starpu-max/tools/starpu_replay.c

/* StarPU --- Runtime system for heterogeneous multicore architectures.
 *
 * Copyright (C) 2017                                     Erwan Leria
 * Copyright (C) 2018                                     Inria
 * Copyright (C) 2017,2018,2019                           CNRS
 * Copyright (C) 2016-2019                                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.
 */

/*
 * This reads a tasks.rec file and replays the recorded task graph.
 * Currently, this version is done to run with simgrid.
 *
 * For further information, contact erwan.leria@inria.fr
 */

#include <starpu.h>
#include <unistd.h>
#include <stdio.h>
#include <math.h>

#include <common/uthash.h>
#include <common/utils.h>
#include <starpu_scheduler.h>
#include <common/rbtree.h>


#define REPLAY_NMAX_DEPENDENCIES 8

#define ARRAY_DUP(in, out, n) memcpy(out, in, n * sizeof(*out))
#define ARRAY_INIT(array, n) memset(array, 0, n * sizeof(*array))

/* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
 * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
 * Declarations of global variables, structures, pointers, ... *
 * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
 * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * */

static int static_workerid;

/* TODO: move to core header while moving starpu_replay_sched to core */
extern void schedRecInit(const char * filename);
extern void applySchedRec(struct starpu_task * starpu_task, long submit_order);

/* Enum for normal and "wontuse" tasks */
enum task_type {NormalTask, WontUseTask};

typedef unsigned long jobid_t;

enum task_type control;
static char *name = NULL;
static char *model = NULL;
static jobid_t jobid;
static jobid_t *dependson;
static long submitorder = -1;
static starpu_tag_t tag;
static int workerid;
static uint32_t footprint;
static double flops, total_flops = 0.;

static double startTime; //start time (The instant when the task starts)
static double endTime; //end time (The instant when the task ends)

static int iteration = -1;

static starpu_data_handle_t handles[STARPU_NMAXBUFS];
static enum starpu_data_access_mode modes[STARPU_NMAXBUFS];
static char normal_reg_signal[STARPU_NMAXBUFS];

/* Use the following arrays when the number of data is greater than STARPU_NMAXBUFS */

starpu_data_handle_t * handles_ptr;
enum starpu_data_access_mode * modes_ptr;
size_t * sizes_set;

static size_t dependson_size;
static size_t ndependson;

static unsigned nb_parameters = 0; /* Number of parameters */
static int alloc_mode; /* If alloc_mode value is 1, then the handles are stored in dyn_handles, else they are in handles */

static int priority = 0;

char * reg_signal = NULL; /* The register signal (0 or 1 coded on 8 bit) is used to know which handle of the task has to be registered in StarPU (in fact to avoid  handle twice)*/

/* Record all tasks, hashed by jobid. */
static struct task
{
	struct starpu_rbtree_node node;
	UT_hash_handle hh;
	jobid_t jobid;
        int iteration;
	long submit_order;
	jobid_t *deps;
	size_t ndependson;
	struct starpu_task task;
	enum task_type type;
	int reg_signal;

} *tasks;

/* Record handles */
static struct handle
{
	UT_hash_handle hh;
	starpu_data_handle_t mem_ptr; /* This value should be the registered handle */
	starpu_data_handle_t handle; /* The key is the original value of the handle in the file */
} * handles_hash;

/* Record models */

static struct perfmodel
{
	UT_hash_handle hh;
	struct starpu_perfmodel perfmodel;
	char * model_name;
} * model_hash;



/*
 * Replay data interface
 * We don't care about many things anyway, essentially only sizes.
 */

struct replay_interface
{
	enum starpu_data_interface_id id;
	starpu_data_handle_t orig_handle;
	size_t size;
	size_t alloc_size;
	size_t max_size;
};

static struct starpu_data_interface_ops replay_interface_ops;
static void register_replay(starpu_data_handle_t handle, unsigned home_node, void *data_interface)
{
	(void) home_node;
	struct replay_interface *replay_interface = data_interface;
	unsigned node;
	for (node = 0; node < STARPU_MAXNODES; node++)
	{
		struct replay_interface *local_interface =
			starpu_data_get_interface_on_node(handle, node);

		local_interface->id = replay_interface->id;
		local_interface->orig_handle = replay_interface->orig_handle;
		local_interface->size = replay_interface->size;
		local_interface->alloc_size = replay_interface->alloc_size;
		local_interface->max_size = replay_interface->max_size;
	}
}

