- if (opt::bookkeeping)
- expected_load -= load_balance(expected_load);
- else
- load -= load_balance(load);
-
- send();
- compute();
-
-// NDS for Need To Send
-#define NDS ((opt::bookkeeping ? expected_load : load) != prev_load_broadcast)
- do {
- // General idea: block on receiving unless there is
- // something to compute, or to send, or we must exit.
-
- // fixme: review this chunk, and remove this NDS macro!
-
- // FIXME: HAD A DEADLOCK HERE...
-
- bool recv_wait = (load == 0 && !NDS);
- DEBUG1("CALL RECEIVE(%s)", recv_wait? "WAIT": "NO_WAIT");
- 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;
-
- } while (one_more && may_receive && load == 0 && !NDS);
- DEBUG0("RECEIVE LOOP ENDED");
-#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;
+ // send
+ comm.ctrl_flush(false);
+ std::for_each(neigh.begin(), neigh.end(),
+ std::bind(&process::ctrl_send, this, _1));
+ prev_load_broadcast = expected_load;
+ mutex.release();
+
+ sleep_until_date(next_iter_after_date, opt::min_lb_iter_duration);
+ }
+
+ XBT_VERB("Going to finalize for %s...", __func__);
+ XBT_DEBUG("send CTRL_CLOSE to %zu neighbor%s",
+ neigh.size(), ESSE(neigh.size()));
+ std::for_each(neigh.begin(), neigh.end(),
+ std::bind(&process::ctrl_close, this, _1));
+ while (ctrl_close_pending) {
+ comm.ctrl_flush(false);
+ XBT_DEBUG("waiting for %d CTRL_CLOSE", ctrl_close_pending);
+ ctrl_receive(-1.0);
+ }
+ comm.ctrl_flush(true);
+}
+
+void process::compute_loop()
+{
+ using std::placeholders::_1;
+
+ double next_iter_after_date = MSG_get_clock() + opt::min_comp_iter_duration;
+ while (still_running()) {
+ // receive (do not block if there is something to compute)
+ data_receive(real_load > 0.0 ? 0.0 : opt::min_comp_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;
+ if (real_load == 0.0)
+ continue;
+
+ double load_ratio =
+ 100.0 * std::fabs(real_load / total_load_average - 1.0);
+ if (convergence >= 0.0) {
+ if (load_ratio > opt::avg_load_ratio) {
+ XBT_VERB("current load has diverged: %g (%.4g%%)",
+ real_load, load_ratio);
+ convergence = -1.0;
+ }
+ } else {
+ if (load_ratio <= opt::avg_load_ratio) {
+ XBT_VERB("current load has converged: %g (%.4g%%)",
+ real_load, load_ratio);
+ convergence = MSG_get_clock();
+ }
+ }
+
+ // 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;
+ comm.data_flush(true);