X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/85c342932c60954a7e1c9430742bfff2b28c31a8..b3b356352e87ae00a20f737c48e19b0c8413455a:/src/simdag/sd_task.c diff --git a/src/simdag/sd_task.c b/src/simdag/sd_task.c index fc156666b9..c55d588f79 100644 --- a/src/simdag/sd_task.c +++ b/src/simdag/sd_task.c @@ -1,11 +1,11 @@ -/* Copyright (c) 2006, 2007, 2008, 2009, 2010, 2011. The SimGrid Team. +/* Copyright (c) 2006-2015. The SimGrid Team. * All rights reserved. */ /* This program is free software; you can redistribute it and/or modify it * under the terms of the license (GNU LGPL) which comes with this package. */ #include "private.h" -#include "simdag/simdag.h" +#include "simgrid/simdag.h" #include "xbt/sysdep.h" #include "xbt/dynar.h" #include "instr/instr_private.h" @@ -55,8 +55,8 @@ void SD_task_recycle_f(void *t) /* scheduling parameters */ task->workstation_nb = 0; task->workstation_list = NULL; - task->computation_amount = NULL; - task->communication_amount = NULL; + task->flops_amount = NULL; + task->bytes_amount = NULL; task->rate = -1; } @@ -90,13 +90,139 @@ SD_task_t SD_task_create(const char *name, void *data, double amount) sd_global->task_number++; -#ifdef HAVE_TRACING TRACE_sd_task_create(task); -#endif return task; } +static XBT_INLINE SD_task_t SD_task_create_sized(const char *name, + void *data, double amount, + int ws_count) +{ + SD_task_t task = SD_task_create(name, data, amount); + task->bytes_amount = xbt_new0(double, ws_count * ws_count); + task->flops_amount = xbt_new0(double, ws_count); + task->workstation_nb = ws_count; + task->workstation_list = xbt_new0(SD_workstation_t, ws_count); + return task; +} + +/** @brief create a end-to-end communication task that can then be auto-scheduled + * + * Auto-scheduling mean that the task can be used with SD_task_schedulev(). This + * allows to specify the task costs at creation, and decouple them from the + * scheduling process where you just specify which resource should deliver the + * mandatory power. + * + * A end-to-end communication must be scheduled on 2 hosts, and the amount + * specified at creation is sent from hosts[0] to hosts[1]. + */ +SD_task_t SD_task_create_comm_e2e(const char *name, void *data, + double amount) +{ + SD_task_t res = SD_task_create_sized(name, data, amount, 2); + res->bytes_amount[2] = amount; + res->kind = SD_TASK_COMM_E2E; + + TRACE_category("COMM_E2E"); + TRACE_sd_set_task_category(res, "COMM_E2E"); + + return res; +} + +/** @brief create a sequential computation task that can then be auto-scheduled + * + * Auto-scheduling mean that the task can be used with SD_task_schedulev(). This + * allows to specify the task costs at creation, and decouple them from the + * scheduling process where you just specify which resource should deliver the + * mandatory power. + * + * A sequential computation must be scheduled on 1 host, and the amount + * specified at creation to be run on hosts[0]. + * + * \param name the name of the task (can be \c NULL) + * \param data the user data you want to associate with the task (can be \c NULL) + * \param flops_amount amount of compute work to be done by the task + * \return the new SD_TASK_COMP_SEQ typed task + */ +SD_task_t SD_task_create_comp_seq(const char *name, void *data, + double flops_amount) +{ + SD_task_t res = SD_task_create_sized(name, data, flops_amount, 1); + res->flops_amount[0] = flops_amount; + res->kind = SD_TASK_COMP_SEQ; + + TRACE_category("COMP_SEQ"); + TRACE_sd_set_task_category(res, "COMP_SEQ"); + + return res; +} + +/** @brief create a parallel computation task that can then be auto-scheduled + * + * Auto-scheduling mean that the task can be used with SD_task_schedulev(). This + * allows to specify the task costs at creation, and decouple them from the + * scheduling process where you just specify which resource should deliver the + * mandatory power. + * + * A parallel computation can be scheduled on any number of host. + * The underlying speedup model is Amdahl's law. + * To be auto-scheduled, \see SD_task_distribute_comp_amdahl has to be called + * first. + * \param name the name of the task (can be \c NULL) + * \param data the user data you want to associate with the task (can be \c NULL) + * \param flops_amount amount of compute work to be done by the task + * \param alpha purely serial fraction of the work to be done (in [0.;1.[) + * \return the new task + */ +SD_task_t SD_task_create_comp_par_amdahl(const char *name, void *data, + double flops_amount, double alpha) +{ + xbt_assert(alpha < 1. && alpha >= 0., + "Invalid parameter: alpha must be in [0.;1.["); + + SD_task_t res = SD_task_create(name, data, flops_amount); + res->alpha = alpha; + res->kind = SD_TASK_COMP_PAR_AMDAHL; + + TRACE_category("COMP_PAR_AMDAHL"); + TRACE_sd_set_task_category(res, "COMP_PAR_AMDAHL"); + + return res; +} + +/** @brief create a complex data redistribution task that can then be + * auto-scheduled + * + * Auto-scheduling mean that the task can be used with SD_task_schedulev(). + * This allows to specify the task costs at creation, and decouple them from + * the scheduling process where you just specify which resource should + * communicate. + * + * A data redistribution can be scheduled on any number of host. + * The assumed distribution is a 1D block distribution. Each host owns the same + * share of the \see amount. + * To be auto-scheduled, \see SD_task_distribute_comm_mxn_1d_block has to be + * called first. + * \param name the name of the task (can be \c NULL) + * \param data the user data you want to associate with the task (can be + * \c NULL) + * \param amount amount of data to redistribute by the task + * \return the new task + */ +SD_task_t SD_task_create_comm_par_mxn_1d_block(const char *name, void *data, + double amount) +{ + SD_task_t res = SD_task_create(name, data, amount); + res->workstation_list=NULL; + res->kind = SD_TASK_COMM_PAR_MXN_1D_BLOCK; + + TRACE_category("COMM_PAR_MXN_1D_BLOCK"); + TRACE_sd_set_task_category(res, "COMM_PAR_MXN_1D_BLOCK"); + + return res; +} + /** * \brief Destroys a task. * @@ -121,19 +247,17 @@ void SD_task_destroy(SD_task_t task) xbt_free(task->name); if (task->surf_action != NULL) - surf_workstation_model->action_unref(task->surf_action); + surf_action_unref(task->surf_action); xbt_free(task->workstation_list); - xbt_free(task->communication_amount); - xbt_free(task->computation_amount); + xbt_free(task->bytes_amount); + xbt_free(task->flops_amount); + + TRACE_sd_task_destroy(task); xbt_mallocator_release(sd_global->task_mallocator,task); sd_global->task_number--; -#ifdef HAVE_TRACING - TRACE_sd_task_destroy(task); -#endif - XBT_DEBUG("Task destroyed."); } @@ -164,6 +288,32 @@ void SD_task_set_data(SD_task_t task, void *data) task->data = data; } +/** + * \brief Sets the rate of a task + * + * This will change the network bandwidth a task can use. This rate + * cannot be dynamically changed. Once the task has started, this call + * is ineffective. This rate depends on both the nominal bandwidth on + * the route onto which the task is scheduled (\see + * SD_task_get_current_bandwidth) and the amount of data to transfer. + * + * To divide the nominal bandwidth by 2, the rate then has to be : + * rate = bandwidth/(2*amount) + * + * \param task a \see SD_TASK_COMM_E2E task (end-to-end communication) + * \param rate the new rate you want to associate with this task. + */ +void SD_task_set_rate(SD_task_t task, double rate) +{ + xbt_assert(task->kind == SD_TASK_COMM_E2E, + "The rate can be modified for end-to-end communications only."); + if(task->start_time<0) { + task->rate = rate; + } else { + XBT_WARN("Task %p has started. Changing rate is ineffective.", task); + } +} + /** * \brief Returns the state of a task * @@ -200,13 +350,11 @@ void __SD_task_set_state(SD_task_t task, e_SD_task_state_t new_state) break; case SD_RUNNING: task->state_set = sd_global->running_task_set; - task->start_time = - surf_workstation_model->action_get_start_time(task->surf_action); + task->start_time = surf_action_get_start_time(task->surf_action); break; case SD_DONE: task->state_set = sd_global->done_task_set; - task->finish_time = - surf_workstation_model->action_get_finish_time(task->surf_action); + task->finish_time = surf_action_get_finish_time(task->surf_action); task->remains = 0; #ifdef HAVE_JEDULE jedule_log_sd_event(task); @@ -317,6 +465,40 @@ double SD_task_get_amount(SD_task_t task) return task->amount; } +/** + * \brief Sets the total amount of work of a task + * For sequential typed tasks (COMP_SEQ and COMM_E2E), it also sets the + * appropriate values in the flops_amount and bytes_amount arrays + * respectively. Nothing more than modifying task->amount is done for paralle + * typed tasks (COMP_PAR_AMDAHL and COMM_PAR_MXN_1D_BLOCK) as the distribution + * of the amount of work is done at scheduling time. + * + * \param task a task + * \param amount the new amount of work to execute + */ +void SD_task_set_amount(SD_task_t task, double amount) +{ + task->amount = amount; + if (task->kind == SD_TASK_COMP_SEQ) + task->flops_amount[0] = amount; + if (task->kind == SD_TASK_COMM_E2E) + task->bytes_amount[2] = amount; +} + +/** + * \brief Returns the alpha parameter of a SD_TASK_COMP_PAR_AMDAHL task + * + * \param task a parallel task assuming Amdahl's law as speedup model + * \return the alpha parameter (serial part of a task in percent) for this task + */ +double SD_task_get_alpha(SD_task_t task) +{ + xbt_assert(SD_task_get_kind(task) == SD_TASK_COMP_PAR_AMDAHL, + "Alpha parameter is not defined for this kink of task"); + return task->alpha; +} + + /** * \brief Returns the remaining amount work to do till the completion of a task * @@ -327,7 +509,7 @@ double SD_task_get_amount(SD_task_t task) double SD_task_get_remaining_amount(SD_task_t task) { if (task->surf_action) - return surf_workstation_model->get_remains(task->surf_action); + return surf_action_get_remains(task->surf_action); else return task->remains; } @@ -346,16 +528,16 @@ void SD_task_dump(SD_task_t task) XBT_INFO("Displaying task %s", SD_task_get_name(task)); statename = bprintf("%s %s %s %s %s %s %s %s", - (task->state & SD_NOT_SCHEDULED ? "not scheduled" : + (task->state == SD_NOT_SCHEDULED ? "not scheduled" : ""), - (task->state & SD_SCHEDULABLE ? "schedulable" : ""), - (task->state & SD_SCHEDULED ? "scheduled" : ""), - (task->state & SD_RUNNABLE ? "runnable" : + (task->state == SD_SCHEDULABLE ? "schedulable" : ""), + (task->state == SD_SCHEDULED ? "scheduled" : ""), + (task->state == SD_RUNNABLE ? "runnable" : "not runnable"), - (task->state & SD_IN_FIFO ? "in fifo" : ""), - (task->state & SD_RUNNING ? "running" : ""), - (task->state & SD_DONE ? "done" : ""), - (task->state & SD_FAILED ? "failed" : "")); + (task->state == SD_IN_FIFO ? "in fifo" : ""), + (task->state == SD_RUNNING ? "running" : ""), + (task->state == SD_DONE ? "done" : ""), + (task->state == SD_FAILED ? "failed" : "")); XBT_INFO(" - state: %s", statename); free(statename); @@ -367,19 +549,31 @@ void SD_task_dump(SD_task_t task) case SD_TASK_COMP_SEQ: XBT_INFO(" - kind: sequential computation"); break; + case SD_TASK_COMP_PAR_AMDAHL: + XBT_INFO(" - kind: parallel computation following Amdahl's law"); + break; + case SD_TASK_COMM_PAR_MXN_1D_BLOCK: + XBT_INFO(" - kind: MxN data redistribution assuming 1D block distribution"); + break; default: XBT_INFO(" - (unknown kind %d)", task->kind); } } + + if (task->category) + XBT_INFO(" - tracing category: %s", task->category); + XBT_INFO(" - amount: %.0f", SD_task_get_amount(task)); + if (task->kind == SD_TASK_COMP_PAR_AMDAHL) + XBT_INFO(" - alpha: %.2f", task->alpha); XBT_INFO(" - Dependencies to satisfy: %d", task->unsatisfied_dependencies); - if (xbt_dynar_length(task->tasks_before)) { + if (!xbt_dynar_is_empty(task->tasks_before)) { XBT_INFO(" - pre-dependencies:"); xbt_dynar_foreach(task->tasks_before, counter, dependency) { XBT_INFO(" %s", SD_task_get_name(dependency->src)); } } - if (xbt_dynar_length(task->tasks_after)) { + if (!xbt_dynar_is_empty(task->tasks_after)) { XBT_INFO(" - post-dependencies:"); xbt_dynar_foreach(task->tasks_after, counter, dependency) { XBT_INFO(" %s", SD_task_get_name(dependency->dst)); @@ -395,9 +589,11 @@ void SD_task_dotty(SD_task_t task, void *out) fprintf(out, " T%p [label=\"%.20s\"", task, task->name); switch (task->kind) { case SD_TASK_COMM_E2E: + case SD_TASK_COMM_PAR_MXN_1D_BLOCK: fprintf(out, ", shape=box"); break; case SD_TASK_COMP_SEQ: + case SD_TASK_COMP_PAR_AMDAHL: fprintf(out, ", shape=circle"); break; default: @@ -433,9 +629,9 @@ void SD_task_dependency_add(const char *name, void *data, SD_task_t src, SD_task_t dst) { xbt_dynar_t dynar; - int length; + unsigned long length; int found = 0; - int i; + unsigned long i; SD_dependency_t dependency; dynar = src->tasks_after; @@ -447,9 +643,10 @@ void SD_task_dependency_add(const char *name, void *data, SD_task_t src, SD_task_get_name(src)); if (!__SD_task_is_not_scheduled(src) && !__SD_task_is_schedulable(src) - && !__SD_task_is_scheduled_or_runnable(src)) + && !__SD_task_is_scheduled_or_runnable(src) && !__SD_task_is_running(src)) THROWF(arg_error, 0, - "Task '%s' must be SD_NOT_SCHEDULED, SD_SCHEDULABLE, SD_SCHEDULED or SD_RUNNABLE", + "Task '%s' must be SD_NOT_SCHEDULED, SD_SCHEDULABLE, SD_SCHEDULED or SD_RUNNABLE" + " or SD_RUNNING", SD_task_get_name(src)); if (!__SD_task_is_not_scheduled(dst) && !__SD_task_is_schedulable(dst) @@ -463,7 +660,7 @@ void SD_task_dependency_add(const char *name, void *data, SD_task_t src, for (i = 0; i < length && !found; i++) { xbt_dynar_get_cpy(dynar, i, &dependency); found = (dependency->dst == dst); - XBT_DEBUG("Dependency %d: dependency->dst = %s", i, + XBT_DEBUG("Dependency %lu: dependency->dst = %s", i, SD_task_get_name(dependency->dst)); } @@ -494,13 +691,27 @@ void SD_task_dependency_add(const char *name, void *data, SD_task_t src, SD_task_get_name(dst)); __SD_task_set_state(dst, SD_SCHEDULED); } +} +/** + * \brief Returns the name given as input when dependency has been created.. + * + * \param src a task + * \param dst a task depending on \a src + * + */ +const char *SD_task_dependency_get_name(SD_task_t src, SD_task_t dst){ + unsigned int i; + SD_dependency_t dependency; - /* __SD_print_dependencies(src); - __SD_print_dependencies(dst); */ + xbt_dynar_foreach(src->tasks_after, i, dependency){ + if (dependency->dst == dst) + return dependency->name; + } + return NULL; } /** - * \brief Indacates whether there is a dependency between two tasks. + * \brief Indicates whether there is a dependency between two tasks. * * \param src a task * \param dst a task depending on \a src @@ -543,9 +754,9 @@ void SD_task_dependency_remove(SD_task_t src, SD_task_t dst) { xbt_dynar_t dynar; - int length; + unsigned long length; int found = 0; - int i; + unsigned long i; SD_dependency_t dependency; /* remove the dependency from src->tasks_after */ @@ -614,9 +825,9 @@ void *SD_task_dependency_get_data(SD_task_t src, SD_task_t dst) { xbt_dynar_t dynar; - int length; + unsigned long length; int found = 0; - int i; + unsigned long i; SD_dependency_t dependency; dynar = src->tasks_after; @@ -702,16 +913,16 @@ void SD_task_unwatch(SD_task_t task, e_SD_task_state_t state) * \param task the task to evaluate * \param workstation_nb number of workstations on which the task would be executed * \param workstation_list the workstations on which the task would be executed - * \param computation_amount computation amount for each workstation - * \param communication_amount communication amount between each pair of workstations + * \param flops_amount computation amount for each workstation + * \param bytes_amount communication amount between each pair of workstations * \see SD_schedule() */ double SD_task_get_execution_time(SD_task_t task, int workstation_nb, const SD_workstation_t * workstation_list, - const double *computation_amount, - const double *communication_amount) + const double *flops_amount, + const double *bytes_amount) { double time, max_time = 0.0; int i, j; @@ -721,17 +932,17 @@ double SD_task_get_execution_time(SD_task_t task, for (i = 0; i < workstation_nb; i++) { time = 0.0; - if (computation_amount != NULL) + if (flops_amount != NULL) time = SD_workstation_get_computation_time(workstation_list[i], - computation_amount[i]); + flops_amount[i]); - if (communication_amount != NULL) + if (bytes_amount != NULL) for (j = 0; j < workstation_nb; j++) { time += SD_route_get_communication_time(workstation_list[i], workstation_list[j], - communication_amount[i * + bytes_amount[i * workstation_nb + j]); } @@ -766,15 +977,15 @@ static XBT_INLINE void SD_task_do_schedule(SD_task_t task) * \param task the task you want to schedule * \param workstation_count number of workstations on which the task will be executed * \param workstation_list the workstations on which the task will be executed - * \param computation_amount computation amount for each workstation - * \param communication_amount communication amount between each pair of workstations + * \param flops_amount computation amount for each workstation + * \param bytes_amount communication amount between each pair of workstations * \param rate task execution speed rate * \see SD_task_unschedule() */ void SD_task_schedule(SD_task_t task, int workstation_count, const SD_workstation_t * workstation_list, - const double *computation_amount, - const double *communication_amount, double rate) + const double *flops_amount, + const double *bytes_amount, double rate) { int communication_nb; task->workstation_nb = 0; @@ -784,25 +995,25 @@ void SD_task_schedule(SD_task_t task, int workstation_count, task->workstation_nb = workstation_count; task->rate = rate; - if (computation_amount) { - task->computation_amount = xbt_realloc(task->computation_amount, + if (flops_amount) { + task->flops_amount = xbt_realloc(task->flops_amount, sizeof(double) * workstation_count); - memcpy(task->computation_amount, computation_amount, + memcpy(task->flops_amount, flops_amount, sizeof(double) * workstation_count); } else { - xbt_free(task->computation_amount); - task->computation_amount = NULL; + xbt_free(task->flops_amount); + task->flops_amount = NULL; } communication_nb = workstation_count * workstation_count; - if (communication_amount) { - task->communication_amount = xbt_realloc(task->communication_amount, + if (bytes_amount) { + task->bytes_amount = xbt_realloc(task->bytes_amount, sizeof(double) * communication_nb); - memcpy(task->communication_amount, communication_amount, + memcpy(task->bytes_amount, bytes_amount, sizeof(double) * communication_nb); } else { - xbt_free(task->communication_amount); - task->communication_amount = NULL; + xbt_free(task->bytes_amount); + task->bytes_amount = NULL; } task->workstation_list = @@ -835,11 +1046,16 @@ void SD_task_unschedule(SD_task_t task) SD_task_get_name(task)); if (__SD_task_is_scheduled_or_runnable(task) /* if the task is scheduled or runnable */ - &&task->kind == SD_TASK_NOT_TYPED) /* Don't free scheduling data for typed tasks */ + && ((task->kind == SD_TASK_COMP_PAR_AMDAHL) || + (task->kind == SD_TASK_COMM_PAR_MXN_1D_BLOCK))) { /* Don't free scheduling data for typed tasks */ __SD_task_destroy_scheduling_data(task); + xbt_free(task->workstation_list); + task->workstation_list=NULL; + task->workstation_nb = 0; + } if (__SD_task_is_running(task)) /* the task should become SD_FAILED */ - surf_workstation_model->action_cancel(task->surf_action); + surf_action_cancel(task->surf_action); else { if (task->unsatisfied_dependencies == 0) __SD_task_set_state(task, SD_SCHEDULABLE); @@ -850,7 +1066,8 @@ void SD_task_unschedule(SD_task_t task) task->start_time = -1.0; } -/* Destroys the data memorized by SD_task_schedule. Task state must be SD_SCHEDULED or SD_RUNNABLE. +/* Destroys the data memorized by SD_task_schedule. + * Task state must be SD_SCHEDULED or SD_RUNNABLE. */ static void __SD_task_destroy_scheduling_data(SD_task_t task) { @@ -860,37 +1077,36 @@ static void __SD_task_destroy_scheduling_data(SD_task_t task) "Task '%s' must be SD_SCHEDULED, SD_RUNNABLE or SD_IN_FIFO", SD_task_get_name(task)); - xbt_free(task->computation_amount); - xbt_free(task->communication_amount); - task->computation_amount = task->communication_amount = NULL; + xbt_free(task->flops_amount); + xbt_free(task->bytes_amount); + task->flops_amount = task->bytes_amount = NULL; } -/* Runs a task. This function is directly called by __SD_task_try_to_run if the task - * doesn't have to wait in fifos. Otherwise, it is called by __SD_task_just_done when - * the task gets out of its fifos. +/* Runs a task. This function is directly called by __SD_task_try_to_run if + * the task doesn't have to wait in FIFOs. Otherwise, it is called by + * __SD_task_just_done when the task gets out of its FIFOs. */ void __SD_task_really_run(SD_task_t task) { int i; - void **surf_workstations; + sg_host_t *hosts; xbt_assert(__SD_task_is_runnable_or_in_fifo(task), "Task '%s' is not runnable or in a fifo! Task state: %d", - SD_task_get_name(task), SD_task_get_state(task)); + SD_task_get_name(task), (int)SD_task_get_state(task)); xbt_assert(task->workstation_list != NULL, "Task '%s': workstation_list is NULL!", SD_task_get_name(task)); - - XBT_DEBUG("Really running task '%s'", SD_task_get_name(task)); + int host_nb = task->workstation_nb; /* set this task as current task for the workstations in sequential mode */ - for (i = 0; i < task->workstation_nb; i++) { + for (i = 0; i < host_nb; i++) { if (SD_workstation_get_access_mode(task->workstation_list[i]) == SD_WORKSTATION_SEQUENTIAL_ACCESS) { - task->workstation_list[i]->current_task = task; + sg_host_sd(task->workstation_list[i])->current_task = task; xbt_assert(__SD_workstation_is_busy(task->workstation_list[i]), "The workstation should be busy now"); } @@ -901,92 +1117,49 @@ void __SD_task_really_run(SD_task_t task) /* start the task */ - /* we have to create a Surf workstation array instead of the SimDag workstation array */ - surf_workstations = xbt_new(void *, task->workstation_nb); + /* Copy the elements of the task into the action */ + hosts = xbt_new(sg_host_t, host_nb); - for (i = 0; i < task->workstation_nb; i++) - surf_workstations[i] = task->workstation_list[i]->surf_workstation; + for (i = 0; i < host_nb; i++) + hosts[i] = task->workstation_list[i]; - /* It's allowed to pass a NULL vector as cost to mean vector of 0.0 (easing user's life). Let's deal with it */ -#define cost_or_zero(array,pos) ((array)?(array)[pos]:0.0) + double *flops_amount = xbt_new0(double, host_nb); + double *bytes_amount = xbt_new0(double, host_nb * host_nb); - task->surf_action = NULL; - if ((task->workstation_nb == 1) - && (cost_or_zero(task->communication_amount, 0) == 0.0)) { - task->surf_action = - surf_workstation_model->extension. - workstation.execute(surf_workstations[0], - cost_or_zero(task->computation_amount, 0)); - } else if ((task->workstation_nb == 1) - && (cost_or_zero(task->computation_amount, 0) == 0.0)) { - - task->surf_action = - surf_workstation_model->extension. - workstation.communicate(surf_workstations[0], surf_workstations[0], - cost_or_zero(task->communication_amount, - 0), task->rate); - } else if ((task->workstation_nb == 2) - && (cost_or_zero(task->computation_amount, 0) == 0.0) - && (cost_or_zero(task->computation_amount, 1) == 0.0)) { - int nb = 0; - double value = 0.0; - - for (i = 0; i < task->workstation_nb * task->workstation_nb; i++) { - if (cost_or_zero(task->communication_amount, i) > 0.0) { - nb++; - value = cost_or_zero(task->communication_amount, i); - } - } - if (nb == 1) { - task->surf_action = - surf_workstation_model->extension. - workstation.communicate(surf_workstations[0], - surf_workstations[1], value, task->rate); - } - } -#undef cost_or_zero - - if (!task->surf_action) { - double *computation_amount = xbt_new(double, task->workstation_nb); - double *communication_amount = xbt_new(double, task->workstation_nb * - task->workstation_nb); - - memcpy(computation_amount, task->computation_amount, sizeof(double) * - task->workstation_nb); - memcpy(communication_amount, task->communication_amount, - sizeof(double) * task->workstation_nb * task->workstation_nb); - - task->surf_action = - surf_workstation_model->extension. - workstation.execute_parallel_task(task->workstation_nb, - surf_workstations, - computation_amount, - communication_amount, - task->amount, task->rate); - } else { - xbt_free(surf_workstations); - } - surf_workstation_model->action_data_set(task->surf_action, task); + if(task->flops_amount) + memcpy(flops_amount, task->flops_amount, sizeof(double) * + host_nb); + if(task->bytes_amount) + memcpy(bytes_amount, task->bytes_amount, + sizeof(double) * host_nb * host_nb); + + task->surf_action = surf_host_model_execute_parallel_task((surf_host_model_t)surf_host_model, + host_nb, + hosts, + flops_amount, + bytes_amount, + task->rate); + + surf_action_set_data(task->surf_action, task); XBT_DEBUG("surf_action = %p", task->surf_action); -#ifdef HAVE_TRACING if (task->category) TRACE_surf_action(task->surf_action, task->category); -#endif __SD_task_destroy_scheduling_data(task); /* now the scheduling data are not useful anymore */ __SD_task_set_state(task, SD_RUNNING); xbt_assert(__SD_task_is_running(task), "Bad state of task '%s': %d", - SD_task_get_name(task), SD_task_get_state(task)); + SD_task_get_name(task), (int)SD_task_get_state(task)); } -/* Tries to run a task. This function is called by SD_simulate() when a scheduled task becomes SD_RUNNABLE - * (ie when its dependencies are satisfied). - * If one of the workstations where the task is scheduled on is busy (in sequential mode), - * the task doesn't start. +/* Tries to run a task. This function is called by SD_simulate() when a + * scheduled task becomes SD_RUNNABLE (i.e., when its dependencies are + * satisfied). + * If one of the workstations where the task is scheduled on is busy (in + * sequential mode), the task doesn't start. * Returns whether the task has started. */ int __SD_task_try_to_run(SD_task_t task) @@ -998,7 +1171,7 @@ int __SD_task_try_to_run(SD_task_t task) xbt_assert(__SD_task_is_runnable(task), "Task '%s' is not runnable! Task state: %d", - SD_task_get_name(task), SD_task_get_state(task)); + SD_task_get_name(task), (int)SD_task_get_state(task)); for (i = 0; i < task->workstation_nb; i++) { @@ -1008,19 +1181,19 @@ int __SD_task_try_to_run(SD_task_t task) XBT_DEBUG("Task '%s' can start: %d", SD_task_get_name(task), can_start); - if (!can_start) { /* if the task cannot start and is not in the fifos yet */ + if (!can_start) { /* if the task cannot start and is not in the FIFOs yet */ for (i = 0; i < task->workstation_nb; i++) { workstation = task->workstation_list[i]; - if (workstation->access_mode == SD_WORKSTATION_SEQUENTIAL_ACCESS) { - XBT_DEBUG("Pushing task '%s' in the fifo of workstation '%s'", + if (sg_host_sd(workstation)->access_mode == SD_WORKSTATION_SEQUENTIAL_ACCESS) { + XBT_DEBUG("Pushing task '%s' in the FIFO of workstation '%s'", SD_task_get_name(task), SD_workstation_get_name(workstation)); - xbt_fifo_push(workstation->task_fifo, task); + xbt_fifo_push(sg_host_sd(workstation)->task_fifo, task); } } __SD_task_set_state(task, SD_IN_FIFO); xbt_assert(__SD_task_is_in_fifo(task), "Bad state of task '%s': %d", - SD_task_get_name(task), SD_task_get_state(task)); + SD_task_get_name(task), (int)SD_task_get_state(task)); XBT_DEBUG("Task '%s' state is now SD_IN_FIFO", SD_task_get_name(task)); } else { __SD_task_really_run(task); @@ -1031,7 +1204,7 @@ int __SD_task_try_to_run(SD_task_t task) /* This function is called by SD_simulate when a task is done. * It updates task->state and task->action and executes if necessary the tasks - * which were waiting in fifos for the end of `task' + * which were waiting in FIFOs for the end of `task' */ void __SD_task_just_done(SD_task_t task) { @@ -1046,7 +1219,7 @@ void __SD_task_just_done(SD_task_t task) xbt_assert(__SD_task_is_running(task), "The task must be running! Task state: %d", - SD_task_get_state(task)); + (int)SD_task_get_state(task)); xbt_assert(task->workstation_list != NULL, "Task '%s': workstation_list is NULL!", SD_task_get_name(task)); @@ -1055,40 +1228,40 @@ void __SD_task_just_done(SD_task_t task) candidates = xbt_new(SD_task_t, 8); __SD_task_set_state(task, SD_DONE); - surf_workstation_model->action_unref(task->surf_action); + surf_action_unref(task->surf_action); task->surf_action = NULL; XBT_DEBUG("Looking for candidates"); /* if the task was executed on sequential workstations, - maybe we can execute the next task of the fifo for each workstation */ + maybe we can execute the next task of the FIFO for each workstation */ for (i = 0; i < task->workstation_nb; i++) { workstation = task->workstation_list[i]; XBT_DEBUG("Workstation '%s': access_mode = %d", - SD_workstation_get_name(workstation), workstation->access_mode); - if (workstation->access_mode == SD_WORKSTATION_SEQUENTIAL_ACCESS) { - xbt_assert(workstation->task_fifo != NULL, - "Workstation '%s' has sequential access but no fifo!", + SD_workstation_get_name(workstation), (int)sg_host_sd(workstation)->access_mode); + if (sg_host_sd(workstation)->access_mode == SD_WORKSTATION_SEQUENTIAL_ACCESS) { + xbt_assert(sg_host_sd(workstation)->task_fifo != NULL, + "Workstation '%s' has sequential access but no FIFO!", SD_workstation_get_name(workstation)); - xbt_assert(workstation->current_task = + xbt_assert(sg_host_sd(workstation)->current_task = task, "Workstation '%s': current task should be '%s'", SD_workstation_get_name(workstation), SD_task_get_name(task)); /* the task is over so we can release the workstation */ - workstation->current_task = NULL; + sg_host_sd(workstation)->current_task = NULL; - XBT_DEBUG("Getting candidate in fifo"); + XBT_DEBUG("Getting candidate in FIFO"); candidate = xbt_fifo_get_item_content(xbt_fifo_get_first_item - (workstation->task_fifo)); + (sg_host_sd(workstation)->task_fifo)); if (candidate != NULL) { XBT_DEBUG("Candidate: '%s'", SD_task_get_name(candidate)); xbt_assert(__SD_task_is_in_fifo(candidate), "Bad state of candidate '%s': %d", SD_task_get_name(candidate), - SD_task_get_state(candidate)); + (int)SD_task_get_state(candidate)); } XBT_DEBUG("Candidate in fifo: %p", candidate); @@ -1096,9 +1269,10 @@ void __SD_task_just_done(SD_task_t task) /* if there was a task waiting for my place */ if (candidate != NULL) { /* Unfortunately, we are not sure yet that we can execute the task now, - because the task can be waiting more deeply in some other workstation's fifos... - So we memorize all candidate tasks, and then we will check for each candidate - whether or not all its workstations are available. */ + because the task can be waiting more deeply in some other + workstation's FIFOs ... + So we memorize all candidate tasks, and then we will check for each + candidate whether or not all its workstations are available. */ /* realloc if necessary */ if (candidate_nb == candidate_capacity) { @@ -1127,17 +1301,17 @@ void __SD_task_just_done(SD_task_t task) xbt_assert(__SD_task_is_in_fifo(candidate), "Bad state of candidate '%s': %d", - SD_task_get_name(candidate), SD_task_get_state(candidate)); + SD_task_get_name(candidate), (int)SD_task_get_state(candidate)); for (j = 0; j < candidate->workstation_nb && can_start; j++) { workstation = candidate->workstation_list[j]; /* I can start on this workstation if the workstation is shared - or if I am the first task in the fifo */ - can_start = workstation->access_mode == SD_WORKSTATION_SHARED_ACCESS + or if I am the first task in the FIFO */ + can_start = sg_host_sd(workstation)->access_mode == SD_WORKSTATION_SHARED_ACCESS || candidate == xbt_fifo_get_item_content(xbt_fifo_get_first_item - (workstation->task_fifo)); + (sg_host_sd(workstation)->task_fifo)); } XBT_DEBUG("Candidate '%s' can start: %d", SD_task_get_name(candidate), @@ -1148,20 +1322,20 @@ void __SD_task_just_done(SD_task_t task) for (j = 0; j < candidate->workstation_nb && can_start; j++) { workstation = candidate->workstation_list[j]; - /* update the fifo */ - if (workstation->access_mode == SD_WORKSTATION_SEQUENTIAL_ACCESS) { - candidate = xbt_fifo_shift(workstation->task_fifo); /* the return value is stored just for debugging */ - XBT_DEBUG("Head of the fifo: '%s'", + /* update the FIFO */ + if (sg_host_sd(workstation)->access_mode == SD_WORKSTATION_SEQUENTIAL_ACCESS) { + candidate = xbt_fifo_shift(sg_host_sd(workstation)->task_fifo); /* the return value is stored just for debugging */ + XBT_DEBUG("Head of the FIFO: '%s'", (candidate != NULL) ? SD_task_get_name(candidate) : "NULL"); xbt_assert(candidate == candidates[i], - "Error in __SD_task_just_done: bad first task in the fifo"); + "Error in __SD_task_just_done: bad first task in the FIFO"); } } /* for each workstation */ /* finally execute the task */ XBT_DEBUG("Task '%s' state: %d", SD_task_get_name(candidate), - SD_task_get_state(candidate)); + (int)SD_task_get_state(candidate)); __SD_task_really_run(candidate); XBT_DEBUG @@ -1171,7 +1345,7 @@ void __SD_task_just_done(SD_task_t task) xbt_assert(__SD_task_is_running(candidate), "Bad state of task '%s': %d", SD_task_get_name(candidate), - SD_task_get_state(candidate)); + (int)SD_task_get_state(candidate)); XBT_DEBUG("Okay, the task is running."); } /* can start */ @@ -1181,19 +1355,21 @@ void __SD_task_just_done(SD_task_t task) xbt_free(candidates); } -/* Remove all dependencies associated with a task. This function is called when the task is destroyed. +/* + * Remove all dependencies associated with a task. This function is called + * when the task is destroyed. */ static void __SD_task_remove_dependencies(SD_task_t task) { /* we must destroy the dependencies carefuly (with SD_dependency_remove) because each one is stored twice */ SD_dependency_t dependency; - while (xbt_dynar_length(task->tasks_before) > 0) { + while (!xbt_dynar_is_empty(task->tasks_before)) { xbt_dynar_get_cpy(task->tasks_before, 0, &dependency); SD_task_dependency_remove(dependency->src, dependency->dst); } - while (xbt_dynar_length(task->tasks_after) > 0) { + while (!xbt_dynar_is_empty(task->tasks_after)) { xbt_dynar_get_cpy(task->tasks_after, 0, &dependency); SD_task_dependency_remove(dependency->src, dependency->dst); } @@ -1210,8 +1386,7 @@ static void __SD_task_remove_dependencies(SD_task_t task) double SD_task_get_start_time(SD_task_t task) { if (task->surf_action) - return surf_workstation_model-> - action_get_start_time(task->surf_action); + return surf_action_get_start_time(task->surf_action); else return task->start_time; } @@ -1230,96 +1405,78 @@ double SD_task_get_start_time(SD_task_t task) double SD_task_get_finish_time(SD_task_t task) { if (task->surf_action) /* should never happen as actions are destroyed right after their completion */ - return surf_workstation_model-> - action_get_finish_time(task->surf_action); + return surf_action_get_finish_time(task->surf_action); else return task->finish_time; } - -static XBT_INLINE SD_task_t SD_task_create_sized(const char *name, - void *data, double amount, - int ws_count) -{ - SD_task_t task = SD_task_create(name, data, amount); - task->communication_amount = xbt_new0(double, ws_count * ws_count); - task->computation_amount = xbt_new0(double, ws_count); - task->workstation_nb = ws_count; - task->workstation_list = xbt_new0(SD_workstation_t, ws_count); - return task; -} - -/** @brief create a end-to-end communication task that can then be auto-scheduled +/** @brief Blah * - * Auto-scheduling mean that the task can be used with SD_task_schedulev(). This - * allows to specify the task costs at creation, and decorelate them from the - * scheduling process where you just specify which resource should deliver the - * mandatory power. - * - * A end-to-end communication must be scheduled on 2 hosts, and the amount - * specified at creation is sent from hosts[0] to hosts[1]. */ -SD_task_t SD_task_create_comm_e2e(const char *name, void *data, - double amount) +void SD_task_distribute_comp_amdahl(SD_task_t task, int ws_count) { - SD_task_t res = SD_task_create_sized(name, data, amount, 2); - res->communication_amount[2] = amount; - res->kind = SD_TASK_COMM_E2E; - return res; -} + int i; + xbt_assert(task->kind == SD_TASK_COMP_PAR_AMDAHL, + "Task %s is not a SD_TASK_COMP_PAR_AMDAHL typed task." + "Cannot use this function.", + SD_task_get_name(task)); + task->flops_amount = xbt_new0(double, ws_count); + task->bytes_amount = xbt_new0(double, ws_count * ws_count); + xbt_free(task->workstation_list); + task->workstation_nb = ws_count; + task->workstation_list = xbt_new0(SD_workstation_t, ws_count); + + for(i=0;iflops_amount[i] = + (task->alpha + (1 - task->alpha)/ws_count) * task->amount; + } +} -/** @brief create a sequential computation task that can then be auto-scheduled - * - * Auto-scheduling mean that the task can be used with SD_task_schedulev(). This - * allows to specify the task costs at creation, and decorelate them from the - * scheduling process where you just specify which resource should deliver the - * mandatory power. - * - * A sequential computation must be scheduled on 1 host, and the amount - * specified at creation to be run on hosts[0]. - */ -SD_task_t SD_task_create_comp_seq(const char *name, void *data, - double amount) -{ - SD_task_t res = SD_task_create_sized(name, data, amount, 1); - res->computation_amount[0] = amount; - res->kind = SD_TASK_COMP_SEQ; - return res; -} /** @brief Auto-schedules a task. * * Auto-scheduling mean that the task can be used with SD_task_schedulev(). This - * allows to specify the task costs at creation, and decorelate them from the + * allows to specify the task costs at creation, and decouple them from the * scheduling process where you just specify which resource should deliver the * mandatory power. * - * To be auto-schedulable, a task must be created with SD_task_create_comm_e2e() or - * SD_task_create_comp_seq(). Check their definitions for the exact semantic of each - * of them. + * To be auto-schedulable, a task must be created with SD_task_create_comm_e2e() + * or SD_task_create_comp_seq(). Check their definitions for the exact semantic + * of each of them. * * @todo * We should create tasks kind for the following categories: * - Point to point communication (done) * - Sequential computation (done) * - group communication (redistribution, several kinds) - * - parallel tasks with no internal communication (one kind per speedup model such as amdal) - * - idem+ internal communication. Task type not enough since we cannot store comm cost alongside to comp one) + * - parallel tasks with no internal communication (one kind per speedup + * model such as Amdahl) + * - idem+ internal communication. Task type not enough since we cannot store + * comm cost alongside to comp one) */ void SD_task_schedulev(SD_task_t task, int count, const SD_workstation_t * list) { - int i; + int i, j; SD_dependency_t dep; unsigned int cpt; xbt_assert(task->kind != 0, "Task %s is not typed. Cannot automatically schedule it.", SD_task_get_name(task)); switch (task->kind) { + case SD_TASK_COMP_PAR_AMDAHL: + SD_task_distribute_comp_amdahl(task, count); case SD_TASK_COMM_E2E: case SD_TASK_COMP_SEQ: - xbt_assert(task->workstation_nb == count,"Got %d locations, but were expecting %d locations",count,task->workstation_nb); + xbt_assert(task->workstation_nb == count, + "Got %d locations, but were expecting %d locations", + count,task->workstation_nb); for (i = 0; i < count; i++) task->workstation_list[i] = list[i]; + if (SD_task_get_kind(task)== SD_TASK_COMP_SEQ && !task->flops_amount){ + /*This task has failed and is rescheduled. Reset the flops_amount*/ + task->flops_amount = xbt_new0(double, 1); + task->flops_amount[0] = task->remains; + } SD_task_do_schedule(task); break; default: @@ -1331,15 +1488,17 @@ void SD_task_schedulev(SD_task_t task, int count, SD_task_get_name(task), SD_workstation_get_name(task->workstation_list[0]), SD_workstation_get_name(task->workstation_list[1]), - task->communication_amount[2]); + task->bytes_amount[2]); } - /* Iterate over all childs and parent being COMM_E2E to say where I am located (and start them if runnable) */ + + /* Iterate over all children and parents being COMM_E2E to say where I am + * located (and start them if runnable) */ if (task->kind == SD_TASK_COMP_SEQ) { XBT_VERB("Schedule computation task %s on %s. It costs %.f flops", SD_task_get_name(task), SD_workstation_get_name(task->workstation_list[0]), - task->computation_amount[0]); + task->flops_amount[0]); xbt_dynar_foreach(task->tasks_before, cpt, dep) { SD_task_t before = dep->src; @@ -1355,7 +1514,7 @@ void SD_task_schedulev(SD_task_t task, int count, SD_task_get_name(before), SD_workstation_get_name(before->workstation_list[0]), SD_workstation_get_name(before->workstation_list[1]), - before->communication_amount[2]); + before->bytes_amount[2]); } } } @@ -1363,7 +1522,6 @@ void SD_task_schedulev(SD_task_t task, int count, SD_task_t after = dep->dst; if (after->kind == SD_TASK_COMM_E2E) { after->workstation_list[0] = task->workstation_list[0]; - //J-N : Why did you comment on these line (this comment add a bug I think)? if (after->workstation_list[1] && (__SD_task_is_not_scheduled(after) || __SD_task_is_schedulable(after))) { @@ -1373,12 +1531,150 @@ void SD_task_schedulev(SD_task_t task, int count, SD_task_get_name(after), SD_workstation_get_name(after->workstation_list[0]), SD_workstation_get_name(after->workstation_list[1]), - after->communication_amount[2]); + after->bytes_amount[2]); } } } } + /* Iterate over all children and parents being MXN_1D_BLOCK to say where I am + * located (and start them if runnable) */ + if (task->kind == SD_TASK_COMP_PAR_AMDAHL) { + XBT_VERB("Schedule computation task %s on %d workstations. %.f flops" + " will be distributed following Amdahl's Law", + SD_task_get_name(task), task->workstation_nb, + task->flops_amount[0]); + xbt_dynar_foreach(task->tasks_before, cpt, dep) { + SD_task_t before = dep->src; + if (before->kind == SD_TASK_COMM_PAR_MXN_1D_BLOCK){ + if (!before->workstation_list){ + XBT_VERB("Sender side of Task %s is not scheduled yet", + SD_task_get_name(before)); + before->workstation_list = xbt_new0(SD_workstation_t, count); + before->workstation_nb = count; + XBT_VERB("Fill the workstation list with list of Task '%s'", + SD_task_get_name(task)); + for (i=0;iworkstation_list[i] = task->workstation_list[i]; + } else { + XBT_VERB("Build communication matrix for task '%s'", + SD_task_get_name(before)); + int src_nb, dst_nb; + double src_start, src_end, dst_start, dst_end; + src_nb = before->workstation_nb; + dst_nb = count; + before->workstation_list = (SD_workstation_t*) xbt_realloc( + before->workstation_list, + (before->workstation_nb+count)*sizeof(s_SD_workstation_t)); + for(i=0; iworkstation_list[before->workstation_nb+i] = + task->workstation_list[i]; + + before->workstation_nb += count; + xbt_free(before->flops_amount); + xbt_free(before->bytes_amount); + before->flops_amount = xbt_new0(double, + before->workstation_nb); + before->bytes_amount = xbt_new0(double, + before->workstation_nb* + before->workstation_nb); + + for(i=0;iamount/src_nb; + src_end = src_start + before->amount/src_nb; + for(j=0; jamount/dst_nb; + dst_end = dst_start + before->amount/dst_nb; + XBT_VERB("(%s->%s): (%.2f, %.2f)-> (%.2f, %.2f)", + SD_workstation_get_name(before->workstation_list[i]), + SD_workstation_get_name(before->workstation_list[src_nb+j]), + src_start, src_end, dst_start, dst_end); + if ((src_end <= dst_start) || (dst_end <= src_start)) { + before->bytes_amount[i*(src_nb+dst_nb)+src_nb+j]=0.0; + } else { + before->bytes_amount[i*(src_nb+dst_nb)+src_nb+j] = + MIN(src_end, dst_end) - MAX(src_start, dst_start); + } + XBT_VERB("==> %.2f", + before->bytes_amount[i*(src_nb+dst_nb)+src_nb+j]); + } + } + + if (__SD_task_is_schedulable(before) || + __SD_task_is_not_scheduled(before)) { + SD_task_do_schedule(before); + XBT_VERB + ("Auto-Schedule redistribution task %s. Send %.f bytes from %d hosts to %d hosts.", + SD_task_get_name(before),before->amount, src_nb, dst_nb); + } + } + } + } + xbt_dynar_foreach(task->tasks_after, cpt, dep) { + SD_task_t after = dep->dst; + if (after->kind == SD_TASK_COMM_PAR_MXN_1D_BLOCK){ + if (!after->workstation_list){ + XBT_VERB("Receiver side of Task '%s' is not scheduled yet", + SD_task_get_name(after)); + after->workstation_list = xbt_new0(SD_workstation_t, count); + after->workstation_nb = count; + XBT_VERB("Fill the workstation list with list of Task '%s'", + SD_task_get_name(task)); + for (i=0;iworkstation_list[i] = task->workstation_list[i]; + } else { + int src_nb, dst_nb; + double src_start, src_end, dst_start, dst_end; + src_nb = count; + dst_nb = after->workstation_nb; + after->workstation_list = (SD_workstation_t*) xbt_realloc( + after->workstation_list, + (after->workstation_nb+count)*sizeof(s_SD_workstation_t)); + for(i=after->workstation_nb - 1; i>=0; i--) + after->workstation_list[count+i] = after->workstation_list[i]; + for(i=0; iworkstation_list[i] = task->workstation_list[i]; + + after->workstation_nb += count; + + xbt_free(after->flops_amount); + xbt_free(after->bytes_amount); + + after->flops_amount = xbt_new0(double, after->workstation_nb); + after->bytes_amount = xbt_new0(double, + after->workstation_nb* + after->workstation_nb); + + for(i=0;iamount/src_nb; + src_end = src_start + after->amount/src_nb; + for(j=0; jamount/dst_nb; + dst_end = dst_start + after->amount/dst_nb; + XBT_VERB("(%d->%d): (%.2f, %.2f)-> (%.2f, %.2f)", + i, j, src_start, src_end, dst_start, dst_end); + if ((src_end <= dst_start) || (dst_end <= src_start)) { + after->bytes_amount[i*(src_nb+dst_nb)+src_nb+j]=0.0; + } else { + after->bytes_amount[i*(src_nb+dst_nb)+src_nb+j] = + MIN(src_end, dst_end)- MAX(src_start, dst_start); + } + XBT_VERB("==> %.2f", + after->bytes_amount[i*(src_nb+dst_nb)+src_nb+j]); + } + } + + if (__SD_task_is_schedulable(after) || + __SD_task_is_not_scheduled(after)) { + SD_task_do_schedule(after); + XBT_VERB + ("Auto-Schedule redistribution task %s. Send %.f bytes from %d hosts to %d hosts.", + SD_task_get_name(after),after->amount, src_nb, dst_nb); + } + } + } + } + } } /** @brief autoschedule a task on a list of workstations