ソースを参照

remove transfer time from exp_len at pop_time && the computation time at pre_exec & update exp_start accordingly

Andra Hugo 12 年 前
コミット
acdb9f0caa
共有3 個のファイルを変更した50 個の追加38 個の削除を含む
  1. 11 8
      sc_hypervisor/src/policies_utils/speed.c
  2. 12 0
      sc_hypervisor/src/sc_hypervisor.c
  3. 27 30
      src/sched_policies/deque_modeling_policy_data_aware.c

+ 11 - 8
sc_hypervisor/src/policies_utils/speed.c

@@ -64,7 +64,7 @@ double sc_hypervisor_get_velocity_per_worker(struct sc_hypervisor_wrapper *sc_w,
                 double curr_time = starpu_timing_now();
                 double elapsed_time = (curr_time - sc_w->start_time) / 1000000.0; /* in seconds */
 		elapsed_time -= sc_w->idle_time[worker];
-		sc_w->idle_time[worker] = 0.0;
+		
 
 /* 		size_t elapsed_data_used = sc_w->elapsed_data[worker]; */
 /*  		enum starpu_worker_archtype arch = starpu_worker_get_type(worker); */
@@ -84,7 +84,6 @@ double sc_hypervisor_get_velocity_per_worker(struct sc_hypervisor_wrapper *sc_w,
 /* 		} */
 			
                 double vel  = (elapsed_flops/elapsed_time);/* in Gflops/s */
-//		printf("%d in ctx %d: vel %lf\n", worker, sc_w->sched_ctx, vel);
 		sc_w->ref_velocity[worker] = sc_w->ref_velocity[worker] > 1.0 ? (sc_w->ref_velocity[worker] + vel) / 2 : vel; 
                 return vel;
         }
@@ -143,12 +142,9 @@ double sc_hypervisor_get_ref_velocity_per_worker_type(struct sc_hypervisor_wrapp
                 enum starpu_worker_archtype req_arch = starpu_worker_get_type(worker);
                 if(arch == req_arch)
                 {
-		
-			if(sc_w->ref_velocity[worker] > 1.0)
-			{
-				ref_velocity += sc_w->ref_velocity[worker];
-				nw++;
-			}
+			if(sc_w->ref_velocity[worker] < 1.0) return -1.0;
+			ref_velocity += sc_w->ref_velocity[worker];
+			nw++;
 		}
 	}
 	
@@ -159,10 +155,17 @@ double sc_hypervisor_get_velocity(struct sc_hypervisor_wrapper *sc_w, enum starp
 {
 
 	double velocity = sc_hypervisor_get_velocity_per_worker_type(sc_w, arch);
+	printf("arch %d vel %lf\n", arch, velocity);
 	if(velocity == -1.0)
+	{
 		velocity = sc_hypervisor_get_ref_velocity_per_worker_type(sc_w, arch);
+		printf("arch %d ref_vel %lf\n", arch, velocity);
+	}
 	if(velocity == -1.0)
+	{
 		velocity = arch == STARPU_CPU_WORKER ? 5.0 : 100.0;
+		printf("arch %d default_vel %lf\n", arch, velocity);
+	}
        
 	return velocity;
 }

+ 12 - 0
sc_hypervisor/src/sc_hypervisor.c

@@ -440,6 +440,16 @@ double sc_hypervisor_get_total_elapsed_flops_per_sched_ctx(struct sc_hypervisor_
 	return ret_val;
 }
 
+static void _reset_idle_time(unsigned sched_ctx)
+{
+	int i;
+	for(i = 0; i < STARPU_NMAXWORKERS; i++)
+	{
+		hypervisor.sched_ctx_w[sched_ctx].idle_time[i] = 0.0;
+	}
+	return;
+}
+
 void _reset_resize_sample_info(unsigned sender_sched_ctx, unsigned receiver_sched_ctx)
 {
 	/* info concerning only the gflops_rate strateg */
@@ -449,9 +459,11 @@ void _reset_resize_sample_info(unsigned sender_sched_ctx, unsigned receiver_sche
 	double start_time =  starpu_timing_now();
 	sender_sc_w->start_time = start_time;
 	_set_elapsed_flops_per_sched_ctx(sender_sched_ctx, 0.0);
+	_reset_idle_time(sender_sched_ctx);
 
 	receiver_sc_w->start_time = start_time;
 	_set_elapsed_flops_per_sched_ctx(receiver_sched_ctx, 0.0);
+	_reset_idle_time(receiver_sched_ctx);
 }
 
 /* actually move the workers: the cpus are moved, gpus are only shared  */

+ 27 - 30
src/sched_policies/deque_modeling_policy_data_aware.c

@@ -179,12 +179,11 @@ static struct starpu_task *dmda_pop_ready_task(unsigned sched_ctx_id)
 	task = _starpu_fifo_pop_first_ready_task(fifo, node);
 	if (task)
 	{
-		double model = task->predicted;
-
-		if(!isnan(model))
+		double transfer_model = task->predicted_transfer;
+		if(!isnan(transfer_model)) 
 		{
-			fifo->exp_len -= model;
-			fifo->exp_start = starpu_timing_now() + model;
+			fifo->exp_len -= transfer_model;
+			fifo->exp_start = starpu_timing_now() + transfer_model;
 			fifo->exp_end = fifo->exp_start + fifo->exp_len;
 		}
 
@@ -217,13 +216,17 @@ static struct starpu_task *dmda_pop_task(unsigned sched_ctx_id)
 	task = _starpu_fifo_pop_local_task(fifo);
 	if (task)
 	{
+		double transfer_model = task->predicted_transfer;
 		double model = task->predicted;
-		if(!isnan(model))
+
+		if(!isnan(transfer_model)) 
 		{
-			fifo->exp_len -= model;
-			fifo->exp_start = starpu_timing_now() + model;
+			fifo->exp_len -= transfer_model;
+			fifo->exp_start = starpu_timing_now() + transfer_model+model;
 			fifo->exp_end = fifo->exp_start + fifo->exp_len;
 		}
+
+
 		  
 #ifdef STARPU_VERBOSE
 		if (task->cl)
@@ -257,11 +260,11 @@ static struct starpu_task *dmda_pop_every_task(unsigned sched_ctx_id)
 	STARPU_PTHREAD_MUTEX_UNLOCK(sched_mutex);
 	while (new_list)
 	{
-		double model = new_list->predicted;
-		if(!isnan(model)) 
+		double transfer_model = new_list->predicted_transfer;
+		if(!isnan(transfer_model)) 
 		{
-			fifo->exp_len -= model;
-			fifo->exp_start = starpu_timing_now() + model;
+			fifo->exp_len -= transfer_model;
+			fifo->exp_start = starpu_timing_now() + transfer_model;
 			fifo->exp_end = fifo->exp_start + fifo->exp_len;
 		}
 
@@ -295,19 +298,19 @@ static int push_task_on_best_worker(struct starpu_task *task, int best_workerid,
 	fifo->exp_start = STARPU_MAX(fifo->exp_start, starpu_timing_now());
 	fifo->exp_end = fifo->exp_start + fifo->exp_len;
 
-	if (starpu_timing_now() + predicted_transfer < fifo->exp_end)
+	if ((starpu_timing_now() + predicted_transfer) < fifo->exp_end)
 	{
 		/* We may hope that the transfer will be finished by
 		 * the start of the task. */
-		predicted_transfer = 0;
+		predicted_transfer = 0.0;
 	}
 	else
 	{
 		/* The transfer will not be finished by then, take the
 		 * remainder into account */
-		predicted_transfer += starpu_timing_now();
-		predicted_transfer -= fifo->exp_end;
+		predicted_transfer = (starpu_timing_now() + predicted_transfer) - fifo->exp_end;
 	}
+
 	if(!isnan(predicted_transfer)) 
 	{
 		fifo->exp_end += predicted_transfer;
@@ -402,15 +405,7 @@ static int _dm_push_task(struct starpu_task *task, unsigned prio, unsigned sched
 		enum starpu_perfmodel_archtype perf_arch = starpu_worker_get_perf_archtype(worker);
 