static void replay_data_register(starpu_data_handle_t *handleptr, starpu_data_handle_t orig_handle, int home_node, size_t size, size_t alloc_size, size_t max_size)
{
	if (replay_interface_ops.interfaceid == STARPU_UNKNOWN_INTERFACE_ID)
	{
		replay_interface_ops.interfaceid = starpu_data_interface_get_next_id();
	}
	struct replay_interface interface = {
		.id = replay_interface_ops.interfaceid,
		.orig_handle = orig_handle,
		.size = size,
		.alloc_size = alloc_size,
		.max_size = max_size,
	};

	starpu_data_register(handleptr, home_node, &interface, &replay_interface_ops);
}

static size_t replay_get_size(starpu_data_handle_t handle)
{
	struct replay_interface *interface =
		starpu_data_get_interface_on_node(handle, STARPU_MAIN_RAM);
	return interface->size;
}

static size_t replay_get_alloc_size(starpu_data_handle_t handle)
{
	struct replay_interface *interface =
		starpu_data_get_interface_on_node(handle, STARPU_MAIN_RAM);
	return interface->alloc_size;
}

static size_t replay_get_max_size(starpu_data_handle_t handle)
{
	struct replay_interface *interface =
		starpu_data_get_interface_on_node(handle, STARPU_MAIN_RAM);
	return interface->max_size;
}

static uint32_t replay_footprint(starpu_data_handle_t handle)
{
	return starpu_hash_crc32c_be(replay_get_size(handle), 0);
}

static int replay_compare(void *data_interface_a, void *data_interface_b)
{
	struct replay_interface *replay_a = data_interface_a;
	struct replay_interface *replay_b = data_interface_b;

	/* Two variables are considered compatible if they have the same size */
	return replay_a->size == replay_b->size;
}

static void display_replay(starpu_data_handle_t handle, FILE *f)
{
	struct replay_interface *replay_interface =
		starpu_data_get_interface_on_node(handle, STARPU_MAIN_RAM);

	fprintf(f, "%lu/%lu/%lu\t",
			(unsigned long) replay_interface->size,
			(unsigned long) replay_interface->alloc_size,
			(unsigned long) replay_interface->max_size);
}

static starpu_ssize_t describe_replay(void *data_interface, char *buf, size_t size)
{
	struct replay_interface *replay_interface = data_interface;
	return snprintf(buf, size, "r%lu/%lu/%lu\t",
			(unsigned long) replay_interface->size,
			(unsigned long) replay_interface->alloc_size,
			(unsigned long) replay_interface->max_size);
}

static starpu_ssize_t allocate_replay_on_node(void *data_interface, unsigned dst_node)
{
	struct replay_interface *replay_interface = data_interface;
	starpu_memory_allocate(dst_node, replay_interface->alloc_size, STARPU_MEMORY_OVERFLOW);
	return 0;
}

static void free_replay_on_node(void *data_interface, unsigned dst_node)
{
	struct replay_interface *replay_interface = data_interface;
	starpu_memory_deallocate(dst_node, replay_interface->alloc_size);
}

static int replay_copy(void *src_interface, unsigned src_node, void *dst_interface, unsigned dst_node, void *async_data)
{
	(void) dst_interface;
	struct replay_interface *src = src_interface;

	/* We don't care about pointers */
	return starpu_interface_copy(1, 0, src_node, 1, 0, dst_node, src->size, async_data);
}

static const struct starpu_data_copy_methods replay_copy_data_methods =
{
	.any_to_any = replay_copy,
};

static struct starpu_data_interface_ops replay_interface_ops =
{
	.register_data_handle = register_replay,
	.allocate_data_on_node = allocate_replay_on_node,
	.free_data_on_node = free_replay_on_node,
	.copy_methods = &replay_copy_data_methods,
	.get_size = replay_get_size,
	.get_alloc_size = replay_get_alloc_size,
	.get_max_size = replay_get_max_size,
	.footprint = replay_footprint,
	.compare = replay_compare,
	.interfaceid = STARPU_UNKNOWN_INTERFACE_ID,
	.interface_size = sizeof(struct replay_interface),
	.display = display_replay,
	.pack_data = NULL,
	.unpack_data = NULL,
	.describe = describe_replay,

	/* We want to observe actual allocations/deallocations */
	.dontcache = 1,
};


/* [SUBMITORDER] The tree of the submit order */

static struct starpu_rbtree tree = STARPU_RBTREE_INITIALIZER;

/* the cmp_fn arg for rb_tree_insert() */
unsigned int diff(struct starpu_rbtree_node * left_elm, struct starpu_rbtree_node * right_elm)
{
	long oleft = ((struct task *) left_elm)->submit_order;
	long oright = ((struct task *) right_elm)->submit_order;
	if (oleft == -1 && oright == -1)
	{
		if (left_elm < right_elm)
			return -1;
		else
			return 1;
	}
	return oleft - oright;
}

/* Settings for the perfmodel */
struct task_arg
{
	uint32_t footprint;
	unsigned narch;
	double perf[];
};

uint32_t get_footprint(struct starpu_task * task)
{
	return ((struct task_arg*) (task->cl_arg))->footprint;
}

double arch_cost_function(struct starpu_task *task, struct starpu_perfmodel_arch *arch, unsigned nimpl)
{
	int device = starpu_perfmodel_arch_comb_get(arch->ndevices, arch->devices);
	STARPU_ASSERT(device != -1);
	(void) nimpl;

	/* Then, get the pointer to the value of the expected time */
	struct task_arg *arg = task->cl_arg;
	if (device < (int) arg->narch)
	{
		double val = arg->perf[device];

		if (!(val == 0 || isnan(val)))
			return val;
	}

	fprintf(stderr, "[starpu] Error, expected_time is 0 or lower (replay.c line : %d)", __LINE__- 6);

	return 0.0;
}

/* End of settings */

static unsigned long nexecuted_tasks;
void dumb_kernel(void *buffers[], void *args) {
	(void) buffers;
	(void) args;
	nexecuted_tasks++;
	if (!(nexecuted_tasks % 1000))
	{
		fprintf(stderr, "\rExecuted task %lu...", nexecuted_tasks);
		fflush(stdout);
	}

	unsigned this_worker = starpu_worker_get_id_check();
	struct starpu_perfmodel_arch *perf_arch = starpu_worker_get_perf_archtype(this_worker, STARPU_NMAX_SCHED_CTXS);

	struct starpu_task *task = starpu_task_get_current();
	unsigned impl = starpu_task_get_implementation(task);

	double length = starpu_task_expected_length(task, perf_arch, impl);

	STARPU_ASSERT_MSG(!_STARPU_IS_ZERO(length) && !isnan(length),
			"Codelet %s does not have a perfmodel, or is not calibrated enough, please re-run in non-simgrid mode until it is calibrated",
		starpu_task_get_name(task));

	starpu_sleep(length / 1000000);
}

/* [CODELET] Initialization of an unique codelet for all the tasks*/
static int can_execute(unsigned worker_id, struct starpu_task *task, unsigned nimpl)
{
	struct starpu_perfmodel_arch * arch = starpu_worker_get_perf_archtype(worker_id, STARPU_NMAX_SCHED_CTXS);
	int device = starpu_perfmodel_arch_comb_get(arch->ndevices, arch->devices);
	if (device == -1)
		/* Doesn't exist yet, thus unknown, assuming it can not work there. */
		return 0;
	(void) nimpl;

	/* Then, get the pointer to the value of the expected time */
	struct task_arg *arg = task->cl_arg;
	if (device < (int) arg->narch)
	{
		double val = arg->perf[device];

		if (!(val == 0 || isnan(val)))
			return 1;
	}

	return 0;
}

static struct starpu_perfmodel myperfmodel =
{
	.type = STARPU_PER_ARCH,
	.arch_cost_function = arch_cost_function,
	.footprint = get_footprint,
};

static struct starpu_codelet cl =
{
	.cpu_funcs = { dumb_kernel },
	.cpu_funcs_name = { "dumb_kernel" },
	.cuda_funcs = { dumb_kernel },
	.opencl_funcs = { dumb_kernel },
	.nbuffers = STARPU_VARIABLE_NBUFFERS,
	.can_execute = can_execute,
	.model = &myperfmodel,
	.flags = STARPU_CODELET_SIMGRID_EXECUTE,
};


/* * * * * * * * * * * * * *
* * * * * Functions * * * * *
* * * * * * * * * * * * * * */


/* The following function checks if the program has to use static or dynamic arrays*/
static int set_alloc_mode(int total_parameters)
{
	return total_parameters <= STARPU_NMAXBUFS;
}

/* According to the allocation mode, modify handles_ptr and modes_ptr in static or dynamic */
static void arrays_managing(int mode)
{
	if (mode)
	{
		handles_ptr = &handles[0];
		modes_ptr = &modes[0];
		reg_signal = &normal_reg_signal[0];
	}
	else
	{
		_STARPU_MALLOC(handles_ptr, sizeof(*handles_ptr) * nb_parameters);
		_STARPU_MALLOC(modes_ptr, sizeof(*modes_ptr) * nb_parameters);
		_STARPU_CALLOC(reg_signal, nb_parameters, sizeof(char *));

	}
}

/* Check if a handle hasn't been registered yet */
static void variable_data_register_check(size_t * array_of_size, int nb_handles)
{
	int h, i;
	starpu_data_handle_t orig_handles[nb_handles];

	ARRAY_DUP(handles_ptr, orig_handles, nb_handles);

	for (h = 0 ; h < nb_handles ; h++)
	{
		if(reg_signal[h]) /* Get the register signal, if it's 1 do ... */
		{
			struct handle * handles_cell;

			for (i = 0; i < h; i++)
			{
				/* Maybe we just registered it in this very h loop */
				if (handles_ptr[h] == orig_handles[i])
				{
					handles_ptr[h] = handles_ptr[i];
					break;
				}
			}

			if (i == h)
			{
				_STARPU_MALLOC(handles_cell, sizeof(*handles_cell));
				STARPU_ASSERT(handles_cell != NULL);

				handles_cell->handle = handles_ptr[h]; /* Get the hidden key (initial handle from the file) to store it as a key*/

				replay_data_register(handles_ptr+h, handles_ptr[h],
						modes_ptr[h] & STARPU_R ? STARPU_MAIN_RAM : -1,
						array_of_size[h], array_of_size[h], array_of_size[h]);

				handles_cell->mem_ptr = handles_ptr[h]; /* Store the new value of the handle into the hash table */

				HASH_ADD(hh, handles_hash, handle, sizeof(handles_ptr[h]), handles_cell);
			}
		}
	}
}

void reset(void)
{
	control = NormalTask;

	if (name != NULL)
	{
		free(name);
		name = NULL;
	}

	if (model != NULL)
	{
		free(model);
		model = NULL;
	}

	if (sizes_set != NULL)
	{
		free(sizes_set);
		sizes_set = NULL;
	}

	if (reg_signal != NULL)
	{
		if (!alloc_mode)
		{
			free(reg_signal);
			reg_signal = NULL;
		}
		else
		{
			ARRAY_INIT(reg_signal, nb_parameters);
		}
	}

	jobid = 0;
	ndependson = 0;
	tag = -1;
	workerid = -1;
	footprint = 0;
	startTime = 0.0;
	endTime = 0.0;

	if (submitorder != -1)
		submitorder = -1;

	iteration = -1;
	nb_parameters = 0;
	alloc_mode = 1;
}

void fix_wontuse_handle(struct task * wontuseTask)
{
	STARPU_ASSERT(wontuseTask);

	if (!wontuseTask->reg_signal)
		/* Data was already registered when we created this task, so it's already a handle */
		return;

	struct handle *handle_tmp;

	/* Data was not registered when we created this task, so this is the application pointer, look it up now */
	HASH_FIND(hh, handles_hash, &wontuseTask->task.handles[0], sizeof(wontuseTask->task.handles[0]), handle_tmp);

	if (handle_tmp)
		wontuseTask->task.handles[0] = handle_tmp->mem_ptr;
	else
		/* This data wasn't actually used, don't care about it */
		wontuseTask->task.handles[0] = NULL;
}

