- // General idea: do not iterate if there is nothing to
- // compute, nor to send.
-
- // fixme: review this chunk, and remove this NDS macro!
-
- bool recv_wait = (load == 0 && !NDS);
- bool close_received = !receive(recv_wait? WAIT: NO_WAIT);
-
- if (opt::exit_on_close && close_received)
- one_more = false;
- else if (opt::maxiter && iter >= opt::maxiter)
- one_more = false;
-
- can_recv = (ctrl_close_pending || data_close_pending);
-
- } while (one_more && can_recv && load == 0 && !NDS);
-#undef NDS
-
- } while (one_more);
- VERB0("Going to finalize...");
- finalize();
-
- /* Open Questions :
- * - definition of load on heterogeneous hosts ?
- * - how to detect convergence ?
- * - how to manage link failures ?
- */
-
- VERB0("Done.");
- INFO3("Final load after %d iteration%s: %g", iter, ESSE(iter), load);
- if (opt::bookkeeping)
- INFO1("Expected load: %g", expected_load);
- return 0;
+ // receive
+ // if there is something to compute, do not block
+ // else, block the duration of an *lb* iteration
+ data_receive(real_load > 0.0 ? 0.0 : opt::min_lb_iter_duration);
+
+ // send
+ comm.data_flush(false);
+ mutex.acquire();
+ real_load += received_load;
+ received_load = 0.0;
+ std::for_each(neigh.begin(), neigh.end(),
+ std::bind(&process::data_send, this, _1));
+ mutex.release();
+
+ ++all_comp_iter;
+
+ } while (real_load == 0.0);
+
+ convergence_check();
+
+ // compute
+ ++comp_iter;
+ double flops = opt::comp_cost(real_load);
+ m_task_t task = MSG_task_create("computation", flops, 0.0, NULL);
+ TRACE_msg_set_task_category(task, TRACE_CAT_COMP);
+ XBT_DEBUG("compute %g flop%s", flops, ESSE(flops));
+ MSG_task_execute(task);
+ add_comp_amount(flops);
+ MSG_task_destroy(task);
+
+ sleep_until_date(next_iter_after_date, opt::min_comp_iter_duration);
+ }
+
+ XBT_VERB("Going to finalize for %s...", __func__);
+ finalizing = true;
+ XBT_DEBUG("send DATA_CLOSE to %zu neighbor%s",
+ neigh.size(), ESSE(neigh.size()));
+ std::for_each(neigh.begin(), neigh.end(),
+ std::bind(&process::data_close, this, _1));
+ while (data_close_pending) {
+ comm.data_flush(false);
+ XBT_DEBUG("waiting for %d DATA_CLOSE", data_close_pending);
+ data_receive(-1.0);
+ }
+ real_load += received_load;
+ received_load = 0.0;
+ total_load_running -= real_load;
+ convergence_check();
+ comm.data_flush(true);
+}
+
+void process::convergence_check()
+{
+ double load_diff = std::fabs(real_load - total_load_average);
+ bool converged = load_diff <= load_diff_threshold;
+
+ if (convergence >= 0.0) {
+ if (!converged) {
+ XBT_VERB("current load has diverged: %g (%.4g%%)",
+ real_load, 100.0 * load_diff / total_load_average);
+ convergence = -1.0;
+ }
+ } else {
+ if (converged) {
+ XBT_VERB("current load has converged: %g (%.4g%%)",
+ real_load, 100.0 * load_diff / total_load_average);
+ convergence = MSG_get_clock();
+ }
+ }