[starpu-devel] Creating thread partitions

[Doru Adrian Thom Popovici <dtpopovici@lbl.gov>]

Hello

I was reading the documentation, and in there it was said that if a context is created without a scheduling policy, then that context basically is a parallel worker.

I tried to implement something as follows:

int ncpus = starpu_cpu_worker_get_count(); // p * t
  int *cpus = (int*) malloc(ncpus * sizeof(int));
  starpu_worker_get_ids_by_type(STARPU_CPU_WORKER, cpus, ncpus);

  int sched_ctx[p];
  for(int i = 0; i < p; ++i) {
    sched_ctx[i] = starpu_sched_ctx_create((cpus + i * t), t, "subctx", 0);
  }

  int big_sched_ctx = starpu_sched_ctx_create(fcpus, p * t, "ctx1", STARPU_SCHED_CTX_SUB_CTXS, sched_ctx, p, STARPU_SCHED_CTX_POLICY_NAME, "lws", 0);

Based on my understanding, I got all the cpus on the system. From there I create some sub-contexts that do not have any scheduling policy. Basically I partition the CPUs in p groups. Each group having t cpus.

Then I create a master context that has a policy. 

If I submit tasks to the big_sched_ctx context, each task is transmitted to the sub-contexts? Is this  assumption correct? I just want p parallel workers with t threads each. I hope I got the documentation correctly.

Best regards,

Doru