CPU number overrided before starpu_mpi_init_conf
Somehow related to 1ae29301: after this commit I noticed that when I define STARPU_NCPU=10
(the machine has 24 hyperthreaded cores), only 9 threads are used by Chameleon tests.
Maybe one part of the problem comes from StarPU side, but the problem from Chameleon side is that it overrides the value of conf->ncpus
with -1 just before calling chameleon_starpu_init()
and then starpu_mpi_init_conf()
. Thus a core is reserved for communications by StarPU, which is incorrect since I reserved one core for MPI by reducing the number of workers for StarPU (one for StarPU scheduler, one for MPI: 12-1-1=10)
Is it a bug in Chameleon ? Or do we need to change only the StarPU's behaviour (ping @furmento) ?
Note: it's working well if I use STARPU_RESERVE_NCPU=2
.