X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/0c4fb41478feb409368d73f4174a9d357d5a4906..b1919c1826467eb2cfbc972842e0c39c5ab4d92c:/src/simdag/sd_task.c diff --git a/src/simdag/sd_task.c b/src/simdag/sd_task.c index e62ad0f2d5..3673944cd3 100644 --- a/src/simdag/sd_task.c +++ b/src/simdag/sd_task.c @@ -174,13 +174,40 @@ SD_task_t SD_task_create_comp_par_amdahl(const char *name, void *data, { xbt_assert(alpha < 1. && alpha >= 0., "Invalid parameter: alpha must be in [0.;1.["); - + SD_task_t res = SD_task_create(name, data, amount); res->alpha = alpha; res->kind = SD_TASK_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; + return res; +} /** * \brief Destroys a task. @@ -212,13 +239,13 @@ void SD_task_destroy(SD_task_t task) xbt_free(task->communication_amount); xbt_free(task->computation_amount); - xbt_mallocator_release(sd_global->task_mallocator,task); - sd_global->task_number--; - #ifdef HAVE_TRACING if (task->category) xbt_free(task->category); #endif + xbt_mallocator_release(sd_global->task_mallocator,task); + sd_global->task_number--; + XBT_DEBUG("Task destroyed."); } @@ -252,14 +279,22 @@ void SD_task_set_data(SD_task_t task, void *data) /** * \brief Sets the rate of a task * - * This will change the percentage of the available power or network bandwidth - * a task can use. + * This will change the network bandwidth a task can use. 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. * - * \param task a task - * \param rate the new rate you want to associate with this task + * 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."); + task->rate = rate; } @@ -474,7 +509,7 @@ void SD_task_dump(SD_task_t task) } } XBT_INFO(" - amount: %.0f", SD_task_get_amount(task)); - XBT_INFO(" - Dependencies to satisfy: %u", task->unsatisfied_dependencies); + XBT_INFO(" - Dependencies to satisfy: %d", task->unsatisfied_dependencies); if (!xbt_dynar_is_empty(task->tasks_before)) { XBT_INFO(" - pre-dependencies:"); xbt_dynar_foreach(task->tasks_before, counter, dependency) { @@ -552,7 +587,7 @@ void SD_task_dependency_add(const char *name, void *data, SD_task_t 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" - " or SD_RUNNING", + " or SD_RUNNING", SD_task_get_name(src)); if (!__SD_task_is_not_scheduled(dst) && !__SD_task_is_schedulable(dst) @@ -597,9 +632,6 @@ 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); } - - /* __SD_print_dependencies(src); - __SD_print_dependencies(dst); */ } /** @@ -938,8 +970,12 @@ 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); + 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); @@ -953,7 +989,8 @@ 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) { @@ -968,9 +1005,9 @@ static void __SD_task_destroy_scheduling_data(SD_task_t task) task->computation_amount = task->communication_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) { @@ -985,12 +1022,11 @@ void __SD_task_really_run(SD_task_t task) "Task '%s': workstation_list is NULL!", SD_task_get_name(task)); - - XBT_DEBUG("Really running task '%s'", SD_task_get_name(task)); + int workstation_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 < workstation_nb; i++) { if (SD_workstation_get_access_mode(task->workstation_list[i]) == SD_WORKSTATION_SEQUENTIAL_ACCESS) { task->workstation_list[i]->current_task = task; @@ -1004,71 +1040,31 @@ 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); + /* we have to create a Surf workstation array instead of the SimDag + * workstation array */ + surf_workstations = xbt_new(void *, workstation_nb); - for (i = 0; i < task->workstation_nb; i++) + for (i = 0; i < workstation_nb; i++) surf_workstations[i] = task->workstation_list[i]->surf_workstation; - /* 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 *computation_amount = xbt_new0(double, workstation_nb); + double *communication_amount = xbt_new0(double, workstation_nb * workstation_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); + if(task->computation_amount) memcpy(computation_amount, task->computation_amount, sizeof(double) * - task->workstation_nb); + workstation_nb); + if(task->communication_amount) memcpy(communication_amount, task->communication_amount, - sizeof(double) * task->workstation_nb * task->workstation_nb); + sizeof(double) * workstation_nb * workstation_nb); - task->surf_action = + task->surf_action = surf_workstation_model->extension. - workstation.execute_parallel_task(task->workstation_nb, + workstation.execute_parallel_task(workstation_nb, surf_workstations, computation_amount, communication_amount, task->rate); - } else { - xbt_free(surf_workstations); - } surf_workstation_model->action_data_set(task->surf_action, task); @@ -1086,10 +1082,11 @@ void __SD_task_really_run(SD_task_t 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) @@ -1111,11 +1108,11 @@ 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'", + 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); @@ -1134,7 +1131,7 @@ 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) { @@ -1164,14 +1161,14 @@ void __SD_task_just_done(SD_task_t task) 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), (int)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!", + "Workstation '%s' has sequential access but no FIFO!", SD_workstation_get_name(workstation)); xbt_assert(workstation->current_task = task, "Workstation '%s': current task should be '%s'", @@ -1181,7 +1178,7 @@ void __SD_task_just_done(SD_task_t task) /* the task is over so we can release the workstation */ 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)); @@ -1199,9 +1196,10 @@ 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) { @@ -1236,7 +1234,7 @@ void __SD_task_just_done(SD_task_t task) 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 */ + or if I am the first task in the FIFO */ can_start = workstation->access_mode == SD_WORKSTATION_SHARED_ACCESS || candidate == xbt_fifo_get_item_content(xbt_fifo_get_first_item @@ -1251,14 +1249,14 @@ 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 */ + /* 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'", + 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 */ @@ -1284,7 +1282,9 @@ 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) { @@ -1356,46 +1356,56 @@ void SD_task_distribute_comp_amdhal(SD_task_t task, int ws_count) for(i=0;icomputation_amount[i] = - (task->alpha + (1 - task->alpha)/ws_count) * task->amount; + (task->alpha + (1 - task->alpha)/ws_count) * task->amount; } } + /** @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_COMM_E2E: case SD_TASK_COMP_PAR_AMDAHL: - xbt_assert(task->computation_amount, "SD_task_distribute_comp_amdhal should be called first."); + SD_task_distribute_comp_amdhal(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->computation_amount){ + /*This task has failed and is rescheduled. Reset the computation amount*/ + task->computation_amount = xbt_new0(double, 1); + task->computation_amount[0] = task->remains; + } SD_task_do_schedule(task); break; default: @@ -1410,14 +1420,9 @@ void SD_task_schedulev(SD_task_t task, int count, task->communication_amount[2]); } - if (task->kind == SD_TASK_COMP_PAR_AMDAHL) { - XBT_VERB("Schedule computation task %s on %d hosts. It costs %.f flops on each host", - SD_task_get_name(task), - task->workstation_nb, - task->computation_amount[0]); - } - - /* 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), @@ -1446,7 +1451,6 @@ 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))) { @@ -1462,6 +1466,134 @@ void SD_task_schedulev(SD_task_t task, int count, } } } + /* 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'Law", + SD_task_get_name(task), task->workstation_nb, + task->computation_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. Fill the workstation list with receiver side", + SD_task_get_name(before)); + before->workstation_list = xbt_new0(SD_workstation_t, count); + before->workstation_nb = count; + 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 = 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; + + before->computation_amount = xbt_new0(double, + before->workstation_nb); + before->communication_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->communication_amount[i*(src_nb+dst_nb)+src_nb+j]=0.0; + } else { + before->communication_amount[i*(src_nb+dst_nb)+src_nb+j] = + MIN(src_end, dst_end) - MAX(src_start, dst_start); + } + XBT_VERB("==> %.2f", + before->communication_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. Fill the workstation list with sender side", + SD_task_get_name(after)); + after->workstation_list = xbt_new0(SD_workstation_t, count); + after->workstation_nb = count; + 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; + + after->computation_amount = xbt_new0(double, after->workstation_nb); + after->communication_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->communication_amount[i*(src_nb+dst_nb)+src_nb+j]=0.0; + } else { + after->communication_amount[i*(src_nb+dst_nb)+src_nb+j] = + MIN(src_end, dst_end)- MAX(src_start, dst_start); + } + XBT_VERB("==> %.2f", + after->communication_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