GuanLab/Leopard

Control threads number

Opened this issue · 2 comments

Is there a way to control a thread number used by the program? When submitting jobs to LSF cluster, every queue has the maximum number of threads allowed for the job, so if the program uses more then allowed, it gets killed

Is it possible to request a limited number of CPUs at the job submission step for LSF? For example, when I run PBS jobs, I can request resource based on my need as follows:
#PBS -l nodes=1:ppn=4,mem=20gb
I don't have too much experience in multi-thread programming :( and not quite sure about how to control the thread number..

The threads number is not the same as CPUs or processes number. It's more like another level of parallelization that, I guess, we cannot control in TensorFlow. In my case, I had a limitation of 200 threads, but Leopard needed 593

Wed Dec 11 23:02:28: Resource usage collected.
                     MEM: 33 Gbytes;  SWAP: 83 Gbytes;  NTHREAD: 593
                     PGID: 21493;  PIDs: 21493 21505 21509 21591 21623 21624 
                     21625 21626 21627 21628 

Btw, what is the best configuration to run Leopard. I mean, CPU, RAM, nodes number?