Multithread

[cristian-vasile.achim]

Hello,

I am pretty new at this. In the _compile.xml script there is code

 <LDFLAGS>
    <append compile_threaded="true"> -fopenmp </append>
 </LDFLAGS>

Is the OpenMP support activated by default, or do I need to change some setting when I create a new case? I know that in the _pes.xml there is the variable nthrds to be changed, but is that enough?

Best,

Cristian

 

[jedwards]

Yes - there are two ways to enable openmp. One is simply to set the value of any NTHRDS_ variable > 1.
The other is to set the xml variable FORCE_BUILD_SMP=TRUE. This second is because many compilers change
the optimization settings when using openmp and in testing we compare runs using different numbers of threads and expect
the same results. In order to insure that we must sometimes compile with openmp enabled even when NTHRDS=1.

Also note that, in most cases, as long as you have not exceeded the maximum allowed number of mpi tasks - adding mpi tasks is
always more performant than adding threads.

 

[cristian-vasile.achim]

Thanks for your reply. That's great.
I tried on our cluster Mahti (with 2 x64 cores per node) and got an increase in the throughput by about 1.3 when using 128 mpi tasks per node, vs 64 tasks per node + nthrds =2 , but that was all.

Cristian

 

[cristian-vasile.achim]

Sorry I meant I got an increase in the throughput by about 1.3 when using 64 tasks per node + nthrds =2 compared to 128 mpi tasks per node.