Add modular-heft-prio scheduler

src/Makefile.am

@@ -293,6 +293,7 @@ libstarpu_@STARPU_EFFECTIVE_VERSION@_la_SOURCES = 		\
 	sched_policies/modular_random.c				\
 	sched_policies/modular_random_prefetching.c			\
 	sched_policies/modular_heft.c				\
+	sched_policies/modular_heft_prio.c			\
 	sched_policies/modular_heft2.c				\
 	sched_policies/modular_ws.c
 

src/core/sched_policy.c

@@ -1,6 +1,6 @@
 /* StarPU --- Runtime system for heterogeneous multicore architectures.
  *
- * Copyright (C) 2010-2016  Université de Bordeaux
+ * Copyright (C) 2010-2017  Université de Bordeaux
  * Copyright (C) 2010-2017  CNRS
  * Copyright (C) 2011, 2016  INRIA
  * Copyright (C) 2016  Uppsala University
@@ -59,6 +59,7 @@ static struct starpu_sched_policy *predefined_policies[] =
 	&_starpu_sched_modular_random_prio_prefetching_policy,
 	&_starpu_sched_modular_ws_policy,
 	&_starpu_sched_modular_heft_policy,
+	&_starpu_sched_modular_heft_prio_policy,
 	&_starpu_sched_modular_heft2_policy,
 	&_starpu_sched_eager_policy,
 	&_starpu_sched_prio_policy,

src/core/sched_policy.h

@@ -1,6 +1,6 @@
 /* StarPU --- Runtime system for heterogeneous multicore architectures.
  *
- * Copyright (C) 2010, 2012-2013, 2015-2016  Université de Bordeaux
+ * Copyright (C) 2010, 2012-2013, 2015-2017  Université de Bordeaux
  * Copyright (C) 2011  INRIA
  *
  * StarPU is free software; you can redistribute it and/or modify
@@ -89,6 +89,7 @@ extern struct starpu_sched_policy _starpu_sched_modular_random_prefetching_polic
 extern struct starpu_sched_policy _starpu_sched_modular_random_prio_prefetching_policy;
 extern struct starpu_sched_policy _starpu_sched_modular_ws_policy;
 extern struct starpu_sched_policy _starpu_sched_modular_heft_policy;
+extern struct starpu_sched_policy _starpu_sched_modular_heft_prio_policy;
 extern struct starpu_sched_policy _starpu_sched_modular_heft2_policy;
 extern struct starpu_sched_policy _starpu_sched_graph_test_policy;
 

src/sched_policies/component_sched.c

@@ -224,6 +224,9 @@ void set_properties(struct starpu_sched_component * component)
 	int worker = starpu_bitmap_first(component->workers_in_ctx);
 	if (worker == -1)
 		return;
+#ifdef STARPU_DEVEL
+#warning FIXME: Not all CUDA devices have the same speed
+#endif
 	uint32_t first_worker = _starpu_get_worker_struct(worker)->worker_mask;
 	unsigned first_memory_node = _starpu_get_worker_struct(worker)->memory_node;
 	int is_homogeneous = 1;

src/sched_policies/modular_heft.c

@@ -1,6 +1,6 @@
 /* StarPU --- Runtime system for heterogeneous multicore architectures.
  *
- * Copyright (C) 2013-2015  Université de Bordeaux
+ * Copyright (C) 2013-2015, 2017  Université de Bordeaux
  * Copyright (C) 2013  INRIA
  * Copyright (C) 2013  Simon Archipoff
  *
@@ -44,7 +44,7 @@ static void initialize_heft_center_policy(unsigned sched_ctx_id)
  *                                    |
  *                              window_component
  *                                    |
- * perfmodel_component <--push-- perfmodel_select_component --push--> eager_component
+ * mct_component <--push-- perfmodel_select_component --push--> eager_component
  *          |                                                    |
  *          |                                                    |
  *          >----------------------------------------------------<

src/sched_policies/modular_heft2.c

@@ -1,6 +1,6 @@
 /* StarPU --- Runtime system for heterogeneous multicore architectures.
  *
- * Copyright (C) 2013-2015  Université de Bordeaux
+ * Copyright (C) 2013-2015, 2017  Université de Bordeaux
  * Copyright (C) 2013  INRIA
  * Copyright (C) 2013  Simon Archipoff
  *
@@ -39,6 +39,32 @@ static void initialize_heft2_center_policy(unsigned sched_ctx_id)
 	if (starpu_sched_ctx_max_priority_is_set(sched_ctx_id) == 0)
 		starpu_sched_ctx_set_max_priority(sched_ctx_id, INT_MAX);
 
