X-Git-Url: https://bilbo.iut-bm.univ-fcomte.fr/and/gitweb/loba.git/blobdiff_plain/f8daec44188cc69bbd5b05a1eebfb813a07ae495..c0eaeb249c8fd246f3a56f18f61e417c277f43c1:/process.cpp diff --git a/process.cpp b/process.cpp index 1414033..97f2a69 100644 --- a/process.cpp +++ b/process.cpp @@ -16,14 +16,19 @@ 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 process::convergence_counter(0); + namespace { void sleep_until_date(double& date, double duration) @@ -58,21 +63,27 @@ process::process(int argc, char* argv[]) 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 + - opt::avg_load_ratio * total_load_average) / 100.0; + average_load_ratio * total_load_average) / 100.0; + proc_mutex->release(); ctrl_close_pending = data_close_pending = neigh.size(); close_received = false; @@ -102,7 +113,9 @@ process::process(int argc, char* argv[]) 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) @@ -111,15 +124,30 @@ process::~process() 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) { @@ -202,28 +230,30 @@ void process::compute_loop() 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 + // 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); - do { - // 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(); - - ++all_comp_iter; + // 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(); - } while (real_load == 0.0); + ++all_comp_iter; + if (real_load == 0.0) + continue; 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); @@ -233,10 +263,13 @@ void process::compute_loop() 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())); @@ -249,27 +282,35 @@ void process::compute_loop() } 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 load_diff = std::fabs(real_load - total_load_average); + double average = total_load_average; // fixme: get locked? + double load_diff = std::fabs(real_load - average); bool converged = load_diff <= load_diff_threshold; - if (convergence >= 0.0) { - if (!converged) { - XBT_VERB("current load has diverged: %g (%.4g%%)", - real_load, 100.0 * load_diff / total_load_average); - convergence = -1.0; - } - } else { - if (converged) { + if (converged) { + if (convergence < 0) { XBT_VERB("current load has converged: %g (%.4g%%)", - real_load, 100.0 * load_diff / total_load_average); + 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; } } } @@ -297,6 +338,10 @@ bool process::still_running() 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; @@ -306,7 +351,7 @@ bool process::still_running() 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.");