/* Function that submits all the tasks (used when the program reaches EOF) */
int submit_tasks(void)
{
	/* Add dependencies */

	const struct starpu_rbtree * tmptree = &tree;
	struct starpu_rbtree_node * currentNode = starpu_rbtree_first(tmptree);
	long last_submitorder = 0;

	while (currentNode != NULL)
	{
		struct task * currentTask = (struct task *) currentNode;

		if (currentTask->type == NormalTask)
		{
			if (currentTask->submit_order != -1)
			{
				STARPU_ASSERT(currentTask->submit_order >= last_submitorder + 1);

				while (currentTask->submit_order > last_submitorder + 1)
				{
					/* Oops, some tasks were not submitted by original application, fake some */
					struct starpu_task *task = starpu_task_create();
					int ret;
					task->cl = NULL;
					ret = starpu_task_submit(task);
					STARPU_ASSERT(ret == 0);
					last_submitorder++;
				}
			}

			if (currentTask->ndependson > 0)
			{
				struct starpu_task * taskdeps[currentTask->ndependson];
				unsigned i, j = 0;

				for (i = 0; i < currentTask->ndependson; i++)
				{
					struct task * taskdep;

					/*  Get the ith jobid of deps_jobid */
					HASH_FIND(hh, tasks, &currentTask->deps[i], sizeof(jobid), taskdep);

					if(taskdep)
					{
						taskdeps[j] = &taskdep->task;
						j ++;
					}
				}

				starpu_task_declare_deps_array(&currentTask->task, j, taskdeps);
			}

			if (!(currentTask->iteration == -1))
				starpu_iteration_push(currentTask->iteration);

			applySchedRec(&currentTask->task, currentTask->submit_order);
			int ret_val = starpu_task_submit(&currentTask->task);

			if (!(currentTask->iteration == -1))
				starpu_iteration_pop();

			if (ret_val != 0)
			{
				fprintf(stderr, "\nWhile submitting task %ld (%s): return %d\n",
						currentTask->submit_order,
						currentTask->task.name? currentTask->task.name : "unknown",
						ret_val);
				return -1;
			}


			//fprintf(stderr, "submitting task %s (%lu, %llu)\n", currentTask->task.name?currentTask->task.name:"anonymous", currentTask->jobid, (unsigned long long) currentTask->task.tag_id);
			if (!(currentTask->submit_order % 1000))
			{
				fprintf(stderr, "\rSubmitted task order %ld...", currentTask->submit_order);
				fflush(stdout);
			}
			if (currentTask->submit_order != -1)
				last_submitorder++;
		}

		else
		{
			fix_wontuse_handle(currentTask); /* Add the handle in the wontuse task */
                        /* FIXME: can not actually work properly since we have
                         * disabled sequential consistency, so we don't have any
                         * easy way to make this wait for the last task that
                         * wrote to the handle. */
			if (currentTask->task.handles[0])
				starpu_data_wont_use(currentTask->task.handles[0]);
		}

		currentNode = starpu_rbtree_next(currentNode);

	}
	fprintf(stderr, " done.\n");

	return 1;
}


/* * * * * * * * * * * * * * * */
/* * * * * * MAIN * * * * * * */
/* * * * * * * * * * * * * * */

static void usage(const char *program)
{
	fprintf(stderr,"Usage: %s [--static-workerid] tasks.rec [sched.rec]\n", program);
	exit(EXIT_FAILURE);
}

int main(int argc, char **argv)
{
	starpu_data_set_default_sequential_consistency_flag(0);

	FILE *rec;
	char *s;
	const char *tasks_rec = NULL;
	const char *sched_rec = NULL;
	unsigned i;
	size_t s_allocated = 128;

	unsigned long nread_tasks = 0;

	_STARPU_MALLOC(s, s_allocated);
	dependson_size = REPLAY_NMAX_DEPENDENCIES; /* Change the value of REPLAY_NMAX_DEPENCIES to modify the number of dependencies */
	_STARPU_MALLOC(dependson, dependson_size * sizeof (* dependson));
	alloc_mode = 1;

	for (i = 1; i < (unsigned) argc; i++)
	{
		if (!strcmp(argv[i], "--help") || !strcmp(argv[i], "-h"))
		{
			usage(argv[0]);
		}
		else if (!strcmp(argv[i], "--static-workerid"))
		{
			static_workerid = 1;
		}
		else
		{
			if (!tasks_rec)
				tasks_rec = argv[i];
			else if (!sched_rec)
				sched_rec = argv[i];
			else
				usage(argv[0]);
		}
	}

	if (!tasks_rec)
		usage(argv[0]);

	if (sched_rec)
		schedRecInit(sched_rec);

	rec = fopen(tasks_rec, "r");
	if (!rec)
	{
		fprintf(stderr,"unable to open file %s: %s\n", tasks_rec, strerror(errno));
		exit(EXIT_FAILURE);
	}

	int ret = starpu_init(NULL);
	if (ret == -ENODEV) goto enodev;

