X-Git-Url: https://bilbo.iut-bm.univ-fcomte.fr/and/gitweb/loba.git/blobdiff_plain/1bba96ae977208a57bd80c7011c2033435a64d44..HEAD:/process.cpp diff --git a/process.cpp b/process.cpp index 6d791bd..38cec99 100644 --- a/process.cpp +++ b/process.cpp @@ -1,10 +1,10 @@ #include -#include +#include #include +#include #include #include #include -#include XBT_LOG_EXTERNAL_DEFAULT_CATEGORY(proc); @@ -14,36 +14,83 @@ 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) + { + double sleep_duration = date - MSG_get_clock(); + if (sleep_duration > 0.0) + MSG_process_sleep(sleep_duration); + date = MSG_get_clock() + duration; + } + +} + 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()); 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)); } - comp = 0.0; - 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; + all_comp_iter = comp_iter = lb_iter = 0; + + lb_thread = new_msg_thread("loba", [this]() { this->load_balance_loop(); }); + e_xbt_log_priority_t logp = xbt_log_priority_verbose; if (!LOG_ISENABLED(logp)) return; @@ -53,294 +100,444 @@ process::process(int argc, char* argv[]) oss << ESSE(neigh.size()) << ": "; std::transform(neigh.begin(), neigh.end() - 1, std::ostream_iterator(oss, ", "), - std::tr1::mem_fn(&neighbor::get_name)); + [](const neighbor& neigh) { return neigh.get_name(); }); oss << neigh.back().get_name(); } - LOG1(logp, "Got %s.", oss.str().c_str()); + XBT_LOG(logp, "Got %s.", oss.str().c_str()); print_loads(false, logp); } process::~process() { + delete lb_thread; + proc_mutex->acquire(); total_load_exit += real_load; - if (opt::bookkeeping) { - INFO4("Final load after %d:%d iterations: %g ; expected: %g", - lb_iter, comp_iter, real_load, expected_load); - } else { - INFO2("Final load after %d iterations: %g", - lb_iter, real_load); - if (lb_iter != comp_iter) - WARN2("lb_iter (%d) and comp_iter (%d) differ!", - lb_iter, comp_iter); - } - VERB1("Total computation for this process: %g", comp); + 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:%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() { - double next_iter_after_date = 0.0; - INFO1("Initial load: %g", real_load); - VERB0("Starting..."); - comp_iter = lb_iter = 0; - while (true) { - double ld = lb_load(); - if (ld > 0.0) { - double now = MSG_get_clock(); - if (now < next_iter_after_date) - MSG_process_sleep(next_iter_after_date - now); - next_iter_after_date = MSG_get_clock() + opt::min_iter_duration; + if (opt::log_rate >= 0) { + XBT_INFO("Initial load: %g", real_load); + XBT_VERB("Initial expected load: %g", expected_load); + } + XBT_VERB("Starting..."); + mutex.acquire(); + lb_thread->start(); + while (lb_iter <= opt::comp_iter_delay) + cond.wait(mutex); + mutex.release(); + double sleep_duration = opt::comp_time_delay - MSG_get_clock(); + if (sleep_duration > 0.0) + MSG_process_sleep(sleep_duration); + compute_loop(); + lb_thread->wait(); + XBT_VERB("Done."); + return 0; +} +void process::load_balance_loop() +{ + double next_iter_after_date = MSG_get_clock() + opt::min_lb_iter_duration; + while (still_running()) { + if (lb_iter == opt::comp_iter_delay) { + mutex.acquire(); ++lb_iter; + cond.signal(); + mutex.release(); + } else { + ++lb_iter; + } - if (opt::log_rate && lb_iter % opt::log_rate == 0) { - if (opt::bookkeeping) - INFO4("(%u:%u) current load: %g ; expected: %g", - lb_iter, comp_iter, real_load, expected_load); - else - INFO2("(%u) current load: %g", - lb_iter, real_load); - } + ctrl_receive(0.0); - ld -= load_balance(ld); + mutex.acquire(); + if (!opt::bookkeeping) + expected_load = real_load - get_sum_of_to_send(); + // nothing to do with opt::bookkeeping - print_loads(true, xbt_log_priority_debug); + if (opt::log_rate && lb_iter % opt::log_rate == 0) { + XBT_INFO("(%u:%u:%u) current load: %g", + lb_iter, comp_iter, all_comp_iter, real_load); + XBT_VERB("... expected load: %g", expected_load); } - lb_load() = ld; - // send load information, and load (data) if any - send_all(); - if (real_load > 0.0) { - ++comp_iter; - compute(); - } + if (expected_load > 0.0) + load_balance(); - if (opt::lb_maxiter && lb_iter >= opt::lb_maxiter) { - VERB2("Reached lb_maxiter: %d/%d", lb_iter, opt::lb_maxiter); - break; - } - if (opt::comp_maxiter && comp_iter >= opt::comp_maxiter) { - VERB2("Reached comp_maxiter: %d/%d", comp_iter, opt::comp_maxiter); - break; + print_loads(true, xbt_log_priority_debug); + + // send + comm.ctrl_flush(false); + for (neighbor& n : neigh) + ctrl_send(n); + prev_load_broadcast = expected_load; + mutex.release(); + + sleep_until_date(next_iter_after_date, opt::min_lb_iter_duration); + } + + XBT_VERB("Going to finalize for %s...", __func__); + XBT_DEBUG("send CTRL_CLOSE to %zu neighbor%s", + neigh.size(), ESSE(neigh.size())); + for (neighbor& n : neigh) + ctrl_close(n); + while (ctrl_close_pending) { + comm.ctrl_flush(false); + XBT_DEBUG("waiting for %d CTRL_CLOSE", ctrl_close_pending); + ctrl_receive(-1.0); + } + comm.ctrl_flush(true); +} + +void process::compute_loop() +{ + 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); + + // send + comm.data_flush(false); + mutex.acquire(); + real_load += received_load; + received_load = 0.0; + for (neighbor& n : neigh) + data_send(n); + 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); + msg_task_t task = MSG_task_create("computation", flops, 0.0, nullptr); + // 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())); + for (neighbor& n : neigh) + data_close(n); + while (data_close_pending) { + comm.data_flush(false); + 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 (opt::time_limit && MSG_get_clock() >= opt::time_limit) { - VERB2("Reached time limit: %g/%g", MSG_get_clock(), opt::time_limit); - break; + 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; } + } +} - // block on receiving unless there is something to compute or - // to send - double timeout; - if (real_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); +bool process::still_running() +{ + static bool last_status = true; - // one of our neighbor is finalizing - if (opt::exit_on_close && close_received) { - VERB0("Close received"); - break; - } + if (!last_status) { + /* nop */ - // have no load and cannot receive anything - if (real_load == 0.0 && !may_receive()) { - VERB0("I'm a poor lonesome process, and I have no load..."); - break; - } + } else if (opt::exit_request) { + XBT_VERB("Global exit requested"); + last_status = false; + + } else if (opt::time_limit && MSG_get_clock() >= opt::time_limit) { + XBT_VERB("Reached time limit: %g/%g", MSG_get_clock(), opt::time_limit); + last_status = false; + } else if (opt::lb_maxiter && lb_iter >= opt::lb_maxiter) { + XBT_VERB("Reached lb_maxiter: %d/%d", lb_iter, opt::lb_maxiter); + last_status = false; + + } else if (opt::comp_maxiter && comp_iter >= opt::comp_maxiter) { + 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; + + } else if (real_load == 0.0 && !data_close_pending) { + XBT_VERB("I'm a poor lonesome process, and I have no load..."); + last_status = false; + + } else if (100.0 * total_load_running / total_load_init <= + opt::load_ratio_threshold) { // fixme: get locked? // 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."); - break; - } else { - DEBUG1("still %g load to balance, continuing...", total_load_running); - } + XBT_VERB("No more load to balance in system."); + last_status = false; } - VERB0("Going to finalize..."); - finalize(); - /* Open Questions : - * - definition of load on heterogeneous hosts ? - * - how to detect convergence ? - * - how to manage link failures ? - */ + return last_status; +} - VERB0("Done."); - return 0; +double process::get_sum_of_to_send() const +{ + return std::accumulate(neigh.begin(), neigh.end(), 0.0, + [](double x, const neighbor& neigh) { + return x + neigh.get_to_send(); + }); } -double process::load_balance(double /*my_load*/) +void process::load_balance() { if (lb_iter == 1) // warn only once - WARN0("process::load_balance() is a no-op!"); - return 0.0; + XBT_WARN("process::load_balance() is a no-op!"); } -void process::compute() +void process::send(neighbor& nb, double amount) { - if (real_load > 0.0) { - 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); - DEBUG2("compute %g flop%s", flops, ESSE(flops)); - MSG_task_execute(task); - comp += flops; - MSG_task_destroy(task); - } else { - DEBUG0("nothing to compute !"); - } + expected_load -= amount; + nb.set_to_send(nb.get_to_send() + amount); + nb.set_load(nb.get_load() + amount); } -void process::send1_no_bookkeeping(neighbor& nb) +void process::ctrl_send(neighbor& nb) { - if (real_load != prev_load_broadcast) - comm.send(nb.get_ctrl_mbox(), new message(message::INFO, real_load)); - double load_to_send = nb.get_to_send(); - if (load_to_send > 0.0) { - comm.send(nb.get_data_mbox(), new message(message::LOAD, load_to_send)); - nb.set_to_send(0.0); + double info_to_send = expected_load; + double debt_to_send; + if (opt::bookkeeping) { // bookkeeping + debt_to_send = nb.get_to_send(); + if (debt_to_send > 0.0) { + nb.set_to_send(0.0); + nb.set_debt(nb.get_debt() + debt_to_send); + } + } else { // !bookkeeping + debt_to_send = 0.0; + } + if (info_to_send != prev_load_broadcast || debt_to_send > 0.0) { + message* msg = + new message(message::message_type::CTRL, info_to_send, debt_to_send); + add_ctrl_send_mesg(msg->get_size()); + comm.ctrl_send(nb.get_ctrl_mbox(), msg); } } -void process::send1_bookkeeping(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) { - if (expected_load != prev_load_broadcast) - comm.send(nb.get_ctrl_mbox(), - new message(message::INFO, expected_load)); double load_to_send; - double new_debt; - double debt_to_send = nb.get_to_send(); - if (debt_to_send > 0.0) { - comm.send(nb.get_ctrl_mbox(), - new message(message::CREDIT, debt_to_send)); - nb.set_to_send(0.0); - new_debt = nb.get_debt() + debt_to_send; - } else { - new_debt = nb.get_debt(); + 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; + 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)); + + // 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); + } 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); } - if (real_load <= new_debt) { - load_to_send = real_load; - nb.set_debt(new_debt - load_to_send); - real_load = 0.0; - } else { - load_to_send = new_debt; - nb.set_debt(0.0); - real_load -= load_to_send; + real_load -= load_to_send; + while (load_to_send > 0.0) { + double amount; + if (opt::max_transfer_amount) + amount = std::min(load_to_send, opt::max_transfer_amount); + else + amount = load_to_send; + message* msg = new message(message::message_type::DATA, amount); + add_data_send_mesg(msg->get_size()); + comm.data_send(nb.get_data_mbox(), msg); + load_to_send -= amount; } - if (load_to_send > 0.0) - comm.send(nb.get_data_mbox(), new message(message::LOAD, load_to_send)); } -void process::send_all() +void process::ctrl_close(neighbor& nb) { - using std::tr1::bind; - using std::tr1::placeholders::_1; + comm.ctrl_send(nb.get_ctrl_mbox(), + new message(message::message_type::CTRL_CLOSE, 0.0)); +} - if (opt::bookkeeping) { - std::for_each(neigh.begin(), neigh.end(), - bind(&process::send1_bookkeeping, this, _1)); - prev_load_broadcast = expected_load; - } else { - std::for_each(neigh.begin(), neigh.end(), - bind(&process::send1_no_bookkeeping, this, _1)); - prev_load_broadcast = real_load; - } - comm.flush(false); +void process::data_close(neighbor& nb) +{ + comm.data_send(nb.get_data_mbox(), + new message(message::message_type::DATA_CLOSE, 