X-Git-Url: https://bilbo.iut-bm.univ-fcomte.fr/and/gitweb/loba.git/blobdiff_plain/124a609e80fa7bb5b7a7a698df9c6abef910d83e..7594a30eb5ef2059cf4a04773c7ce101378d4fca:/process.cpp?ds=sidebyside diff --git a/process.cpp b/process.cpp index 11129e4..97f2a69 100644 --- a/process.cpp +++ b/process.cpp @@ -1,4 +1,5 @@ #include +#include #include #include #include @@ -15,10 +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) @@ -36,6 +46,13 @@ process::process(int argc, char* argv[]) if (argc < 2 || !(std::istringstream(argv[1]) >> real_load)) throw std::invalid_argument("bad or missing initial load parameter"); + double iload = std::trunc(real_load); + if (opt::integer_transfer && real_load != iload) { + XBT_WARN("Initial load %g is not an integer. Truncate it.", + real_load); + real_load = iload; + } + neigh.assign(argv + 2, argv + argc); pneigh.reserve(neigh.size()); @@ -48,14 +65,31 @@ process::process(int argc, char* argv[]) prev_load_broadcast = -1; // force sending of load on first send_all() expected_load = real_load; - total_load_running += 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; 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)); @@ -79,13 +113,39 @@ 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) 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", + average_load_ratio, convergence); + else + XBT_INFO("Convergence within %g%% was not achieved", + 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() @@ -124,13 +184,16 @@ void process::load_balance_loop() ++lb_iter; } + ctrl_receive(0.0); + mutex.acquire(); if (!opt::bookkeeping) expected_load = real_load - get_sum_of_to_send(); // 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); } @@ -147,7 +210,6 @@ void process::load_balance_loop() mutex.release(); sleep_until_date(next_iter_after_date, opt::min_lb_iter_duration); - ctrl_receive(0.0); } XBT_VERB("Going to finalize for %s...", __func__); @@ -157,7 +219,7 @@ void process::load_balance_loop() std::bind(&process::ctrl_close, this, _1)); while (ctrl_close_pending) { comm.ctrl_flush(false); - XBT_DEBUG("waiting for %d CTRL CLOSE", ctrl_close_pending); + XBT_DEBUG("waiting for %d CTRL_CLOSE", ctrl_close_pending); ctrl_receive(-1.0); } comm.ctrl_flush(true); @@ -168,26 +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 - mutex.acquire(); - if (real_load > 0.0) - data_receive(0.0); - else - data_receive(opt::min_comp_iter_duration); - mutex.release(); + // 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); 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; 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); @@ -197,27 +263,58 @@ 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__); - // last send, for not losing load scheduled to be sent - std::for_each(neigh.begin(), neigh.end(), - std::bind(&process::data_send, this, _1)); + // Note: idle duration is not counted during finalization finalizing = true; - total_load_running -= real_load; XBT_DEBUG("send DATA_CLOSE to %zu neighbor%s", neigh.size(), ESSE(neigh.size())); std::for_each(neigh.begin(), neigh.end(), std::bind(&process::data_close, this, _1)); while (data_close_pending) { comm.data_flush(false); - XBT_DEBUG("waiting for %d DATA CLOSE", data_close_pending); + XBT_DEBUG("waiting for %d DATA_CLOSE", data_close_pending); data_receive(-1.0); } + 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; @@ -241,6 +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; @@ -250,7 +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."); @@ -303,26 +404,37 @@ void process::ctrl_send(neighbor& nb) } } +double process::compute_load_to_send(double desired) +{ + if (opt::integer_transfer) + desired = std::floor(desired); + return desired >= opt::min_transfer_amount ? desired : 0.0; +} + void process::data_send(neighbor& nb) { double load_to_send; - if (opt::bookkeeping) { - load_to_send = std::min(real_load, nb.get_debt()); - if (load_to_send >= opt::min_transfer_amount) { + if (opt::bookkeeping) { // bookkeeping + 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; + + 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); - real_load -= load_to_send; - } else { - load_to_send = 0.0; - } - } else { - load_to_send = nb.get_to_send(); - if (load_to_send >= opt::min_transfer_amount) { - nb.set_to_send(0.0); - real_load -= load_to_send; - } else { - load_to_send = 0.0; - } + } else { // !bookkeeping + load_to_send = compute_load_to_send(nb.get_to_send()); + if (load_to_send > 0.0) + nb.set_to_send(nb.get_to_send() - load_to_send); } + real_load -= load_to_send; while (load_to_send > 0.0) { double amount; if (opt::max_transfer_amount) @@ -380,14 +492,18 @@ void process::handle_message(message* msg, m_host_t from) 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(); - real_load += ld; - if (finalizing) - total_load_running -= ld; + received_load += ld; + n->set_credit(n->get_credit() - ld); break; } case message::CTRL_CLOSE: