X-Git-Url: https://bilbo.iut-bm.univ-fcomte.fr/and/gitweb/loba.git/blobdiff_plain/5279f43fc9be71815091e79c76d5bc04c8c88421..9936cb5ae7eaf8907cbff5722b206087b18053ca:/process.cpp diff --git a/process.cpp b/process.cpp index d8ec500..34ddbea 100644 --- a/process.cpp +++ b/process.cpp @@ -5,32 +5,50 @@ #include #include #include + +XBT_LOG_EXTERNAL_DEFAULT_CATEGORY(proc); + #include "misc.h" #include "options.h" +#include "tracing.h" + #include "process.h" -XBT_LOG_EXTERNAL_DEFAULT_CATEGORY(simu); +double process::total_load_init = 0.0; +double process::total_load_running = 0.0; +double process::total_load_exit = 0.0; process::process(int argc, char* argv[]) { - if (argc < 2 || !(std::istringstream(argv[1]) >> load)) - throw std::invalid_argument("bad or missing initial load"); + if (argc < 2 || !(std::istringstream(argv[1]) >> real_load)) + throw std::invalid_argument("bad or missing initial load parameter"); neigh.assign(argv + 2, argv + argc); - expected_load = load; - ctrl_close_pending = data_close_pending = neigh.size(); - if (neigh.size() == 1) { - comm.next_close_on_ctrl_is_last(); - comm.next_close_on_data_is_last(); + 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()); + pneigh.push_back(ptr); + rev_neigh.insert(std::make_pair(host, ptr)); } - if (neigh.size() > 0) - comm.listen(); + + comp = 0.0; + + 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; + + ctrl_close_pending = data_close_pending = neigh.size(); + close_received = false; + finalizing = false; + + comp_iter = lb_iter = 0; e_xbt_log_priority_t logp = xbt_log_priority_verbose; if (!LOG_ISENABLED(logp)) return; - LOG1(logp, "My initial load is: %g", load); std::ostringstream oss; oss << neigh.size() << " neighbor"; if (!neigh.empty()) { @@ -40,42 +58,109 @@ process::process(int argc, char* argv[]) std::tr1::mem_fn(&neighbor::get_name)); oss << neigh.back().get_name(); } - LOG1(logp, "Got %s.", oss.str().c_str()); - print_loads(logp); + XBT_LOG(logp, "Got %s.", oss.str().c_str()); + print_loads(false, logp); +} + +process::~process() +{ + total_load_exit += real_load; + if (opt::bookkeeping) { + XBT_INFO("Final load after %d:%d iterations: %g ; expected: %g", + lb_iter, comp_iter, real_load, expected_load); + } else { + XBT_INFO("Final load after %d iterations: %g", + lb_iter, real_load); + if (lb_iter != comp_iter) + XBT_WARN("lb_iter (%d) and comp_iter (%d) differ!", + lb_iter, comp_iter); + } + XBT_VERB("Total computation for this process: %g", comp); } int process::run() { - bool one_more = true; - unsigned iter = 0; - VERB0("Starting..."); - while (one_more) { - bool close_received; - - if (opt::bookkeeping) - INFO3("(%u) current load: %g ; expected: %g", - iter, load, expected_load); - else - INFO2("(%u) current load: %g", - iter, load); + double next_iter_after_date = 0.0; + XBT_INFO("Initial load: %g", real_load); + XBT_VERB("Starting..."); + while (true) { + if (get_load() > 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_lb_iter_duration; - compute(); - close_received = !receive(false); + ++lb_iter; - /* - * compute load balancing; - * send tasks to neighbors; - */ + if (opt::log_rate && lb_iter % opt::log_rate == 0) { + if (opt::bookkeeping) + XBT_INFO("(%u:%u) current load: %g ; expected: %g", + lb_iter, comp_iter, real_load, expected_load); + else + XBT_INFO("(%u) current load: %g", + lb_iter, real_load); + } - comm.flush(false); - ++iter; + load_balance(); - if (opt::exit_on_close && close_received) - one_more = false; - if (opt::maxiter && iter >= opt::maxiter) - one_more = false; + print_loads(true, xbt_log_priority_debug); + } + + // send load information, and load (data) if any + send_all(); + if (real_load > 0.0) { + ++comp_iter; + compute(); + } + + if (opt::lb_maxiter && lb_iter >= opt::lb_maxiter) { + XBT_VERB("Reached lb_maxiter: %d/%d", lb_iter, opt::lb_maxiter); + break; + } + if (opt::comp_maxiter && comp_iter >= opt::comp_maxiter) { + XBT_VERB("Reached comp_maxiter: %d/%d", comp_iter, opt::comp_maxiter); + break; + } + if (opt::time_limit && MSG_get_clock() >= opt::time_limit) { + XBT_VERB("Reached time limit: %g/%g", MSG_get_clock(), opt::time_limit); + break; + } + + // block on receiving unless there is something to compute or + // to send + double timeout; + if (real_load != 0 || get_load() != prev_load_broadcast) + timeout = 0.0; + else if (opt::min_lb_iter_duration) + timeout = opt::min_lb_iter_duration; + else + timeout = 1.0; + receive(timeout); + + // one of our neighbor is finalizing + if (opt::exit_on_close && close_received) { + XBT_VERB("Close received"); + break; + } + + // have no load and cannot receive anything + if (real_load == 0.0 && !may_receive()) { + XBT_VERB("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) { + XBT_VERB("No more load to balance in system."); + break; + } else { + XBT_DEBUG("still %g load to balance, continuing...", total_load_running); + } } - VERB0("Going to finalize..."); + XBT_VERB("Going to finalize..."); finalize(); /* Open Questions : @@ -84,94 +169,186 @@ int process::run() * - how to manage link failures ? */ - VERB0("Done."); + XBT_VERB("Done."); return 0; } +void process::load_balance() +{ + if (lb_iter == 1) // warn only once + XBT_WARN("process::load_balance() is a no-op!"); +} + void process::compute() { - // fixme: shall we do something special when duration is 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)); - MSG_task_execute(task); - MSG_task_destroy(task); + 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); + XBT_DEBUG("compute %g flop%s", flops, ESSE(flops)); + MSG_task_execute(task); + comp += flops; + MSG_task_destroy(task); + } else { + XBT_DEBUG("nothing to compute !"); + } +} + +void process::send(neighbor& nb, double amount) +{ + set_load(get_load() - amount); + nb.set_to_send(nb.get_to_send() + amount); + nb.set_load(nb.get_load() + amount); // fixme: make this optional? } +void process::send1_no_bookkeeping(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); + } +} -// Returns false if a CLOSE message was received. -bool process::receive(bool wait_for_close) +void process::send1_bookkeeping(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 (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; + } + if (load_to_send > 0.0) + comm.send(nb.get_data_mbox(), new message(message::LOAD, load_to_send)); +} + +void process::send_all() +{ + using std::tr1::bind; + using std::tr1::placeholders::_1; + + 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::receive(double timeout) { - bool result = true; message* msg; m_host_t from; - while ((ctrl_close_pending || - data_close_pending) && comm.recv(msg, from, wait_for_close)) { - DEBUG2("received %s from %s", - msg->to_string().c_str(), MSG_host_get_name(from)); + + XBT_DEBUG("%sblocking receive (%g)", "\0non-" + !timeout, timeout); + while (may_receive() && comm.recv(msg, from, timeout)) { switch (msg->get_type()) { - case message::INFO: - // fixme: update neighbor - // need a map m_host_t -> neighbor& + 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: - load += msg->get_amount(); + case message::LOAD: { + double ld = msg->get_amount(); + real_load += ld; + if (finalizing) + total_load_running -= ld; break; + } case message::CTRL_CLOSE: - if (--ctrl_close_pending == 1) - comm.next_close_on_ctrl_is_last(); - DEBUG1("ctrl_close_pending = %d", ctrl_close_pending); - result = false; + ctrl_close_pending--; + close_received = true; break; case message::DATA_CLOSE: - if (--data_close_pending == 1) - comm.next_close_on_data_is_last(); - DEBUG1("data_close_pending = %d", data_close_pending); - result = false; + data_close_pending--; + close_received = true; break; } delete msg; + timeout = 0.0; // only wait on first recv } - return result; + comm.flush(false); +} + +void process::finalize1(neighbor& nb) +{ + 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)); } void process::finalize() { - DEBUG2("send CLOSE to %d neighbor%s.", - (int )neigh.size(), ESSE(neigh.size())); - std::vector::iterator n; - for (n = neigh.begin() ; n != neigh.end() ; ++n) { - comm.send(n->get_ctrl_mbox(), new message(message::CTRL_CLOSE, 0.0)); - comm.send(n->get_data_mbox(), new message(message::DATA_CLOSE, 0.0)); - } + using std::tr1::bind; + using std::tr1::placeholders::_1; + + finalizing = true; + total_load_running -= real_load; + + XBT_DEBUG("send CLOSE to %zu neighbor%s", neigh.size(), ESSE(neigh.size())); + std::for_each(neigh.begin(), neigh.end(), + bind(&process::finalize1, this, _1)); - DEBUG2("wait for CLOSE from %d neighbor%s.", - (int )neigh.size(), ESSE(neigh.size())); - receive(true); + while (may_receive()) { + comm.flush(false); + XBT_DEBUG("waiting for %d CTRL and %d DATA CLOSE", + ctrl_close_pending, data_close_pending); + receive(-1.0); + } comm.flush(true); } -void process::print_loads(e_xbt_log_priority_t logp) +#define print_loads_generic(vec, verbose, logp, cat) \ + if (_XBT_LOG_ISENABLEDV((*cat), logp)) { \ + using std::tr1::bind; \ + using std::tr1::placeholders::_1; \ + XBT_XCLOG(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 { - if (!LOG_ISENABLED(logp)) - return; + print_loads_generic(neigh, verbose, logp, cat); +} - std::ostringstream oss; - if (neigh.empty()) { - oss << "no neighbor!"; - } else { - std::transform(neigh.begin(), neigh.end() - 1, - std::ostream_iterator(oss, ", "), - std::tr1::mem_fn(&neighbor::get_load)); - oss << neigh.back().get_load(); - } - LOG1(logp, "Neighbor loads: %s", oss.str().c_str()); +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); } +#undef print_loads_generic + // Local variables: // mode: c++ // End: