From 6a266736a5a4f2e30a529662180050e8bc2b1a49 Mon Sep 17 00:00:00 2001 From: Arnaud Giersch Date: Thu, 10 Feb 2011 18:52:54 +0100 Subject: [PATCH] Simulate computations in a separate thread. This is work in progress. The code compiles and runs, but may not work as expected... --- process.cpp | 429 +++++++++++++++++++++++++++------------------------- process.h | 43 +++--- 2 files changed, 242 insertions(+), 230 deletions(-) diff --git a/process.cpp b/process.cpp index 39646cd..fe08b33 100644 --- a/process.cpp +++ b/process.cpp @@ -46,6 +46,10 @@ process::process(int argc, char* argv[]) comp_iter = lb_iter = 0; + compute_thread = new_msg_thread("compute", + std::tr1::bind(&process::compute_loop, + this)); + e_xbt_log_priority_t logp = xbt_log_priority_verbose; if (!LOG_ISENABLED(logp)) return; @@ -64,6 +68,7 @@ process::process(int argc, char* argv[]) process::~process() { + delete compute_thread; total_load_exit += real_load; if (opt::bookkeeping) { XBT_INFO("Final load after %d:%d iterations: %g ; expected: %g", @@ -80,108 +85,85 @@ process::~process() int process::run() { - 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; - - ++lb_iter; - - 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); - } + compute_thread->start(); + load_balance_loop(); + compute_thread->wait(); + XBT_VERB("Done."); + return 0; +} - load_balance(); +void process::load_balance_loop() +{ + using std::tr1::bind; + using std::tr1::placeholders::_1; - print_loads(true, xbt_log_priority_debug); - } + while (still_running()) { + ++lb_iter; - // send load information, and load (data) if any - send_all(); - if (real_load > 0.0) { - ++comp_iter; - compute(); + 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:%u) current load: %g", + lb_iter, comp_iter, real_load); } - 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; - } + if (get_load() > 0.0) + load_balance(); + + print_loads(true, xbt_log_priority_debug); + + // send + std::for_each(neigh.begin(), neigh.end(), + bind(&process::ctrl_send, this, _1)); // 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; - } + ctrl_receive(opt::min_lb_iter_duration); - // 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); - } + comm.ctrl_flush(false); } - XBT_VERB("Going to finalize..."); - finalize(); - - /* Open Questions : - * - definition of load on heterogeneous hosts ? - * - how to detect convergence ? - * - how to manage link failures ? - */ - XBT_VERB("Done."); - return 0; + XBT_VERB("Going to finalize for %s...", __func__); + XBT_DEBUG("send CTRL_CLOSE to %zu neighbor%s", + neigh.size(), ESSE(neigh.size())); + std::for_each(neigh.begin(), neigh.end(), + bind(&process::ctrl_close, this, _1)); + 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::load_balance() +void process::compute_loop() { - if (lb_iter == 1) // warn only once - XBT_WARN("process::load_balance() is a no-op!"); -} + using std::tr1::bind; + using std::tr1::placeholders::_1; -void process::compute() -{ - if (real_load > 0.0) { + double next_iter_after_date = 0.0; + while (still_running()) { + // receive + double sleep_duration = real_load + ? std::max(MSG_get_clock() - next_iter_after_date, 0.0) + : opt::min_comp_iter_duration; + data_receive(sleep_duration); + + comm.data_flush(false); + + if (real_load == 0.0) + continue; + + // send + std::for_each(neigh.begin(), neigh.end(), + bind(&process::data_send, this, _1)); + + // compute + ++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); @@ -189,147 +171,73 @@ void process::compute() 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.ctrl_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.data_send(nb.get_data_mbox(), - new message(message::LOAD, load_to_send)); - nb.set_to_send(0.0); + next_iter_after_date = MSG_get_clock() + opt::min_comp_iter_duration; } -} -void process::send1_bookkeeping(neighbor& nb) -{ - if (expected_load != prev_load_broadcast) - comm.ctrl_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.ctrl_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; + XBT_VERB("Going to finalize for %s...", __func__); + XBT_DEBUG("send DATA_CLOSE to %zu neighbor%s", + neigh.size(), ESSE(neigh.size())); + std::for_each(neigh.begin(), neigh.end(), + bind(&process::data_close, this, _1)); + while (data_close_pending) { + comm.data_flush(false); + XBT_DEBUG("waiting for %d DATA CLOSE", data_close_pending); + data_receive(-1.0); } - if (load_to_send > 0.0) - comm.data_send(nb.get_data_mbox(), - new message(message::LOAD, load_to_send)); + comm.data_flush(true); } -void process::send_all() +bool process::still_running() { - using std::tr1::bind; - using std::tr1::placeholders::_1; + static bool last_status = true; - 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.ctrl_flush(false); - comm.data_flush(false); -} + if (!last_status) { + /* nop */ -void process::receive(double timeout) -{ - message* msg; - m_host_t from; + } 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; - XBT_DEBUG("%sblocking receive (%g)", "\0non-" + !timeout, timeout); - while (may_receive() && (comm.ctrl_recv(msg, from, timeout) || - comm.data_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 + } 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_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: 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() ? + XBT_VERB("No more load to balance in system."); + last_status = false; } - comm.ctrl_flush(false); - comm.data_flush(false); + + return last_status; } -void process::finalize1(neighbor& nb) +void process::load_balance() { - comm.ctrl_send(nb.get_ctrl_mbox(), new message(message::CTRL_CLOSE, 0.0)); - comm.data_send(nb.get_data_mbox(), new message(message::DATA_CLOSE, 0.0)); + if (lb_iter == 1) // warn only once + XBT_WARN("process::load_balance() is a no-op!"); } -void process::finalize() +void process::send(neighbor& nb, double amount) { - 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)); - - while (may_receive()) { - comm.ctrl_flush(false); - comm.data_flush(false); - XBT_DEBUG("waiting for %d CTRL and %d DATA CLOSE", - ctrl_close_pending, data_close_pending); - receive(-1.0); - } - - comm.ctrl_flush(true); - comm.data_flush(true); + set_load(get_load() - amount); + nb.set_to_send(nb.get_to_send() + amount); + nb.set_load(nb.get_load() + amount); } #define print_loads_generic(vec, verbose, logp, cat) \ @@ -357,6 +265,109 @@ void process::print_loads_p(bool verbose, #undef print_loads_generic +void process::ctrl_send(neighbor& nb) +{ + double info_to_send = get_load(); + if (info_to_send != prev_load_broadcast) + comm.ctrl_send(nb.get_ctrl_mbox(), + new message(message::INFO, info_to_send)); + if (opt::bookkeeping) { + double 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); + comm.ctrl_send(nb.get_ctrl_mbox(), + new message(message::CREDIT, debt_to_send)); + } + } +} + +void process::data_send(neighbor& nb) +{ + double load_to_send; + if (opt::bookkeeping) { + if (real_load <= nb.get_debt()) { + load_to_send = real_load; + nb.set_debt(nb.get_debt() - load_to_send); + real_load = 0.0; + } else { + load_to_send = nb.get_debt(); + nb.set_debt(0.0); + real_load -= load_to_send; + } + } else { + load_to_send = nb.get_to_send(); + nb.set_to_send(0.0); + } + if (load_to_send > 0.0) + comm.data_send(nb.get_data_mbox(), + new message(message::LOAD, load_to_send)); +} + +void process::ctrl_close(neighbor& nb) +{ + comm.ctrl_send(nb.get_ctrl_mbox(), new message(message::CTRL_CLOSE, 0.0)); +} + +void process::data_close(neighbor& nb) +{ + comm.data_send(nb.get_data_mbox(), new message(message::DATA_CLOSE, 0.0)); +} + +void process::ctrl_receive(double timeout) +{ + message* msg; + m_host_t from; + + XBT_DEBUG("%sblocking receive on ctrl (%g)", "\0non-" + !timeout, timeout); + while (ctrl_close_pending && comm.ctrl_recv(msg, from, timeout)) { + handle_message(msg, from); + timeout = 0.0; + } +} + +void process::data_receive(double timeout) +{ + message* msg; + m_host_t from; + + XBT_DEBUG("%sblocking receive on data (%g)", "\0non-" + !timeout, timeout); + while (data_close_pending && comm.data_recv(msg, from, timeout)) { + handle_message(msg, from); + timeout = 0.0; + } +} + +void process::handle_message(message* msg, m_host_t from) +{ + 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; +} + // Local variables: // mode: c++ // End: diff --git a/process.h b/process.h index 38a6c42..dbf9b05 100644 --- a/process.h +++ b/process.h @@ -18,6 +18,7 @@ #include #include "communicator.h" #include "misc.h" +#include "msg_thread.h" #include "neighbor.h" #include "options.h" @@ -42,18 +43,22 @@ protected: pneigh_type pneigh; // list of pointers to neighbors that // we are free to reorder - template - void pneigh_sort_by_load(const Compare& comp); - // Get and set current load, which may be real load, or expected // load if opt::bookkeeping is true. double get_load() const; void set_load(double load); + // The load balancing algorithm comes here... + virtual void load_balance(); + // Register some amount of load to send to given neighbor. void send(neighbor& nb, double amount); void send(neighbor* nb, double amount) { send(*nb, amount); } + // Sort pneigh by applying comp to their loads + template + void pneigh_sort_by_load(const Compare& comp); + // Calls neighbor::print(verbose, logp, cat) for each member of neigh. void print_loads(bool verbose = false, e_xbt_log_priority_t logp = xbt_log_priority_info, @@ -92,31 +97,27 @@ private: double real_load; // current load double expected_load; // expected load in bookkeeping mode - // The load balancing algorithm comes here... - virtual void load_balance(); + // Load-balancing loop + void load_balance_loop(); - // Virtually do some computation - void compute(); + // Simulate computation loop + msg_thread* compute_thread; + void compute_loop(); - // Send procedures, with helpers for bookkeeping mode or not - void send1_no_bookkeeping(neighbor& nb); - void send1_bookkeeping(neighbor& nb); - void send_all(); + bool still_running(); - // Returns true if there remains neighbors to listen for - bool may_receive() const { - return ctrl_close_pending || data_close_pending; - } + // Send procedures + void ctrl_send(neighbor& nb); + void data_send(neighbor& nb); + void ctrl_close(neighbor& nb); + void data_close(neighbor& nb); // Receive procedure // Parameter "timeout" may be 0 for non-blocking operation, -1 for // infinite waiting, or any positive timeout. - void receive(double timeout); - - // Finalize sends a "close" message to each neighbor and wait for - // all of them to answer. - void finalize1(neighbor& nb); - void finalize(); + void ctrl_receive(double timeout); + void data_receive(double timeout); + void handle_message(message* msg, m_host_t from); }; inline -- 2.39.5