lsb_hostinfo

Add additional field to show Total CPU per host and Total CPU in use per host. Currently sees the number of "tasks".

Our real issue is we have started using a combination of OpenMP/MPI for some of our users jobs. When running a standard OpenMP or MPI job, the “BSUB -n” value correctly shows the number of tasks/slots. The issue is with these new hybrid jobs. Since we are no longer referencing ‘BSUB -n' as tasks/slots but rather just the number of tasks, and utilizing the ptile/OMP_NUM_THREAD to calculate out the number of CPUs being used.

The reason we are pushing on the affinity is by matching the BSUB -R Affinity to the OMP_NUM_THREAD we can utilize this additional information to resolve our issue. Again affinity is not our “true” issue. Our issue is calculating the total number of CPUs that are in use on each host. We originally did this by the -n or the ‘numRun' in the lsb_hostinfo. As this was a 1:1 statement, this worked, but it does not with the new jobs being ran in the environment.

I have attached two documents. Both jobs are using 24 CPU. The Parallel job is the older OpenMP/MPI that “BSUB -n” is both the number of tasks and slots. The Hybrid job is the newer method where “BSUB -n” is the number of tasks, and we need to set additional argument and variables to finish the calculation.

If we look at the just the output of a bjobs
Hybrid: 11461 cv7457 RUN standard cirrustest- 2*compute HELLO Jun 14 14:12
Parallel: 11460 cv7457 RUN standard cirrustest- 24*compute abaqus Jun 14 14:12

While both jobs are running 24 CPU, the hybrid is only showing 2 tasks, from the BSUB -n 2. If we run our modified bjobs command, the command looks at the number of affinity being ran for this job, in this case 24, and reports 24 instead of 2.