	/* Read line by line, and on empty line submit the task with the accumulated information */
	reset();

	double start = starpu_timing_now();
	int linenum = 0;

	while(1)
	{
		char *ln;

		if (!fgets(s, s_allocated, rec))
		{
			fprintf(stderr, " done.\n");
			int submitted = submit_tasks();

			if (submitted == -1)
			{
				goto enodev;
			}

			goto eof;
		}

		while (!(ln = strchr(s, '\n')))
		{
			/* fprintf(stderr,"buffer size %d too small, doubling it\n", s_allocated); */
			_STARPU_REALLOC(s, s_allocated * 2);

			if (!fgets(s + s_allocated-1, s_allocated+1, rec))
			{
				fprintf(stderr, "\n");
				int submitted = submit_tasks();

				if (submitted == -1)
				{
					goto enodev;
				}

				goto eof;
			}

			s_allocated *= 2;
		}

		linenum++;

		if (ln == s)
		{
			/* Empty line, do task */

			struct task * task;
			_STARPU_MALLOC(task, sizeof(*task));

			starpu_task_init(&task->task);
			task->deps = NULL;

			task->submit_order = submitorder;

			starpu_rbtree_node_init(&task->node);
			starpu_rbtree_insert(&tree, &task->node, diff);


			task->jobid = jobid;
			task->iteration = iteration;

			if (name != NULL)
				task->task.name = strdup(name);

			task->type = control;

			if (control == NormalTask)
			{
				if (workerid >= 0)
				{
					task->task.priority = priority;
					task->task.cl = &cl;
					if (static_workerid)
					{
						task->task.workerid = workerid;
						task->task.execute_on_a_specific_worker = 1;
					}

					if (alloc_mode)
					{
						/* Duplicating the handles stored (and registered in the current context) into the task */

						ARRAY_DUP(modes_ptr, task->task.modes, nb_parameters);
						ARRAY_DUP(modes_ptr, task->task.cl->modes, nb_parameters);
						variable_data_register_check(sizes_set, nb_parameters);
						ARRAY_DUP(handles_ptr, task->task.handles, nb_parameters);
					}
					else
					{
						task->task.dyn_modes = modes_ptr;
						_STARPU_MALLOC(task->task.cl->dyn_modes, (sizeof(*task->task.cl->dyn_modes) * nb_parameters));
						ARRAY_DUP(modes_ptr, task->task.cl->dyn_modes, nb_parameters);
						variable_data_register_check(sizes_set, nb_parameters);
						task->task.dyn_handles = handles_ptr;
					}

					task->task.nbuffers = nb_parameters;

					struct perfmodel * realmodel;

					HASH_FIND_STR(model_hash, model, realmodel);

					if (realmodel == NULL)
					{
						int len = strlen(model);
						_STARPU_CALLOC(realmodel, 1, sizeof(struct perfmodel));

						_STARPU_MALLOC(realmodel->model_name, sizeof(char) * (len+1));
						realmodel->model_name = strcpy(realmodel->model_name, model);

						starpu_perfmodel_init(&realmodel->perfmodel);

						int error = starpu_perfmodel_load_symbol(model, &realmodel->perfmodel);

						if (!error)
						{
							HASH_ADD_STR(model_hash, model_name, realmodel);
						}
						else
						{

							fprintf(stderr, "[starpu][Warning] Error loading perfmodel symbol %s\n", model);
							fprintf(stderr, "[starpu][Warning] Taking only measurements from the given execution, and forcing execution on worker %d\n", workerid);
							starpu_perfmodel_unload_model(&realmodel->perfmodel);
							free(realmodel->model_name);
							free(realmodel);
							realmodel = NULL;
						}

