/* * StarPU * Copyright (C) Université Bordeaux 1, CNRS 2008-2010 (see AUTHORS file) * * This program is free software; you can redistribute it and/or modify * it under the terms of the GNU Lesser General Public License as published by * the Free Software Foundation; either version 2.1 of the License, or (at * your option) any later version. * * This program is distributed in the hope that it will be useful, but * WITHOUT ANY WARRANTY; without even the implied warranty of * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. * * See the GNU Lesser General Public License in COPYING.LGPL for more details. */ #include #include #include #include #define STARPU_MPI_VERBOSE 1 #ifdef STARPU_MPI_VERBOSE # define _STARPU_MPI_DEBUG(fmt, args ...) { int rank; MPI_Comm_rank(MPI_COMM_WORLD, &rank); \ fprintf(stderr, "[%d][starpu_mpi][%s] " fmt , rank, __func__ ,##args); \ fflush(stderr); } #else # define _STARPU_MPI_DEBUG(fmt, args ...) #endif /* TODO find a better way to select the polling method (perhaps during the * configuration) */ //#define USE_STARPU_ACTIVITY 1 static void submit_mpi_req(void *arg); static void handle_request_termination(struct starpu_mpi_req_s *req); /* The list of requests that have been newly submitted by the application */ static starpu_mpi_req_list_t new_requests; /* The list of detached requests that have already been submitted to MPI */ static starpu_mpi_req_list_t detached_requests; static pthread_mutex_t detached_requests_mutex; static pthread_cond_t cond; static pthread_mutex_t mutex; static pthread_t progress_thread; static int running = 0; /* * Isend */ static void starpu_mpi_isend_func(struct starpu_mpi_req_s *req) { void *ptr = starpu_mpi_handle_to_ptr(req->data_handle); _STARPU_MPI_DEBUG("post MPI isend tag %x dst %d ptr %p req %p\n", req->mpi_tag, req->srcdst, ptr, &req->request); starpu_mpi_handle_to_datatype(req->data_handle, &req->datatype); //MPI_Isend(ptr, 1, req->datatype, req->srcdst, req->mpi_tag, req->comm, &req->request); MPI_Isend(ptr, 1, req->datatype, req->srcdst, req->mpi_tag, req->comm, &req->request); TRACE_MPI_ISEND(req->srcdst, req->mpi_tag, 0); /* somebody is perhaps waiting for the MPI request to be posted */ PTHREAD_MUTEX_LOCK(&req->req_mutex); req->submitted = 1; PTHREAD_COND_BROADCAST(&req->req_cond); PTHREAD_MUTEX_UNLOCK(&req->req_mutex); } static struct starpu_mpi_req_s *_starpu_mpi_isend_common(starpu_data_handle data_handle, int dest, int mpi_tag, MPI_Comm comm, unsigned detached, void (*callback)(void *), void *arg) { _STARPU_MPI_DEBUG("--> starpu_mpi_isend_common\n"); struct starpu_mpi_req_s *req = calloc(1, sizeof(struct starpu_mpi_req_s)); STARPU_ASSERT(req); /* Initialize the request structure */ req->submitted = 0; req->completed = 0; PTHREAD_MUTEX_INIT(&req->req_mutex, NULL); PTHREAD_COND_INIT(&req->req_cond, NULL); req->request_type = SEND_REQ; req->data_handle = data_handle; req->srcdst = dest; req->mpi_tag = mpi_tag; req->comm = comm; req->func = starpu_mpi_isend_func; req->detached = detached; req->callback = callback; req->callback_arg = arg; /* Asynchronously request StarPU to fetch the data in main memory: when * it is available in main