[MITgcm-support] Flops?

[Constantinos Evangelinos]

On Wednesday 02 November 2005 17:11, Paola Cessi wrote:

> How do I find out the Flops achieved in a computation?

a) You run the code on an IBM SP with hpmcount installed and proceed as 
follows:

serial)    hpmcount -o some_prefix ./mitgcmuv 
parallel) poe /usr/local/bin/hpmcount -g $NPROCS -o some_prefix ./mitgcmuv

b) On a platform with PAPIex installed you do

serial)    papiex -f some_prefix ./mitgcmuv
parallel) mpirun -np $NPROCS `which papiex` -f some_prefix -- ./mitgcmuv

c) On a platform with PerfSuite installed

serial)   psrun ./mitgcmuv
	    psprocess psrun.*.xml
parallel) mpirun -np $NPROCS `which psrun` -f ./mitgcmuv
            psprocess –c psrun.*.xml > combined.xml
            psprocess combined.xml

d) You wait a little for some extra testing and you will be able to use the 
builtin PAPI and IBM LIBhpm capabilities that I've coded in MITgcm.
	
Constantinos

-- 
Dr. Constantinos Evangelinos
Department of Earth, Atmospheric and Planetary Sciences
Massachusetts Institute of Technology