					}

					struct starpu_perfmodel_arch *arch = starpu_worker_get_perf_archtype(workerid, 0);

					unsigned comb = starpu_perfmodel_arch_comb_add(arch->ndevices, arch->devices);
					unsigned narch = starpu_perfmodel_get_narch_combs();

					struct task_arg *arg;
					_STARPU_MALLOC(arg, sizeof(struct task_arg) + sizeof(double) * narch);
					arg->footprint = footprint;
					arg->narch = narch;
					double * perfTime  = arg->perf;

					if (realmodel == NULL)
					{
						/* Erf, do without perfmodel, for execution there */
						task->task.workerid = workerid;
						task->task.execute_on_a_specific_worker = 1;
						for (i = 0; i < narch ; i++)
						{
							if (i == comb)
								perfTime[i] = endTime - startTime;
							else
								perfTime[i] = NAN;
						}
					}
					else
					{
						int one = 0;
						for (i = 0; i < narch ; i++)
						{
							arch = starpu_perfmodel_arch_comb_fetch(i);
							perfTime[i] = starpu_perfmodel_history_based_expected_perf(&realmodel->perfmodel, arch, footprint);
							if (!(perfTime[i] == 0 || isnan(perfTime[i])))
								one = 1;
						}
						if (!one)
						{
							fprintf(stderr, "We do not have any performance measurement for symbol '%s' for footprint %x, we can not execute this", model, footprint);
							exit(EXIT_FAILURE);
						}
					}

					task->task.cl_arg = arg;
					task->task.flops = flops;
					total_flops += flops;
				}

				task->task.cl_arg_size = 0;
				task->task.tag_id = tag;
				task->task.use_tag = 1;

				task->ndependson = ndependson;
				if (ndependson > 0)
				{
					_STARPU_MALLOC(task->deps, ndependson * sizeof (* task->deps));
					ARRAY_DUP(dependson, task->deps, ndependson);
				}
			}

			else
			{
				STARPU_ASSERT(nb_parameters == 1);
				task->reg_signal = reg_signal[0];
				ARRAY_DUP(handles_ptr, task->task.handles, nb_parameters);
			}

			/* Add this task to task hash */
			HASH_ADD(hh, tasks, jobid, sizeof(jobid), task);

			nread_tasks++;
			if (!(nread_tasks % 1000))
			{
				fprintf(stderr, "\rRead task %lu...", nread_tasks);
				fflush(stdout);
			}

			reset();
		}

		/* Record various information */
#define TEST(field) (!strncmp(s, field": ", strlen(field) + 2))

		else if(TEST("Control"))
		{
			char * c = s+9;

			if(!strncmp(c, "WontUse", 7))
			{
				control = WontUseTask;
				nb_parameters = 1;
				alloc_mode = set_alloc_mode(nb_parameters);
				arrays_managing(alloc_mode);
			}
			else
				control = NormalTask;
		}
		else if (TEST("Name"))
		{
			*ln = 0;
			name = strdup(s+6);
		}
		else if (TEST("Model"))
		{
			*ln = 0;
			model = strdup(s+7);
		}
		else if (TEST("JobId"))
			jobid = atol(s+7);
		else if(TEST("SubmitOrder"))
			submitorder = atoi(s+13);
		else if (TEST("DependsOn"))
		{
			char *c = s + 11;

			for (ndependson = 0; *c != '\n'; ndependson++)
			{
				if (ndependson >= dependson_size)
				{
					dependson_size *= 2;
					_STARPU_REALLOC(dependson, dependson_size * sizeof(*dependson));
				}
				dependson[ndependson] = strtol(c, &c, 10);
			}
		}
		else if (TEST("Tag"))
		{
			tag = strtol(s+5, NULL, 16);
		}
		else if (TEST("WorkerId"))
		{
			workerid = atoi(s+10);
		}
		else if (TEST("Footprint"))
		{
			footprint = strtoul(s+11, NULL, 16);
		}
		else if (TEST("Parameters"))
		{
			/* Parameters line format is PARAM1 PARAM2 (...)PARAMi (...)PARAMn */
			char * param_str = s + 12;
			int count = 0;

			for (i = 0 ; param_str[i] != '\n'; i++)
			{
				if (param_str[i] == ' ') /* Checking the number of ' ' (space), assuming that the file is not corrupted */
				{
					count++;
				}
			}

			nb_parameters = count + 1; /* There is one space per paramater except for the last one, that's why we have to add +1 (dirty programming) */

			/* This part of the algorithm will determine if it needs static or dynamic arrays */
			alloc_mode = set_alloc_mode(nb_parameters);
			arrays_managing(alloc_mode);

		}
		else if (TEST("Handles"))
		{
			char *buffer = s + 9;
			const char *delim = " ";
			char *token = strtok(buffer, delim);

			for (i = 0 ; i < nb_parameters ; i++)
			{
				STARPU_ASSERT(token);
				struct handle *handles_cell; /* A cell of the hash table for the handles */
				starpu_data_handle_t  handle_value = (starpu_data_handle_t) strtol(token, NULL, 16); /* Get the ith handle on the line (in the file) */