+/* The scheduling strategy look like this :
+ *
+ *                                    |
+ *                              window_component
+ *                                    |
+ * heft_component <--push-- perfmodel_select_component --push--> eager_component
+ *          |                                                    |
+ *          |                                                    |
+ *          >----------------------------------------------------<
+ *                    |                                |
+ *              best_impl_component                    best_impl_component
+ *                    |                                |
+ *                prio_component                        prio_component
+ *                    |                                |
+ *               worker_component                   worker_component
+ *
+ * A window contain the tasks that failed to be pushed, so as when the prio_components reclaim
+ * tasks by calling can_push to their parent (classically, just after a successful pop have
+ * been made by its associated worker_component), this call goes up to the window_component which
+ * pops a task from its local queue and try to schedule it by pushing it to the
+ * decision_component. 
+ * Finally, the task will be pushed to the prio_component which is the direct
+ * parent in the tree of the worker_component the task has been scheduled on. This
+ * component will push the task on its local queue if no one of the two thresholds
+ * have been reached for it, or send a push_error signal to its parent.
+ */
 	struct starpu_sched_tree * t = starpu_sched_tree_create(sched_ctx_id);
 
 	struct starpu_sched_component * window_component = starpu_sched_component_prio_create(t, NULL);

src/sched_policies/modular_heft_prio.c

@@ -0,0 +1,154 @@
+/* StarPU --- Runtime system for heterogeneous multicore architectures.
+ *
+ * Copyright (C) 2013-2015, 2017  Université de Bordeaux
+ * Copyright (C) 2013  INRIA
+ * Copyright (C) 2013  Simon Archipoff
+ *
+ * StarPU is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU Lesser General Public License as published by
+ * the Free Software Foundation; either version 2.1 of the License, or (at
+ * your option) any later version.
+ *
+ * StarPU is distributed in the hope that it will be useful, but
+ * WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
+ *
+ * See the GNU Lesser General Public License in COPYING.LGPL for more details.
+ */
+
+#include <starpu_sched_component.h>
+#include <starpu_scheduler.h>
+#include <float.h>
+#include <limits.h>
+
+/* The two thresolds concerns the prio components, which contains queues
+ * who can handle the priority of StarPU tasks. You can tune your
+ * scheduling by benching those values and choose which one is the
+ * best for your current application. 
+ * The current value of the ntasks_threshold is the best we found
+ * so far across several types of applications (cholesky, LU, stencil).
+ */
+#define _STARPU_SCHED_NTASKS_THRESHOLD_DEFAULT 30
+#define _STARPU_SCHED_EXP_LEN_THRESHOLD_DEFAULT 1000000000.0
+
+static void initialize_heft_prio_policy(unsigned sched_ctx_id)
+{
+	/* The application may use any integer */
+	if (starpu_sched_ctx_min_priority_is_set(sched_ctx_id) == 0)
+		starpu_sched_ctx_set_min_priority(sched_ctx_id, INT_MIN);
+	if (starpu_sched_ctx_max_priority_is_set(sched_ctx_id) == 0)
+		starpu_sched_ctx_set_max_priority(sched_ctx_id, INT_MAX);
+
+/* The scheduling strategy look like this :
+ *
+ *                                    |
+ *                              window_component
+ *                                    |
+ *  mct_component <--push-- perfmodel_select_component --push--> eager_component
+ *  |     |     |                                                  |
+ * eager eager eager                                               |
+ *  |     |     |                                                  |
+ *  >--------------------------------------------------------------<
+ *                    |                                |
+ *              best_impl_component              best_impl_component
+ *                    |                                |
+ *                prio_component                  prio_component
+ *                    |                               |
+ *               worker_component                   worker_component
+ *
+ * A window contain the tasks that failed to be pushed, so as when the prio_components reclaim
+ * tasks by calling can_push to their parent (classically, just after a successful pop have
