No time saving with OpenMP on linux

[rookie2010]

Hi there,

My company developed a fortran code with OpemMP and we built 32-bit packages on Windows* (Vista Business 32-bit
with ifort 10.0.025) and linux! (SuSE Enterprise Server 9, 64-bit with ifort 11.1.056).

For identical case, we ran on Windows* (2 dual-core Intel CPUs - 4 total CPUs with 4 GB RAM) and linux ! (2 AMD
CPUs with 4GB RAM). We use input file to set up number of CPUs for execution.

On Windows*, 2 CPUs ran 8min and 38 sec; 1 CPU ran 12 min and 40 sec. On linux!, there is NO SAVING at all with
2 CPUs. Sometimes 2 CPUs job even ran longer than 1 CPU job.

I've tried to set OMP_NUM_THREADS to 2 on linux! but the situation (NO SAVING with 2 CPUs) did not change.
I've checked and ulimit for stack is unlimited.

Do I have to do any setting/configuration with AMD CPUs on linux? Any opinion is very much appreciated.

Reggie

[Ron_Green]

AMD should make no difference, nothing special you have to do.

How are you timing this? 'time' from command line, or cpu_time() in the code? or ??

[rookie2010]

Ron,

Thanks for the posting. I raised AMD issue because I read on Wikipedia "Some vendors
recommended setting the processor affinity on OpenMP threads to associate them with
particular processor cores." (http://en.wikipedia.org/wiki/OpenMP)

The way I got execution time is to submit jobs on linux using script like:
date
code_exe input_1cpu.inp
date
code_exe input_2cpu.inp
date

code_exe willuse 1st arguement as input file specifying 1 or 2 CPUs. And Ievaluate overall time
from date displays since code_exe runs on the foreground.

I just submitted the batch script which runs a different (and bigger) case. But the jobs won't
finish until later tonight. We'll see if there is any time saving for this case with 2CPUs.

Best regards,

Reggie

[jimdempseyatthecove]

Reggie,

At the beginning of your program, insert code to obtain the number of threads

!$omp parallel
!$omp master
write(*,*) "Number of threads ", omp_get_num_threads()
!$omp end master
!$omp end parallel

This will tell you if (on linux) you are indeed getting the number of threads you request.

If this indicates 1 thread then the potential problems are:

a) you linked in the wrong libraries
b) you compiled without -openmp
c) something in your Linux configuration is restricting the number of threads

Jim Dempsey

[rookie2010]

Thanks a lot for the help from Ronald and Jim.

The overnight jobs show reasonable results. 1 cpu used 4 hr 16 min and 2 cpus used 2 hr 25 min.
I'm happy as long as openmp helps with multiple CPUs.

Shorter jobs with 1 and 2 cpus did not show reasonable results could be due to the load at job
execution is large and 2 cpus did not gain advantage. I've tried various combination of flags
(-openmp for all codes, not just those with openmp directives; with or without OMP_NUM_THREADS;
differentulimit) and learned quite a bit.

I'm glad to know (from Ronaldand now confirm) that ifort does not discriminate AMD cpus and
alsoI also appreciate the tips from Jim.