#include <stdexcept>
#include <sstream>
#include <xbt/log.h>
-#include <xbt/time.h>
XBT_LOG_EXTERNAL_DEFAULT_CATEGORY(proc);
#include "process.h"
+mutex_t *process::proc_mutex;
+
double process::total_load_init = 0.0;
double process::total_load_running = 0.0;
double process::total_load_exit = 0.0;
int process::process_counter = 0;
double process::total_load_average;
+double process::average_load_ratio;
+double process::load_diff_threshold;
+
+std::atomic<int> process::convergence_counter(0);
namespace {
pneigh.reserve(neigh.size());
for (unsigned i = 0 ; i < neigh.size() ; i++) {
neighbor* ptr = &neigh[i];
- m_host_t host = MSG_get_host_by_name(ptr->get_name());
+ msg_host_t host = MSG_get_host_by_name(ptr->get_name());
pneigh.push_back(ptr);
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;
+ idle_duration = 0.0;
convergence = -1.0;
+ proc_mutex->acquire();
process_counter++;
+ convergence_counter++;
+ total_load_init += real_load;
+ total_load_running += real_load;
total_load_average = total_load_running / process_counter;
+ if (opt::avg_load_ratio >= 0.0)
+ average_load_ratio = opt::avg_load_ratio;
+ else
+ average_load_ratio = 100.0 *
+ (process_counter / -opt::avg_load_ratio) / total_load_average;
+ load_diff_threshold = (opt::load_ratio_threshold +
+ average_load_ratio * total_load_average) / 100.0;
+ proc_mutex->release();
ctrl_close_pending = data_close_pending = neigh.size();
close_received = false;
process::~process()
{
delete lb_thread;
+ proc_mutex->acquire();
total_load_exit += real_load;
+ proc_mutex->release();
xbt_assert(received_load == 0.0,
"received_load is %g, but should be 0.0 !", received_load);
if (opt::log_rate < 0)
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);
+ average_load_ratio, convergence);
else
XBT_INFO("Convergence within %g%% was not achieved",
- opt::avg_load_ratio);
+ average_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);
}
+double process::get_iter_deviation() const
+{
+ double average_cost = opt::comp_cost(total_load_average); // fixme: get locked?
+ // Do not count idle periods
+ double comp_iter_opt = acc.comp_amount / average_cost;
+/*
+ // Add iterations that could have been achieved while beeing idle
+ // (kept for documentation)
+ double self_speed = MSG_get_host_speed(MSG_host_self());
+ double average_duration = average_cost / self_speed;
+ comp_iter_opt += idle_duration / average_duration;
+*/
+ return comp_iter - comp_iter_opt;
+}
+
int process::run()
{
if (opt::log_rate >= 0) {
using std::placeholders::_1;
double next_iter_after_date = MSG_get_clock() + opt::min_comp_iter_duration;
+ double idle_since_date = 0.0;
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);
+ // receive
+ // if there is something to compute, do not block
+ // else, block the duration of an *lb* iteration
+ data_receive(real_load > 0.0 ? 0.0 : opt::min_lb_iter_duration);
// send
comm.data_flush(false);
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();
- }
- }
+ convergence_check();
// compute
+ idle_duration += MSG_get_clock() - idle_since_date;
++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);
+ msg_task_t task = MSG_task_create("computation", flops, 0.0, NULL);
+ // MSG_task_set_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);
+ idle_since_date = MSG_get_clock();
+
sleep_until_date(next_iter_after_date, opt::min_comp_iter_duration);
}
XBT_VERB("Going to finalize for %s...", __func__);
+ // Note: idle duration is not counted during finalization
finalizing = true;
XBT_DEBUG("send DATA_CLOSE to %zu neighbor%s",
neigh.size(), ESSE(neigh.size()));
}
real_load += received_load;
received_load = 0.0;
+ proc_mutex->acquire();
total_load_running -= real_load;
+ proc_mutex->release();
+ convergence_check();
comm.data_flush(true);
}
+void process::convergence_check()
+{
+ double average = total_load_average; // fixme: get locked?
+ double load_diff = std::fabs(real_load - average);
+ bool converged = load_diff <= load_diff_threshold;
+
+ if (converged) {
+ if (convergence < 0) {
+ XBT_VERB("current load has converged: %g (%.4g%%)",
+ real_load, 100.0 * load_diff / average);
+ convergence = MSG_get_clock();
+ local_convergence_counter = opt::exit_on_convergence;
+ }
+ if (local_convergence_counter > 0 && --local_convergence_counter == 0)
+ --convergence_counter;
+ } else {
+ if (convergence >= 0.0) {
+ XBT_VERB("current load has diverged: %g (%.4g%%)",
+ real_load, 100.0 * load_diff / average);
+ convergence = -1.0;
+ if (local_convergence_counter == 0)
+ ++convergence_counter;
+ }
+ }
+}
+
bool process::still_running()
{
static bool last_status = true;
XBT_VERB("Reached comp_maxiter: %d/%d", comp_iter, opt::comp_maxiter);
last_status = false;
+ } else if (opt::exit_on_convergence && convergence_counter == 0) {
+ XBT_VERB("Global convergence detected");
+ last_status = false;
+
} else if (opt::exit_on_close && close_received) {
XBT_VERB("Close received");
last_status = false;
last_status = false;
} else if (100.0 * total_load_running / total_load_init <=
- opt::load_ratio_threshold) {
+ opt::load_ratio_threshold) { // fixme: get locked?
// 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.");
else
excess_load = real_load;
- double balance = nb.get_debt() - nb.get_credit();
+ double balance;
+ if (nb.get_credit() > 0.0)
+ balance = nb.get_debt() - nb.get_credit();
+ else
+ balance = nb.get_debt();
load_to_send = std::min(excess_load,
std::max(0.0, balance));
void process::ctrl_receive(double timeout)
{
message* msg;
- m_host_t from;
+ msg_host_t from;
XBT_DEBUG("%sblocking receive on ctrl (%g)", "\0non-" + !timeout, timeout);
while (ctrl_close_pending && comm.ctrl_recv(msg, from, timeout)) {
void process::data_receive(double timeout)
{
message* msg;
- m_host_t from;
+ msg_host_t from;
XBT_DEBUG("%sblocking receive on data (%g)", "\0non-" + !timeout, timeout);
while (data_close_pending && comm.data_recv(msg, from, timeout)) {
}
}
-void process::handle_message(message* msg, m_host_t from)
+void process::handle_message(message* msg, msg_host_t from)
{
switch (msg->get_type()) {
case message::CTRL: {
