- 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)
- bool can_recv;
- do {
- // 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;
+ ctrl_receive(0.0);
+
+ mutex.acquire();
+ if (!opt::bookkeeping)
+ expected_load = real_load - get_sum_of_to_send();
+ // nothing to do with opt::bookkeeping
+
+ if (opt::log_rate && lb_iter % opt::log_rate == 0) {
+ XBT_INFO("(%u:%u) current load: %g", lb_iter, comp_iter, real_load);
+ XBT_VERB("... expected load: %g", expected_load);
+ }
+
+ if (expected_load > 0.0)
+ load_balance();
+
+ print_loads(true, xbt_log_priority_debug);
+
+ // 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();
+
+ if (real_load == 0.0)
+ continue;
+
+ // 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);
+}
+
+bool process::still_running()
+{
+ static bool last_status = true;
+
+ if (!last_status) {
+ /* nop */
+
+ } else if (opt::exit_request) {
+ XBT_VERB("Global exit requested");
+ last_status = false;
+
+ } else if (opt::time_limit && MSG_get_clock() >= opt::time_limit) {
+ XBT_VERB("Reached time limit: %g/%g", MSG_get_clock(), opt::time_limit);
+ last_status = false;
+
+ } else if (opt::lb_maxiter && lb_iter >= opt::lb_maxiter) {
+ XBT_VERB("Reached lb_maxiter: %d/%d", lb_iter, opt::lb_maxiter);
+ last_status = false;
+
+ } else if (opt::comp_maxiter && comp_iter >= opt::comp_maxiter) {
+ XBT_VERB("Reached comp_maxiter: %d/%d", comp_iter, opt::comp_maxiter);
+ last_status = false;
+
+ } else if (opt::exit_on_close && close_received) {
+ XBT_VERB("Close received");
+ last_status = false;
+
+ } else if (real_load == 0.0 && !data_close_pending) {
+ XBT_VERB("I'm a poor lonesome process, and I have no load...");
+ last_status = false;
+
+ } else if (100.0 * total_load_running / total_load_init <=
+ opt::load_ratio_threshold) {
+ // fixme: this check should be implemented with a distributed
+ // algorithm, and not a shared global variable!
+ XBT_VERB("No more load to balance in system.");
+ last_status = false;
+ }
+
+ return last_status;
