double process::total_load_running = 0.0;
double process::total_load_exit = 0.0;
+int process::process_counter = 0;
+double process::total_load_average;
+
namespace {
void sleep_until_date(double& date, double duration)
rev_neigh.insert(std::make_pair(host, ptr));
}
+ // Note: there should not be race condition with the current
+ // version of Simgrid, when updating the global variables.
+
prev_load_broadcast = -1; // force sending of load on first send_all()
expected_load = real_load;
total_load_running += real_load;
total_load_init += real_load;
received_load = 0.0;
+ convergence = -1.0;
+
+ process_counter++;
+ total_load_average = total_load_running / process_counter;
+
ctrl_close_pending = data_close_pending = neigh.size();
close_received = false;
finalizing = false;
- comp_iter = lb_iter = 0;
+ all_comp_iter = comp_iter = lb_iter = 0;
lb_thread = new_msg_thread("loba",
std::bind(&process::load_balance_loop, this));
"received_load is %g, but should be 0.0 !", received_load);
if (opt::log_rate < 0)
return;
- XBT_INFO("Final load after %d:%d iterations: %g",
- lb_iter, comp_iter, real_load);
+ XBT_INFO("Final load after %d:%d:%d iterations: %g",
+ lb_iter, comp_iter, all_comp_iter, real_load);
+ if (convergence >= 0.0)
+ XBT_INFO("Convergence within %g%% was achieved at time %g",
+ opt::avg_load_ratio, convergence);
+ else
+ XBT_INFO("Convergence within %g%% was not achieved",
+ opt::avg_load_ratio);
XBT_VERB("Expected load was: %g", expected_load);
XBT_VERB("Total computation for this process: %g", get_comp_amount());
+ print_loads(true, xbt_log_priority_debug);
}
int process::run()
// 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_INFO("(%u:%u:%u) current load: %g",
+ lb_iter, comp_iter, all_comp_iter, real_load);
XBT_VERB("... expected load: %g", expected_load);
}
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);
{
double load_to_send;
if (opt::bookkeeping) { // bookkeeping
- double excess_load;
+ double excess_load; // load amount we are able to send
if (opt::egocentric)
excess_load = std::max(0.0, real_load - expected_load);
else
excess_load = real_load;
- load_to_send = compute_load_to_send(std::min(excess_load,
- nb.get_debt()));
+
+ double balance = nb.get_debt() - nb.get_credit();
+ load_to_send = std::min(excess_load,
+ std::max(0.0, balance));
+
+ // adjust load to send (rounding, truncation, etc.)
+ load_to_send = compute_load_to_send(load_to_send);
if (load_to_send > 0.0)
nb.set_debt(nb.get_debt() - load_to_send);
} else { // !bookkeeping
case message::CTRL: {
neighbor* n = rev_neigh[from];
n->set_load(msg->get_amount() + n->get_to_send());
- expected_load += msg->get_credit(); // may be 0.0 if !opt::bookkeeping
+ if (opt::bookkeeping) {
+ double credit = msg->get_credit();
+ expected_load += credit;
+ n->set_credit(n->get_credit() + credit);
+ }
break;
}
case message::DATA: {
+ neighbor* n = rev_neigh[from];
double ld = msg->get_amount();
received_load += ld;
+ n->set_credit(n->get_credit() - ld);
break;
}
case message::CTRL_CLOSE: