best regards
> Ihsan Erikat
> Physics department-Jordan University
> Pw_forum mailing list
> Pw_forum at pwscf.org
Axel Kohlmeyer akohlmey at cmm.chem.upenn.edu http://www.cmm.upenn.edu
Center for Molecular Modeling -- University of Pennsylvania
Department of Chemistry, 231 S.34th Street, Philadelphia, PA 19104-6323
tel: 1-215-898-1582, fax: 1-215-573-6233, office-tel: 1-215-898-5425
If you make something idiot-proof, the universe creates a better idiot.
More information about the users