We used to run MPI progression from workers, and thus had to wake all workers when we queue a new request. But we currently run everything from the MPI thread, and already wake it. Waking everybody is now pure overhead.
@@ -1101,8 +1101,6 @@ static void _starpu_mpi_handle_detached_request(struct _starpu_mpi_req *req)
_starpu_mpi_req_list_push_back(&detached_requests, req);
STARPU_PTHREAD_MUTEX_UNLOCK(&detached_requests_mutex);
- starpu_wake_all_blocked_workers();
-
STARPU_PTHREAD_MUTEX_LOCK(&progress_mutex);
STARPU_PTHREAD_COND_SIGNAL(&progress_cond);
STARPU_PTHREAD_MUTEX_UNLOCK(&progress_mutex);