0.0)); } -void process::receive(double timeout) +void process::ctrl_receive(double timeout) { message* msg; - m_host_t from; - - 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]; - n->set_load(msg->get_amount()); - break; - } - case message::CREDIT: - expected_load += msg->get_amount(); - break; - case message::LOAD: { - double ld = msg->get_amount(); - real_load += ld; - if (finalizing) - total_load_running -= ld; - break; - } - case message::CTRL_CLOSE: - ctrl_close_pending--; - close_received = true; - break; - case message::DATA_CLOSE: - data_close_pending--; - close_received = true; - break; - } - delete msg; - timeout = 0.0; // only wait on first recv + msg_host_t from; + + XBT_DEBUG("%sblocking receive on ctrl (%g)", "\0non-" + !timeout, timeout); + while (ctrl_close_pending && comm.ctrl_recv(msg, from, timeout)) { + if (msg->get_type() != message::message_type::CTRL_CLOSE) + add_ctrl_recv_mesg(msg->get_size()); + handle_message(msg, from); + timeout = 0.0; } - comm.flush(false); } -void process::finalize1(neighbor& nb) +void process::data_receive(double timeout) { - comm.send(nb.get_ctrl_mbox(), new message(message::CTRL_CLOSE, 0.0)); - comm.send(nb.get_data_mbox(), new message(message::DATA_CLOSE, 0.0)); + message* msg; + msg_host_t from; + + XBT_DEBUG("%sblocking receive on data (%g)", "\0non-" + !timeout, timeout); + while (data_close_pending && comm.data_recv(msg, from, timeout)) { + if (msg->get_type() != message::message_type::DATA_CLOSE) + add_data_recv_mesg(msg->get_size()); + handle_message(msg, from); + timeout = 0.0; + } } -void process::finalize() +void process::handle_message(message* msg, msg_host_t from) { - using std::tr1::bind; - using std::tr1::placeholders::_1; - - finalizing = true; - total_load_running -= real_load; - - 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 %lu neighbor%s", - (unsigned long )neigh.size(), ESSE(neigh.size())); - while (may_receive()) { - comm.flush(false); - receive(-1.0); + switch (msg->get_type()) { + case message::message_type::CTRL: { + neighbor* n = rev_neigh[from]; + n->set_load(msg->get_amount() + n->get_to_send()); + if (opt::bookkeeping) { + double credit = msg->get_credit(); + expected_load += credit; + n->set_credit(n->get_credit() + credit); + } + break; } - - comm.flush(true); + case message::message_type::DATA: { + neighbor* n = rev_neigh[from]; + double ld = msg->get_amount(); + received_load += ld; + n->set_credit(n->get_credit() - ld); + break; + } + case message::message_type::CTRL_CLOSE: + ctrl_close_pending--; + close_received = true; + break; + case message::message_type::DATA_CLOSE: + data_close_pending--; + close_received = true; + break; + } + delete msg; } -#define print_loads_generic(vec, verbose, logp, cat) \ - if (_XBT_LOG_ISENABLEDV((*cat), logp)) { \ - using std::tr1::bind; \ - using std::tr1::placeholders::_1; \ - XCLOG0(cat, logp, "Neighbor loads:"); \ - std::for_each(vec.begin(), vec.end(), \ - bind(&neighbor::print, _1, verbose, logp, cat)); \ - } else ((void)0) - void process::print_loads(bool verbose, e_xbt_log_priority_t logp, xbt_log_category_t cat) const { - print_loads_generic(neigh, verbose, logp, cat); + if (!_XBT_LOG_ISENABLEDV((*cat), logp)) + return; + XBT_XCLOG(cat, logp, "My load: %g (real); %g (expected). Neighbor loads:", + real_load, expected_load); + for (const neighbor& n : neigh) + n.print(verbose, logp, cat); } void process::print_loads_p(bool verbose, e_xbt_log_priority_t logp, xbt_log_category_t cat) const { - print_loads_generic(pneigh, verbose, logp, cat); + if (!_XBT_LOG_ISENABLEDV((*cat), logp)) + return; + XBT_XCLOG(cat, logp, "My load: %g (real); %g (expected). Neighbor loads:", + real_load, expected_load); + for (const neighbor* n : pneigh) + n->print(verbose, logp, cat); } #undef print_loads_generic