 		/* Sometimes workers didn't take the tasks as early as we expected */
-		starpu_pthread_mutex_t *sched_mutex;
-		starpu_pthread_cond_t *sched_cond;
-		starpu_worker_get_sched_condition(worker, &sched_mutex, &sched_cond);
-
-		STARPU_PTHREAD_MUTEX_LOCK(sched_mutex);
-		fifo->exp_start = STARPU_MAX(fifo->exp_start, starpu_timing_now());
-		fifo->exp_end = fifo->exp_start + fifo->exp_len;
-		STARPU_PTHREAD_MUTEX_UNLOCK(sched_mutex);
-
+		double exp_start = STARPU_MAX(fifo->exp_start, starpu_timing_now());
 
 		for (nimpl = 0; nimpl < STARPU_MAXIMPLEMENTATIONS; nimpl++)
 		{
@@ -477,7 +472,7 @@ static int _dm_push_task(struct starpu_task *task, unsigned prio, unsigned sched
 			if (unknown)
 				continue;
 
-			exp_end = fifo->exp_start + fifo->exp_len + local_length;
+			exp_end = exp_start + fifo->exp_len + local_length;
 
 			if (best == -1 || exp_end < best_exp_end)
 			{
@@ -558,7 +553,7 @@ static void compute_all_performance_predictions(struct starpu_task *task,
 			}
 
 			STARPU_ASSERT_MSG(fifo != NULL, "worker %d ctx %d\n", worker, sched_ctx_id);
-			exp_end[worker_ctx][nimpl] = fifo->exp_start + fifo->exp_len;
+			exp_end[worker_ctx][nimpl] = exp_start + fifo->exp_len;
 			if (exp_end[worker_ctx][nimpl] > max_exp_end)
 				max_exp_end = exp_end[worker_ctx][nimpl];
 
@@ -580,7 +575,6 @@ static void compute_all_performance_predictions(struct starpu_task *task,
 				if (conversion_time > 0.0)
 					local_task_length[worker_ctx][nimpl] += conversion_time;
 			}
-			
 			double ntasks_end = fifo->ntasks / starpu_worker_get_relative_speedup(perf_arch);
 
 			/*
@@ -635,6 +629,9 @@ static void compute_all_performance_predictions(struct starpu_task *task,
 
 			exp_end[worker_ctx][nimpl] = exp_start + fifo->exp_len + local_task_length[worker_ctx][nimpl];
 
+
+			printf("%d/%d: exp_start %lf exp_len %lf exp_end %lf local_len %lf local_pen %lf \n", worker, sched_ctx_id, exp_start,  fifo->exp_len, exp_end[worker_ctx][nimpl], 
+			       local_task_length[worker_ctx][nimpl], local_data_penalty[worker_ctx][nimpl]);			
 			if (exp_end[worker_ctx][nimpl] < best_exp_end)
 			{
 				/* a better solution was found */
@@ -916,9 +913,9 @@ static void dmda_pre_exec_hook(struct starpu_task *task)
 	/* Once the task is executing, we can update the predicted amount
 	 * of work. */
 	STARPU_PTHREAD_MUTEX_LOCK(sched_mutex);
-	if(!isnan(transfer_model))
+	if(!isnan(model))
 	{
-		fifo->exp_len-= transfer_model;
+		fifo->exp_len-= model;
 		fifo->exp_start = starpu_timing_now() + model;
 		fifo->exp_end= fifo->exp_start + fifo->exp_len;
 	}