rev_neigh.insert(std::make_pair(host, ptr));
}
+ comp = 0.0;
+
prev_load_broadcast = -1; // force sending of load on first send()
expected_load = load;
total_load_running += load;
int process::run()
{
+ double next_iter_after_date = 0.0;
INFO1("Initial load: %g", load);
VERB0("Starting...");
comp_iter = lb_iter = 0;
while (true) {
- if (load > 0.0) {
- ++comp_iter;
- if (opt::log_rate && comp_iter % opt::log_rate == 0) {
+ 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(delay);
+ }
+ next_iter_after_date = MSG_get_clock() + opt::min_iter_duration;
+ }
+
+ double ld = lb_load();
+ if (ld > 0.0) {
+ ++lb_iter;
+
+ if (opt::log_rate && lb_iter % opt::log_rate == 0) {
if (opt::bookkeeping)
INFO4("(%u:%u) current load: %g ; expected: %g",
- comp_iter, lb_iter, load, expected_load);
+ lb_iter, comp_iter, load, expected_load);
else
INFO2("(%u) current load: %g",
- comp_iter, load);
+ lb_iter, load);
}
- if (opt::bookkeeping)
- expected_load -= load_balance(expected_load);
- else
- load -= load_balance(load);
+ ld -= load_balance(ld);
print_loads(true, xbt_log_priority_debug);
+ }
+ lb_load() = ld;
- send();
+ // send load information, and load (data) if any
+ send();
+ if (load > 0.0) {
+ ++comp_iter;
compute();
-
- } else {
- // send load information, and load when bookkeeping
- send();
}
- if (opt::comp_maxiter && comp_iter >= opt::comp_maxiter)
+ if (opt::lb_maxiter && lb_iter >= opt::lb_maxiter) {
+ VERB2("Reached lb_maxiter: %d/%d", lb_iter, opt::lb_maxiter);
break;
- if (opt::lb_maxiter && lb_iter >= opt::lb_maxiter)
+ }
+ if (opt::comp_maxiter && comp_iter >= opt::comp_maxiter) {
+ VERB2("Reached comp_maxiter: %d/%d", comp_iter, opt::comp_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 wait = (load == 0 &&
- ((opt::bookkeeping ? expected_load : load)
- == prev_load_broadcast));
- receive(wait);
+ double timeout;
+ if (load != 0 || lb_load() != prev_load_broadcast)
+ timeout = 0.0;
+ else if (opt::min_iter_duration)
+ timeout = opt::min_iter_duration;
+ else
+ timeout = 1.0;
+ receive(timeout);
// 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;
+ } else {
+ DEBUG1("still %g load to balance, continuing...", total_load_running);
}
-
}
VERB0("Going to finalize...");
finalize();
*/
VERB0("Done.");
- INFO3("Final load after %d iteration%s: %g",
- comp_iter, ESSE(comp_iter), load);
- if (opt::bookkeeping)
- INFO1("Expected load: %g", expected_load);
+ if (opt::bookkeeping) {
+ INFO4("Final load after %d:%d iterations: %g ; expected: %g",
+ lb_iter, comp_iter, load, expected_load);
+ } else {
+ INFO2("Final load after %d iterations: %g",
+ lb_iter, load);
+ if (lb_iter != comp_iter)
+ WARN2("lb_iter (%d) and comp_iter (%d) differ!",
+ lb_iter, comp_iter);
+ }
return 0;
}
double process::load_balance(double /*my_load*/)
{
if (lb_iter == 1) // warn only once
- WARN0("process::load_balance is a no-op!");
+ WARN0("process::load_balance() is a no-op!");
return 0.0;
}
void process::compute()
{
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));
+ double flops = opt::comp_cost(load);
+ m_task_t task = MSG_task_create("computation", flops, 0.0, NULL);
+ DEBUG2("compute %g flop%s", flops, ESSE(flops));
MSG_task_execute(task);
+ comp += flops;
MSG_task_destroy(task);
} else {
DEBUG0("nothing to compute !");
comm.flush(false);
}
-void process::receive(bool wait)
+void process::receive(double timeout)
{
message* msg;
m_host_t from;
- while (may_receive() && comm.recv(msg, from, wait)) {
+ DEBUG2("%sblocking receive (%g)", "\0non-" + !timeout, timeout);
+ while (may_receive() && comm.recv(msg, from, timeout)) {
switch (msg->get_type()) {
case message::INFO: {
neighbor* n = rev_neigh[from];
break;
}
delete msg;
- wait = false; // only wait on first recv
+ timeout = 0.0; // only wait on first recv
}
comm.flush(false);
}
(unsigned long )neigh.size(), ESSE(neigh.size()));
while (may_receive()) {
comm.flush(false);
- receive(true);
+ receive(-1.0);
}
comm.flush(true);
