starpu-max/src/sched_policies/component_prio.c

/* StarPU --- Runtime system for heterogeneous multicore architectures.
 *
 * Copyright (C) 2013,2017                                Inria
 * Copyright (C) 2014-2017                                CNRS
 * Copyright (C) 2014-2015,2017-2018                           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
 * 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 <common/fxt.h>
#include <core/workers.h>

#include "prio_deque.h"

#ifdef STARPU_USE_FXT
#define STARPU_TRACE_SCHED_COMPONENT_PUSH_PRIO(component,ntasks,exp_len) do {                                 \
	int workerid = STARPU_NMAXWORKERS + 1;									\
	if((component->nchildren == 1) && starpu_sched_component_is_worker(component->children[0])) \
		workerid = starpu_sched_component_worker_get_workerid(component->children[0]); \
	_STARPU_TRACE_SCHED_COMPONENT_PUSH_PRIO(workerid, ntasks, exp_len); \
} while (0)

#define STARPU_TRACE_SCHED_COMPONENT_POP_PRIO(component,ntasks,exp_len) do {                                 \
	int workerid = STARPU_NMAXWORKERS + 1;									\
	if((component->nchildren == 1) && starpu_sched_component_is_worker(component->children[0])) \
		workerid = starpu_sched_component_worker_get_workerid(component->children[0]); \
	_STARPU_TRACE_SCHED_COMPONENT_POP_PRIO(workerid, ntasks, exp_len); \
} while (0)
#else
#define STARPU_TRACE_SCHED_COMPONENT_PUSH_PRIO(component,ntasks,exp_len) do { } while (0)
#define STARPU_TRACE_SCHED_COMPONENT_POP_PRIO(component,ntasks,exp_len) do { } while (0)
#endif

struct _starpu_prio_data
{
	struct _starpu_prio_deque prio;
	starpu_pthread_mutex_t mutex;
	unsigned ntasks_threshold;
	double exp_len_threshold;
};

static void prio_component_deinit_data(struct starpu_sched_component * component)
{
	STARPU_ASSERT(component && component->data);
	struct _starpu_prio_data * f = component->data;
	_starpu_prio_deque_destroy(&f->prio);
	STARPU_PTHREAD_MUTEX_DESTROY(&f->mutex);
	free(f);
}

static double prio_estimated_end(struct starpu_sched_component * component)
{
	STARPU_ASSERT(component && component->data);
	struct _starpu_prio_data * data = component->data;
	struct _starpu_prio_deque * prio = &data->prio;
	return starpu_sched_component_estimated_end_min_add(component, prio->exp_len);
}

static double prio_estimated_load(struct starpu_sched_component * component)
{
	STARPU_ASSERT(component && component->data);
	STARPU_ASSERT(starpu_bitmap_cardinal(component->workers_in_ctx) != 0);
	struct _starpu_prio_data * data = component->data;
	struct _starpu_prio_deque * prio = &data->prio;
	starpu_pthread_mutex_t * mutex = &data->mutex;
	double relative_speedup = 0.0;
	double load = starpu_sched_component_estimated_load(component);
	if(STARPU_SCHED_COMPONENT_IS_HOMOGENEOUS(component))
	{
		int first_worker = starpu_bitmap_first(component->workers_in_ctx);
		relative_speedup = starpu_worker_get_relative_speedup(starpu_worker_get_perf_archtype(first_worker, component->tree->sched_ctx_id));
		STARPU_COMPONENT_MUTEX_LOCK(mutex);
		load += prio->ntasks / relative_speedup;
		STARPU_COMPONENT_MUTEX_UNLOCK(mutex);
		return load;
	}
	else
	{
		int i;
		for(i = starpu_bitmap_first(component->workers_in_ctx);
		    i != -1;
		    i = starpu_bitmap_next(component->workers_in_ctx, i))
			relative_speedup += starpu_worker_get_relative_speedup(starpu_worker_get_perf_archtype(i, component->tree->sched_ctx_id));
		relative_speedup /= starpu_bitmap_cardinal(component->workers_in_ctx);
		STARPU_ASSERT(!_STARPU_IS_ZERO(relative_speedup));
		STARPU_COMPONENT_MUTEX_LOCK(mutex);
		load += prio->ntasks / relative_speedup;
		STARPU_COMPONENT_MUTEX_UNLOCK(mutex);
	}
	return load;
}

static int prio_push_local_task(struct starpu_sched_component * component, struct starpu_task * task, unsigned is_pushback)
{
	STARPU_ASSERT(component && component->data && task);
	STARPU_ASSERT(starpu_sched_component_can_execute_task(component,task));
	struct _starpu_prio_data * data = component->data;
	struct _starpu_prio_deque * prio = &data->prio;
	starpu_pthread_mutex_t * mutex = &data->mutex;
	int ret;
	const double now = starpu_timing_now();
	STARPU_COMPONENT_MUTEX_LOCK(mutex);
	double exp_len;
	if(!isnan(task->predicted))
		exp_len = prio->exp_len + task->predicted;
	else
		exp_len = prio->exp_len;

