total_load_init += load;
ctrl_close_pending = data_close_pending = neigh.size();
- if (neigh.size() == 1) {
- comm.next_close_on_ctrl_is_last();
- comm.next_close_on_data_is_last();
- }
close_received = false;
- may_receive = (neigh.size() > 0); // the same as (ctrl_close_pending ||
- // data_close_pending)
finalizing = false;
- if (may_receive)
- comm.listen();
e_xbt_log_priority_t logp = xbt_log_priority_verbose;
if (!LOG_ISENABLED(logp))
int process::run()
{
+ double next_iter_after_date = 0.0;
INFO1("Initial load: %g", load);
VERB0("Starting...");
- iter = 0;
+ comp_iter = lb_iter = 0;
while (true) {
+ if (opt::min_iter_duration) {
+ double now = MSG_get_clock();
+ if (now < next_iter_after_date){
+ double delay = next_iter_after_date - now;
+ DEBUG1("sleeping for %g s", delay);
+ MSG_process_sleep(next_iter_after_date - now);
+ }
+ next_iter_after_date = MSG_get_clock() + opt::min_iter_duration;
+ }
+
if (load > 0.0) {
- ++iter;
- if (opt::log_rate && iter % opt::log_rate == 0) {
+ ++comp_iter;
+ if (opt::log_rate && comp_iter % opt::log_rate == 0) {
if (opt::bookkeeping)
- INFO3("(%u) current load: %g ; expected: %g",
- iter, load, expected_load);
+ INFO4("(%u:%u) current load: %g ; expected: %g",
+ comp_iter, lb_iter, load, expected_load);
else
INFO2("(%u) current load: %g",
- iter, load);
+ comp_iter, load);
}
if (opt::bookkeeping)
send();
compute();
- if (opt::maxiter && iter >= opt::maxiter)
- break;
} else {
// send load information, and load when bookkeeping
send();
}
+ if (opt::comp_maxiter && comp_iter >= opt::comp_maxiter) {
+ VERB2("Reached comp_maxiter: %d/%d", comp_iter, opt::comp_maxiter);
+ break;
+ }
+ if (opt::lb_maxiter && lb_iter >= opt::lb_maxiter) {
+ VERB2("Reached lb_maxiter: %d/%d", lb_iter, opt::lb_maxiter);
+ break;
+ }
+ if (opt::time_limit && MSG_get_clock() >= opt::time_limit) {
+ VERB2("Reached time limit: %g/%g", MSG_get_clock(), opt::time_limit);
+ break;
+ }
+
// block on receiving unless there is something to compute or
// to send
- bool recv_wait = (load == 0 &&
- ((opt::bookkeeping ? expected_load : load)
- == prev_load_broadcast));
- DEBUG1("CALL RECEIVE(%s)", recv_wait? "WAIT": "NO_WAIT");
- receive(recv_wait? WAIT: NO_WAIT);
+ bool wait = (load == 0 &&
+ ((opt::bookkeeping ? expected_load : load)
+ == prev_load_broadcast));
+ receive(wait);
// one of our neighbor is finalizing
- if (opt::exit_on_close && close_received)
+ if (opt::exit_on_close && close_received) {
+ VERB0("Close received");
break;
+ }
// have no load and cannot receive anything
- if (load == 0.0 && !may_receive)
+ if (load == 0.0 && !may_receive()) {
+ VERB0("I'm a poor lonesome process, and I have no load...");
break;
+ }
// fixme: this check should be implemented with a distributed
// algorithm, and not a shared global variable!
+ // fixme: should this chunk be moved before call to receive() ?
if (100.0 * total_load_running / total_load_init <=
opt::load_ratio_threshold) {
- VERB0("No more load to balance in system, stopping.");
+ VERB0("No more load to balance in system.");
break;
}
-
}
VERB0("Going to finalize...");
finalize();
*/
VERB0("Done.");
- INFO3("Final load after %d iteration%s: %g", iter, ESSE(iter), load);
- if (opt::bookkeeping)
- INFO1("Expected load: %g", expected_load);
+ if (opt::bookkeeping) {
+ INFO4("Final load after %d:%d iterations: %g ; expected: %g",
+ comp_iter, lb_iter, load, expected_load);
+ } else {
+ INFO2("Final load after %d iterations: %g",
+ comp_iter, load);
+ if (comp_iter != lb_iter)
+ WARN2("comp_iter (%d) and lb_iter (%d) differ!",
+ comp_iter, lb_iter);
+ }
return 0;
}
double process::load_balance(double /*my_load*/)
{
- if (iter == 1)
- WARN0("process::load_balance is a no-op!");
+ if (lb_iter == 1) // warn only once
+ WARN0("process::load_balance() is a no-op!");
return 0.0;
}
if (load > 0.0) {
double duration = opt::comp_cost(load);
m_task_t task = MSG_task_create("computation", duration, 0.0, NULL);
- DEBUG2("compute %g flop%s.", duration, ESSE(duration));
+ DEBUG2("compute %g flop%s", duration, ESSE(duration));
MSG_task_execute(task);
MSG_task_destroy(task);
} else {
bind(&process::send1_no_bookkeeping, this, _1));
prev_load_broadcast = load;
}
+ comm.flush(false);
}
-void process::receive(recv_wait_mode wait)
+void process::receive(bool wait)
{
- // DEBUG1("go for receive(%s)",
- // "NO_WAIT\0WAIT\0\0\0\0WAIT_FOR_CLOSE" + 8 * wait);
message* msg;
m_host_t from;
- bool do_wait = (wait != NO_WAIT);
- while (may_receive && comm.recv(msg, from, do_wait)) {
+
+ while (may_receive() && comm.recv(msg, from, wait)) {
switch (msg->get_type()) {
case message::INFO: {
neighbor* n = rev_neigh[from];
break;
}
case message::CTRL_CLOSE:
- if (--ctrl_close_pending == 1)
- comm.next_close_on_ctrl_is_last();
- // DEBUG1("ctrl_close_pending = %d", ctrl_close_pending);
+ ctrl_close_pending--;
close_received = true;
break;
case message::DATA_CLOSE:
- if (--data_close_pending == 1)
- comm.next_close_on_data_is_last();
- // DEBUG1("data_close_pending = %d", data_close_pending);
+ data_close_pending--;
close_received = true;
break;
}
delete msg;
- may_receive = (ctrl_close_pending || data_close_pending);
- do_wait = (wait == WAIT_FOR_CLOSE);
+ wait = false; // only wait on first recv
}
+ comm.flush(false);
}
void process::finalize1(neighbor& nb)
finalizing = true;
total_load_running -= load;
- DEBUG2("send CLOSE to %d neighbor%s.",
- (int )neigh.size(), ESSE(neigh.size()));
+ DEBUG2("send CLOSE to %lu neighbor%s",
+ (unsigned long )neigh.size(), ESSE(neigh.size()));
std::for_each(neigh.begin(), neigh.end(),
bind(&process::finalize1, this, _1));
- DEBUG2("wait for CLOSE from %d neighbor%s.",
- (int )neigh.size(), ESSE(neigh.size()));
- receive(WAIT_FOR_CLOSE);
+ DEBUG2("wait for CLOSE from %lu neighbor%s",
+ (unsigned long )neigh.size(), ESSE(neigh.size()));
+ while (may_receive()) {
+ comm.flush(false);
+ receive(true);
+ }
comm.flush(true);
}
