Hi HPC Forum,
When I run the hybrid model on the AIX, Pwr6, the number of MPI tasks are only considered by the LoadLeveller not the OpenMp/pthreads when it is calculating the total number of JOBS for a class. Though the LL allocates the nodes for the OpenMp/pthreads, the MPI tasks are only taken into account.
For a certain class there are 8 nodes allocated.
1) 128 MPI tasks if allocated takes a max. of 2 nodes : which correctly handles by LL
I can run 4 (2 nodes x 4) such type of jobs at a time.
2) If I have 128 MPI x 2 Threads = 256 Procs allocates a max. of 4 nodes: Which correctly handles by LL
For this also I can run 4 (4 nodes x 4) such type of jobs at a time (actually we cannot run more than 2 JOBS simultaneously).
The problem is here, the LL is not able to remember/identify the threading in the previous JOB. Because of this, 16 nodes are being used of which the quota was only 8 nodes.
Is there a solution to this problem?
I have checked it with my model many times and I can exceed the total no. nodes allocated by the admin under the present circumstances?
In fact if this is a BUG! then I have few more which are in the order...