+ * been made by its associated worker_component), this call goes up to the window_component which
+ * pops a task from its local queue and try to schedule it by pushing it to the
+ * decision_component. 
+ * Finally, the task will be pushed to the prio_component which is the direct
+ * parent in the tree of the worker_component the task has been scheduled on. This
+ * component will push the task on its local queue if no one of the two thresholds
+ * have been reached for it, or send a push_error signal to its parent.
+ */
+	struct starpu_sched_tree * t = starpu_sched_tree_create(sched_ctx_id);
+
+	struct starpu_sched_component * window_component = starpu_sched_component_prio_create(t, NULL);
+
+	struct starpu_sched_component * perfmodel_component = starpu_sched_component_mct_create(t, NULL);
+	struct starpu_sched_component * no_perfmodel_component = starpu_sched_component_eager_create(t, NULL);
+	struct starpu_sched_component * calibrator_component = starpu_sched_component_eager_calibration_create(t, NULL);
+	
+	struct starpu_sched_component_perfmodel_select_data perfmodel_select_data =
+		{
+			.calibrator_component = calibrator_component,
+			.no_perfmodel_component = no_perfmodel_component,
+			.perfmodel_component = perfmodel_component,
+		};
+
+	struct starpu_sched_component * perfmodel_select_component = starpu_sched_component_perfmodel_select_create(t, &perfmodel_select_data);
+
+	t->root = window_component;
+	starpu_sched_component_connect(window_component, perfmodel_select_component);
+
+	starpu_sched_component_connect(perfmodel_select_component, perfmodel_component);
+	starpu_sched_component_connect(perfmodel_select_component, calibrator_component);
+	starpu_sched_component_connect(perfmodel_select_component, no_perfmodel_component);
+
+	struct starpu_sched_component_prio_data prio_data =
+		{
+			.ntasks_threshold = starpu_get_env_number_default("STARPU_NTASKS_THRESHOLD", _STARPU_SCHED_NTASKS_THRESHOLD_DEFAULT),
+			.exp_len_threshold = starpu_get_env_float_default("STARPU_EXP_LEN_THRESHOLD", _STARPU_SCHED_EXP_LEN_THRESHOLD_DEFAULT),
+		};
+
+	unsigned i, j;
+
+	struct starpu_sched_component * eagers[starpu_memory_nodes_get_count()];
+
+	/* Create one eager component per memory node, below mct */
+	for(i = 0; i < starpu_memory_nodes_get_count(); i++)
+	{
+		for(j = 0; j < starpu_worker_get_count() + starpu_combined_worker_get_count(); j++)
+			if (starpu_worker_get_memory_node(j) == i)
+				break;
+		if (j == starpu_worker_get_count() + starpu_combined_worker_get_count())
+			/* Don't create a component for this memory node with no worker */
+			continue;
+		eagers[i] = starpu_sched_component_eager_create(t, NULL);
+		starpu_sched_component_connect(perfmodel_component, eagers[i]);
+	}
+
+	for(i = 0; i < starpu_worker_get_count() + starpu_combined_worker_get_count(); i++)
+	{
+		struct starpu_sched_component * worker_component = starpu_sched_component_worker_get(sched_ctx_id, i);
+		struct starpu_sched_component * prio_component = starpu_sched_component_prio_create(t, &prio_data);
+		struct starpu_sched_component * impl_component = starpu_sched_component_best_implementation_create(t, NULL);
+
+		starpu_sched_component_connect(prio_component, worker_component);
+		starpu_sched_component_connect(impl_component, prio_component);
+
+		starpu_sched_component_connect(eagers[starpu_worker_get_memory_node(i)], impl_component);
+		starpu_sched_component_connect(no_perfmodel_component, impl_component);
+		starpu_sched_component_connect(calibrator_component, impl_component);
+	}
+
+	starpu_sched_tree_update_workers(t);
+	starpu_sched_ctx_set_policy_data(sched_ctx_id, (void*)t);
+}
+
+static void deinitialize_heft_prio_policy(unsigned sched_ctx_id)
+{
+	struct starpu_sched_tree *t = (struct starpu_sched_tree*)starpu_sched_ctx_get_policy_data(sched_ctx_id);
+	starpu_sched_tree_destroy(t);
+}
+
+struct starpu_sched_policy _starpu_sched_modular_heft_prio_policy =
+{
+	.init_sched = initialize_heft_prio_policy,
+	.deinit_sched = deinitialize_heft_prio_policy,
+	.add_workers = starpu_sched_tree_add_workers,
+	.remove_workers = starpu_sched_tree_remove_workers,
+	.push_task = starpu_sched_tree_push_task,
+	.pop_task = starpu_sched_tree_pop_task,
+	.pre_exec_hook = starpu_sched_component_worker_pre_exec_hook,
+	.post_exec_hook = starpu_sched_component_worker_post_exec_hook,
+	.pop_every_task = NULL,
+	.policy_name = "modular-heft-prio",
+	.policy_description = "heft+prio modular policy",
+	.worker_type = STARPU_WORKER_LIST,
+};