	if((data->ntasks_threshold != 0) && (data->exp_len_threshold != 0.0) &&
			((prio->ntasks >= data->ntasks_threshold) || (exp_len >= data->exp_len_threshold)))
	{
		static int warned;
		if(task->predicted > data->exp_len_threshold && !warned)
		{
			_STARPU_DISP("Warning : a predicted task length (%lf) exceeds the expected length threshold (%lf) of a prio component queue, you should reconsider the value of this threshold. This message will not be printed again for further thresholds exceeding.\n",task->predicted,data->exp_len_threshold);
			warned = 1;
		}
		ret = 1;
		STARPU_COMPONENT_MUTEX_UNLOCK(mutex);
	}
	else
	{
		if(is_pushback)
			ret = _starpu_prio_deque_push_front_task(prio,task);
		else
		{
			ret = _starpu_prio_deque_push_back_task(prio,task);
			starpu_sched_component_prefetch_on_node(component, task);
			STARPU_TRACE_SCHED_COMPONENT_PUSH_PRIO(component, prio->ntasks, exp_len);
		}

		if(!isnan(task->predicted_transfer))
		{
			double end = prio_estimated_end(component);
			double tfer_end = now + task->predicted_transfer;
			if(tfer_end < end)
				task->predicted_transfer = 0.0;
			else
				task->predicted_transfer = tfer_end - end;
			exp_len += task->predicted_transfer;
		}

		if(!isnan(task->predicted))
		{
			prio->exp_len = exp_len;
			prio->exp_end = prio->exp_start + prio->exp_len;
		}
		STARPU_ASSERT(!isnan(prio->exp_end));
		STARPU_ASSERT(!isnan(prio->exp_len));
		STARPU_ASSERT(!isnan(prio->exp_start));

		STARPU_COMPONENT_MUTEX_UNLOCK(mutex);
		if(!is_pushback)
			component->can_pull(component);
	}

	return ret;
}

static int prio_push_task(struct starpu_sched_component * component, struct starpu_task * task)
{
	int ret = prio_push_local_task(component, task, 0);
	return ret;
}

static struct starpu_task * prio_pull_task(struct starpu_sched_component * component, struct starpu_sched_component * to STARPU_ATTRIBUTE_UNUSED)
{
	STARPU_ASSERT(component && component->data);
	struct _starpu_prio_data * data = component->data;
	struct _starpu_prio_deque * prio = &data->prio;
	starpu_pthread_mutex_t * mutex = &data->mutex;
	const double now = starpu_timing_now();
	STARPU_COMPONENT_MUTEX_LOCK(mutex);
	struct starpu_task * task = _starpu_prio_deque_pop_task(prio);
	if(task)
	{
		if(!isnan(task->predicted))
		{
			const double exp_len = prio->exp_len - task->predicted;
			prio->exp_start = now + task->predicted;
			if (exp_len >= 0.0)
			{
				prio->exp_len = exp_len;
			}
			else
			{
				/* exp_len can become negative due to rounding errors */
				prio->exp_len = 0.0;
			}
		}
		STARPU_ASSERT_MSG(prio->exp_len>=0, "prio->exp_len=%lf\n",prio->exp_len);
		if(!isnan(task->predicted_transfer))
		{
			if (prio->exp_len > task->predicted_transfer)
			{
				prio->exp_start += task->predicted_transfer;
				prio->exp_len -= task->predicted_transfer;
			}
			else
			{
				prio->exp_start += prio->exp_len;
				prio->exp_len = 0;
			}
		}

		prio->exp_end = prio->exp_start + prio->exp_len;
		if(prio->ntasks == 0)
			prio->exp_len = 0.0;

		STARPU_TRACE_SCHED_COMPONENT_POP_PRIO(component, prio->ntasks, prio->exp_len);
	}
	STARPU_ASSERT(!isnan(prio->exp_end));
	STARPU_ASSERT(!isnan(prio->exp_len));
	STARPU_ASSERT(!isnan(prio->exp_start));
	STARPU_COMPONENT_MUTEX_UNLOCK(mutex);

	// When a pop is called, a can_push is called for pushing tasks onto
	// the empty place of the queue left by the popped task.

	starpu_sched_component_send_can_push_to_parents(component);

	if(task)
		return task;

	return NULL;
}

/* When a can_push is caught by this function, we try to pop and push
 * tasks from our local queue as much as possible, until a
 * push fails, which means that the worker prio_components are
 * currently "full".
 */
static int prio_can_push(struct starpu_sched_component * component, struct starpu_sched_component * to STARPU_ATTRIBUTE_UNUSED)
{
	STARPU_ASSERT(component && starpu_sched_component_is_prio(component));
	int res = 0;
	struct starpu_task * task;

	task = starpu_sched_component_pump_downstream(component, &res);

	if(task)
	{
		int ret = prio_push_local_task(component,task,1);
		STARPU_ASSERT(!ret);
	}

	return res;
}

int starpu_sched_component_is_prio(struct starpu_sched_component * component)
{
	return component->push_task == prio_push_task;
}

struct starpu_sched_component * starpu_sched_component_prio_create(struct starpu_sched_tree *tree, struct starpu_sched_component_prio_data * params)
{
	struct starpu_sched_component * component = starpu_sched_component_create(tree, "prio");
	struct _starpu_prio_data *data;
	_STARPU_MALLOC(data, sizeof(*data));
	_starpu_prio_deque_init(&data->prio);
	STARPU_PTHREAD_MUTEX_INIT(&data->mutex,NULL);
	component->data = data;
	component->estimated_end = prio_estimated_end;
	component->estimated_load = prio_estimated_load;
	component->push_task = prio_push_task;
	component->pull_task = prio_pull_task;
	component->can_push = prio_can_push;
	component->deinit_data = prio_component_deinit_data;

	if(params)
	{
		data->ntasks_threshold=params->ntasks_threshold;
		data->exp_len_threshold=params->exp_len_threshold;
	}
	else
	{
		data->ntasks_threshold=0;
		data->exp_len_threshold=0.0;
	}

	return component;
}