				HASH_FIND(hh, handles_hash, &handle_value, sizeof(handle_value), handles_cell); /* Find if the handle_value was already registered as a key in the hash table */

				/* If it wasn't, then add it to the hash table */
				if (handles_cell == NULL)
				{
					/* Hide the initial handle from the file into the handles array to find it when necessary */
					handles_ptr[i] = handle_value;
					reg_signal[i] = 1;
				}
				else
				{
					handles_ptr[i] = handles_cell->mem_ptr;
					reg_signal[i] = 0;
				}

				token = strtok(NULL, delim);
			}
		}
		else if (TEST("Modes"))
		{
			char * buffer = s + 7;
			unsigned mode_i = 0;
			const char * delim = " ";
			char * token = strtok(buffer, delim);

			while (token != NULL && mode_i < nb_parameters)
			{
				/* Subject to the names of starpu modes enumerator are not modified */
				if (!strncmp(token, "RW", 2))
				{
					*(modes_ptr+mode_i) = STARPU_RW;
					mode_i++;
				}
				else if (!strncmp(token, "R", 1))
				{
					*(modes_ptr+mode_i) = STARPU_R;
					mode_i++;
				}
				else if (!strncmp(token, "W", 1))
				{
					*(modes_ptr+mode_i) = STARPU_W;
					mode_i++;
				}
				/* Other cases produce a warning*/
				else
				{
					fprintf(stderr, "[Warning] A mode is different from R/W (jobid task : %lu)", jobid);
				}
				token = strtok(NULL, delim);
			}
		}
		else if (TEST("Sizes"))
		{
			char *  buffer = s + 7;
			const char * delim = " ";
			char * token = strtok(buffer, delim);
			unsigned k = 0;

			_STARPU_MALLOC(sizes_set, nb_parameters * sizeof(size_t));

			while (token != NULL && k < nb_parameters)
			{
				sizes_set[k] = strtol(token, NULL, 10);
				token = strtok(NULL, delim);

				k++;
			}
		}
		else if (TEST("StartTime"))
		{
			startTime = strtod(s+11, NULL);
		}
		else if (TEST("EndTime"))
		{
			endTime = strtod(s+9, NULL);
		}
		else if (TEST("GFlop"))
		{
			flops = 1000000000 * strtod(s+7, NULL);
		}
		else if (TEST("Iteration"))
		{
			iteration = (unsigned) strtol(s+11, NULL, 10);
		}
		else if (TEST("Priority"))
		{
			priority = strtol(s + 10, NULL, 10);
		}
	}

eof:

	starpu_task_wait_for_all();
	fprintf(stderr, " done.\n");

	printf("%g ms", (starpu_timing_now() - start) / 1000.);
	if (total_flops != 0.)
		printf("\t%g GF/s", (total_flops / (starpu_timing_now() - start)) / 1000.);
	printf("\n");

	/* FREE allocated memory */

	free(dependson);
	free(s);

	/* End of FREE */

	struct handle *handle=NULL, *handletmp=NULL;
	HASH_ITER(hh, handles_hash, handle, handletmp)
	{
		starpu_data_unregister(handle->mem_ptr);
		HASH_DEL(handles_hash, handle);
		free(handle);
        }

	struct perfmodel *model_s=NULL, *modeltmp=NULL;
	HASH_ITER(hh, model_hash, model_s, modeltmp)
	{
		starpu_perfmodel_unload_model(&model_s->perfmodel);
		HASH_DEL(model_hash, model_s);
		free(model_s->model_name);
		free(model_s);
        }

	struct task *task=NULL, *tasktmp=NULL;
	HASH_ITER(hh, tasks, task, tasktmp)
	{
		free(task->task.cl_arg);
		free((char*)task->task.name);

		if (task->task.dyn_handles != NULL)
		{
			free(task->task.dyn_handles);
			free(task->task.dyn_modes);
		}

		HASH_DEL(tasks, task);
		starpu_task_clean(&task->task);
		free(task->deps);
		starpu_rbtree_remove(&tree, &task->node);
		free(task);
        }

	starpu_shutdown();
	return 0;

enodev:
	starpu_shutdown();
	return 77;
}