tests/microbenchs/parallel_dependent_homogeneous_tasks_data.sh

@@ -2,7 +2,7 @@
 #
 # StarPU --- Runtime system for heterogeneous multicore architectures.
 #
-# Copyright (C) 2016  Université de Bordeaux
+# Copyright (C) 2016-2017  Université de Bordeaux
 # Copyright (C) 2016  CNRS
 #
 # StarPU is free software; you can redistribute it and/or modify
@@ -18,6 +18,6 @@
 
 source $(dirname $0)/microbench.sh
 
-XSUCCESS="modular-heft modular-heft2 lws dmda dmdar dmdas dmdasd pheft"
+XSUCCESS="modular-heft modular-heft-prio modular-heft2 lws dmda dmdar dmdas dmdasd pheft"
 
 test_scheds parallel_dependent_homogeneous_tasks_data

tests/microbenchs/parallel_independent_heterogeneous_tasks.sh

@@ -2,7 +2,7 @@
 #
 # StarPU --- Runtime system for heterogeneous multicore architectures.
 #
-# Copyright (C) 2016  Université de Bordeaux
+# Copyright (C) 2016-2017  Université de Bordeaux
 # Copyright (C) 2016  CNRS
 #
 # StarPU is free software; you can redistribute it and/or modify
@@ -18,6 +18,6 @@
 
 source $(dirname $0)/microbench.sh
 
-XFAIL="modular-eager-prefetching modular-prio-prefetching modular-random-prefetching modular-random-prio-prefetching modular-ws modular-heft modular-heft2 random peager heteroprio graph_test"
+XFAIL="modular-eager-prefetching modular-prio-prefetching modular-random-prefetching modular-random-prio-prefetching modular-ws modular-heft modular-heft-prio modular-heft2 random peager heteroprio graph_test"
 
 test_scheds parallel_independent_heterogeneous_tasks

tests/microbenchs/parallel_independent_heterogeneous_tasks_data.sh

@@ -2,7 +2,7 @@
 #
 # StarPU --- Runtime system for heterogeneous multicore architectures.
 #
-# Copyright (C) 2016  Université de Bordeaux
+# Copyright (C) 2016-2017  Université de Bordeaux
 # Copyright (C) 2016  CNRS
 #
 # StarPU is free software; you can redistribute it and/or modify
@@ -18,6 +18,6 @@
 
 source $(dirname $0)/microbench.sh
 
-XSUCCESS="modular-heft modular-heft2 dmda dmdar dmdas dmdasd pheft"
+XSUCCESS="modular-heft modular-heft-prio modular-heft2 dmda dmdar dmdas dmdasd pheft"
 
 test_scheds parallel_independent_heterogeneous_tasks_data

tests/microbenchs/parallel_independent_homogeneous_tasks.sh

@@ -2,7 +2,7 @@
 #
 # StarPU --- Runtime system for heterogeneous multicore architectures.
 #
-# Copyright (C) 2016  Université de Bordeaux
+# Copyright (C) 2016-2017  Université de Bordeaux
 # Copyright (C) 2016  CNRS
 #
 # StarPU is free software; you can redistribute it and/or modify
@@ -18,6 +18,6 @@
 
 source $(dirname $0)/microbench.sh
 
-XFAIL="modular-eager-prefetching modular-prio-prefetching modular-random-prefetching modular-random-prio-prefetching modular-ws modular-heft modular-heft2 random peager heteroprio graph_test"
+XFAIL="modular-eager-prefetching modular-prio-prefetching modular-random-prefetching modular-random-prio-prefetching modular-ws modular-heft modular-heft-prio modular-heft2 random peager heteroprio graph_test"
 
 test_scheds parallel_independent_homogeneous_tasks

tests/microbenchs/parallel_independent_homogeneous_tasks_data.sh

@@ -3,7 +3,7 @@
 # StarPU --- Runtime system for heterogeneous multicore architectures.
 #
 # Copyright (C) 2016  CNRS
-# Copyright (C) 2016  Université de Bordeaux
+# Copyright (C) 2016-2017  Université de Bordeaux
 #
 # StarPU is free software; you can redistribute it and/or modify
 # it under the terms of the GNU Lesser General Public License as published by
@@ -18,6 +18,6 @@
 
 source $(dirname $0)/microbench.sh
 
-XSUCCESS="modular-heft modular-heft2 dmda dmdar dmdas dmdasd pheft"
+XSUCCESS="modular-heft modular-heft-prio modular-heft2 dmda dmdar dmdas dmdasd pheft"
 
 test_scheds parallel_independent_homogeneous_tasks_data