memory, submit_mpi_req(req) is called and * the request is actually submitted */ starpu_data_acquire_cb(data_handle, STARPU_R, submit_mpi_req, (void *)req); _STARPU_MPI_DEBUG("<-- starpu_mpi_isend_common\n"); return req; } int starpu_mpi_isend(starpu_data_handle data_handle, starpu_mpi_req *public_req, int dest, int mpi_tag, MPI_Comm comm) { STARPU_ASSERT(public_req); struct starpu_mpi_req_s *req; req = _starpu_mpi_isend_common(data_handle, dest, mpi_tag, comm, 0, NULL, NULL); STARPU_ASSERT(req); *public_req = req; return 0; } /* * Isend (detached) */ int starpu_mpi_isend_detached(starpu_data_handle data_handle, int dest, int mpi_tag, MPI_Comm comm, void (*callback)(void *), void *arg) { _STARPU_MPI_DEBUG("--> starpu_mpi_isend_detached\n"); _starpu_mpi_isend_common(data_handle, dest, mpi_tag, comm, 1, callback, arg); _STARPU_MPI_DEBUG("<-- starpu_mpi_isend_detached\n"); return 0; } /* * Irecv */ static void starpu_mpi_irecv_func(struct starpu_mpi_req_s *req) { void *ptr = starpu_mpi_handle_to_ptr(req->data_handle); STARPU_ASSERT(ptr); starpu_mpi_handle_to_datatype(req->data_handle, &req->datatype); _STARPU_MPI_DEBUG("post MPI irecv tag %x src %d ptr %p req %p datatype %d\n", req->mpi_tag, req->srcdst, ptr, &req->request, req->datatype); MPI_Irecv(ptr, 1, req->datatype, req->srcdst, req->mpi_tag, req->comm, &req->request); /* somebody is perhaps waiting for the MPI request to be posted */ PTHREAD_MUTEX_LOCK(&req->req_mutex); req->submitted = 1; PTHREAD_COND_BROADCAST(&req->req_cond); PTHREAD_MUTEX_UNLOCK(&req->req_mutex); } static struct starpu_mpi_req_s *_starpu_mpi_irecv_common(starpu_data_handle data_handle, int source, int mpi_tag, MPI_Comm comm, unsigned detached, void (*callback)(void *), void *arg) { struct starpu_mpi_req_s *req = calloc(1, sizeof(struct starpu_mpi_req_s)); STARPU_ASSERT(req); /* Initialize the request structure */ req->submitted = 0; PTHREAD_MUTEX_INIT(&req->req_mutex, NULL); PTHREAD_COND_INIT(&req->req_cond, NULL); req->request_type = RECV_REQ; req->data_handle = data_handle; req->srcdst = source; req->mpi_tag = mpi_tag; req->comm = comm; req->detached = detached; req->callback = callback; req->callback_arg = arg; req->func = starpu_mpi_irecv_func; /* Asynchronously request StarPU to fetch the data in main memory: when * it is available in main memory, submit_mpi_req(req) is called and * the request is actually submitted */ starpu_data_acquire_cb(data_handle, STARPU_W, submit_mpi_req, (void *)req); return req; } int starpu_mpi_irecv(starpu_data_handle data_handle, starpu_mpi_req *public_req, int source, int mpi_tag, MPI_Comm comm) { STARPU_ASSERT(public_req); struct starpu_mpi_req_s *req; req = _starpu_mpi_irecv_common(data_handle, source, mpi_tag, comm, 0, NULL, NULL); STARPU_ASSERT(req); *public_req = req; return 0; } /* * Irecv (detached) */ int starpu_mpi_irecv_detached(starpu_data_handle data_handle, int source, int mpi_tag, MPI_Comm comm, void (*callback)(void *), void *arg) { _starpu_mpi_irecv_common(data_handle, source, mpi_tag, comm, 1, callback, arg); return 0; } /* * Recv */ int starpu_mpi_recv(starpu_data_handle data_handle, int source, int mpi_tag, MPI_Comm comm, MPI_Status *status) { starpu_mpi_req req; starpu_mpi_irecv(data_handle, &req, source, mpi_tag, comm); starpu_mpi_wait(&req, status); return 0; } /* * Send */ int starpu_mpi_send(starpu_data_handle data_handle, int dest, int mpi_tag, MPI_Comm comm) { starpu_mpi_req req; MPI_Status status; memset(&status, 0, sizeof(MPI_Status)); starpu_mpi_isend(data_handle, &req, dest, mpi_tag, comm); starpu_mpi_wait(&req, &status); return 0; } /* * Wait */ static void starpu_mpi_wait_func(struct starpu_mpi_req_s *waiting_req) { /* Which is the mpi request we are waiting for ? */ struct starpu_mpi_req_s *req = waiting_req->other_request; req->ret = MPI_Wait(&req->request, waiting_req->status); handle_request_termination(req); } int starpu_mpi_wait(starpu_mpi_req *public_req, MPI_Status *status) { int ret; struct starpu_mpi_req_s waiting_req; memset(&waiting_req, 0, sizeof(struct starpu_mpi_req_s)); struct starpu_mpi_req_s *req = *public_req; /* We cannot try to complete a MPI request that was not actually posted * to MPI yet. */ PTHREAD_MUTEX_LOCK(&req->req_mutex); while (!req->submitted) PTHREAD_COND_WAIT(&req->req_cond, &req->req_mutex); PTHREAD_MUTEX_UNLOCK(&req->req_mutex); /* Initialize the request structure */ PTHREAD_MUTEX_INIT(&waiting_req.req_mutex, NULL); PTHREAD_COND_INIT(&waiting_req.req_cond, NULL); waiting_req.status = status; waiting_req.other_request = req; waiting_req.func = starpu_mpi_wait_func; submit_mpi_req(&waiting_req); /* We wait for the MPI request to finish */ PTHREAD_MUTEX_LOCK(&req->req_mutex); while (!req->completed) PTHREAD_COND_WAIT(&req->req_cond, &req->req_mutex); PTHREAD_MUTEX_UNLOCK(&req->req_mutex); ret = req->ret; /* The internal request structure was automatically allocated */ *public_req = NULL; free(req); return ret; } /* * Test */ static void starpu_mpi_test_func(struct starpu_mpi_req_s *testing_req) { /* Which is the mpi request we are testing for ? */ struct starpu_mpi_req_s *req = testing_req->other_request; //_STARPU_MPI_DEBUG("Test request %p - mpitag %x - TYPE %s %d\n", &req->request, req->mpi_tag, (req->request_type == RECV_REQ)?"recv : source":"send : dest", req->srcdst); int ret = MPI_Test(&req->request, testing_req->flag, testing_req->status); if (*testing_req->flag) { testing_req->ret = ret; handle_request_termination(req); } PTHREAD_MUTEX_LOCK(&testing_req->req_mutex); testing_req->completed = 1; pthread_cond_signal(&testing_req->req_cond); PTHREAD_MUTEX_UNLOCK(&testing_req->req_mutex); } int starpu_mpi_test(starpu_mpi_req *public_req, int *flag, MPI_Status *status) { int ret = 0; STARPU_ASSERT(public_req); struct starpu_mpi_req_s *req = *public_req; STARPU_ASSERT(!req->detached); PTHREAD_MUTEX_LOCK(&req->req_mutex); unsigned submitted = req->submitted; PTHREAD_MUTEX_UNLOCK(&req->req_mutex); if (submitted) { struct starpu_mpi_req_s testing_req; memset(&testing_req, 0, sizeof(struct starpu_mpi_req_s)); /* Initialize the request structure */ PTHREAD_MUTEX_INIT(&testing_req.req_mutex, NULL); PTHREAD_COND_INIT(&testing_req.req_cond, NULL); testing_req.flag = flag; testing_req.status = status; testing_req.other_request = req; testing_req.func = starpu_mpi_test_func; testing_req.completed = 0; submit_mpi_req(&testing_req); /* We wait for the test request to finish */ PTHREAD_MUTEX_LOCK(&testing_req.req_mutex); while (!testing_req.completed) PTHREAD_COND_WAIT(&testing_req.req_cond, &testing_req.req_mutex); PTHREAD_MUTEX_UNLOCK(&testing_req.req_mutex); ret = testing_req.ret; if (*testing_req.flag) { /* The request was completed so we free the internal * request structure which was automatically allocated * */ *public_req = NULL; free(req); } } else { *flag = 0; } return ret; } /* * Requests */ static void handle_request_termination(struct starpu_mpi_req_s *req) { MPI_Type_free(&req->datatype); starpu_data_release(req->data_handle); _STARPU_MPI_DEBUG("complete MPI (%s %d) req %p - tag %x\n", (req->request_type == RECV_REQ)?"recv : source":"send : dest", req->srcdst, &req->request, req->mpi_tag); if (req->request_type == RECV_REQ) { TRACE_MPI_IRECV_END(req->srcdst, req->mpi_tag); } /* Execute the specified callback, if any */ if (req->callback) req->callback(req->callback_arg); /* tell anyone potentiallly waiting on the request that it is * terminated now */ PTHREAD_MUTEX_LOCK(&req->req_mutex); req->completed = 1; PTHREAD_COND_BROADCAST(&req->req_cond); PTHREAD_MUTEX_UNLOCK(&req->req_mutex); } static void submit_mpi_req(void *arg) { struct starpu_mpi_req_s *req = arg; PTHREAD_MUTEX_LOCK(&mutex); starpu_mpi_req_list_push_front(new_requests, req); PTHREAD_COND_BROADCAST(&cond); PTHREAD_MUTEX_UNLOCK(&mutex); } /* * Scheduler hook */ static unsigned progression_hook_func(void *arg __attribute__((unused))) { unsigned may_block = 1; PTHREAD_MUTEX_LOCK(&mutex); if (!starpu_mpi_req_list_empty(detached_requests)) { pthread_cond_signal(&cond); may_block = 0; } PTHREAD_MUTEX_UNLOCK(&mutex); return may_block; } /* * Progression loop */ static void test_detached_requests(void) { int flag; MPI_Status status; struct starpu_mpi_req_s *req, *next_req; PTHREAD_MUTEX_LOCK(&detached_requests_mutex); for (req = starpu_mpi_req_list_begin(detached_requests); req != starpu_mpi_req_list_end(detached_requests); req = next_req) { next_req = starpu_mpi_req_list_next(req); PTHREAD_MUTEX_UNLOCK(&detached_requests_mutex); int ret = MPI_Test(&req->request, &flag, &status); STARPU_ASSERT(ret == MPI_SUCCESS); //_STARPU_MPI_DEBUG("Test request %p - mpitag %x - TYPE %s %d\n", &req->request, req->mpi_tag, (req->request_type == RECV_REQ)?"recv : source":"send : dest", req->srcdst); if (flag) { handle_request_termination(req); } PTHREAD_MUTEX_LOCK(&detached_requests_mutex); if (flag) starpu_mpi_req_list_erase(detached_requests, req); #warning TODO fix memleak /* Detached requests are automatically allocated by the lib */ //if (req->detached) // free(req); } PTHREAD_MUTEX_UNLOCK(&detached_requests_mutex); } static void handle_new_request(struct starpu_mpi_req_s *req) { STARPU_ASSERT(req); /* submit the request to MPI */ req->func(req); if (req->detached) { PTHREAD_MUTEX_LOCK(&mutex); starpu_mpi_req_list_push_front(detached_requests, req); PTHREAD_MUTEX_UNLOCK(&mutex); starpu_wake_all_blocked_workers(); /* put the submitted request into the list of pending requests * so that it can be handled by the progression mechanisms */ PTHREAD_MUTEX_LOCK(&mutex); pthread_cond_signal(&cond); PTHREAD_MUTEX_UNLOCK(&mutex); } } static void *progress_thread_func(void *arg __attribute__((unused))) { /* notify the main thread that the progression thread is ready */ PTHREAD_MUTEX_LOCK(&mutex); running = 1; pthread_cond_signal(&cond); PTHREAD_MUTEX_UNLOCK(&mutex); PTHREAD_MUTEX_LOCK(&mutex); while (running) { /* shall we block ? */ unsigned block = starpu_mpi_req_list_empty(new_requests); #ifndef USE_STARPU_ACTIVITY block = block && starpu_mpi_req_list_empty(detached_requests); #endif if (block) { _STARPU_MPI_DEBUG("NO MORE REQUESTS TO HANDLE\n"); PTHREAD_COND_WAIT(&cond, &mutex); } if (!running) break; /* test whether there are some terminated "detached request" */ PTHREAD_MUTEX_UNLOCK(&mutex); test_detached_requests(); PTHREAD_MUTEX_LOCK(&mutex); /* get one request */ struct starpu_mpi_req_s *req; while (!starpu_mpi_req_list_empty(new_requests)) { req = starpu_mpi_req_list_pop_back(new_requests); /* handling a request is likely to block for a while * (on a sync_data_with_mem call), we want to let the * application submit requests in the meantime, so we * release the lock. */ PTHREAD_MUTEX_UNLOCK(&mutex); handle_new_request(req); PTHREAD_MUTEX_LOCK(&mutex); } } PTHREAD_MUTEX_UNLOCK(&mutex); return NULL; } /* * (De)Initialization methods */ #ifdef USE_STARPU_ACTIVITY static int hookid = - 1; #endif static void _starpu_mpi_add_sync_point_in_fxt(void) { #ifdef STARPU_USE_FXT int rank; int worldsize; MPI_Comm_rank(MPI_COMM_WORLD, &rank); MPI_Comm_size(MPI_COMM_WORLD, &worldsize); int barrier_ret = MPI_Barrier(MPI_COMM_WORLD); STARPU_ASSERT(barrier_ret == MPI_SUCCESS); /* We generate a "unique" key so that we can make sure that different * FxT traces come from the same MPI run. */ int random_number; /* XXX perhaps we don't want to generate a new seed if the application * specified some reproductible behaviour ? */ if (rank == 0) { srand(time(NULL)); random_number = rand(); } MPI_Bcast(&random_number, 1, MPI_INT, 0, MPI_COMM_WORLD); TRACE_MPI_BARRIER(rank, worldsize, random_number); _STARPU_MPI_DEBUG("unique key %x\n", random_number); #endif } int starpu_mpi_initialize(void) { PTHREAD_MUTEX_INIT(&mutex, NULL); PTHREAD_COND_INIT(&cond, NULL); new_requests = starpu_mpi_req_list_new(); PTHREAD_MUTEX_INIT(&detached_requests_mutex, NULL); detached_requests = starpu_mpi_req_list_new(); int ret = pthread_create(&progress_thread, NULL, progress_thread_func, NULL); PTHREAD_MUTEX_LOCK(&mutex); while (!running) PTHREAD_COND_WAIT(&cond, &mutex); PTHREAD_MUTEX_UNLOCK(&mutex); #ifdef USE_STARPU_ACTIVITY hookid = starpu_progression_hook_register(progression_hook_func, NULL); STARPU_ASSERT(hookid >= 0); #endif _starpu_mpi_add_sync_point_in_fxt(); return 0; } int starpu_mpi_shutdown(void) { void *value; /* kill the progression thread */ PTHREAD_MUTEX_LOCK(&mutex); running = 0; PTHREAD_COND_BROADCAST(&cond); PTHREAD_MUTEX_UNLOCK(&mutex); pthread_join(progress_thread, &value); #ifdef USE_STARPU_ACTIVITY starpu_progression_hook_deregister(hookid); #endif /* free the request queues */ starpu_mpi_req_list_delete(detached_requests); starpu_mpi_req_list_delete(new_requests); return 0; }