-
Peter W. Draper authoredReported by @jborrow, closes #618.
Peter W. Draper authoredReported by @jborrow, closes #618.
scheduler.c 70.93 KiB
/*******************************************************************************
* This file is part of SWIFT.
* Copyright (c) 2012 Pedro Gonnet (pedro.gonnet@durham.ac.uk)
* Matthieu Schaller (matthieu.schaller@durham.ac.uk)
* 2016 Peter W. Draper (p.w.draper@durham.ac.uk)
*
* This program 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 3 of the License, or
* (at your option) any later version.
*
* This program 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 General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public License
* along with this program. If not, see <http://www.gnu.org/licenses/>.
*
******************************************************************************/
/* Config parameters. */
#include "../config.h"
/* Some standard headers. */
#include <limits.h>
#include <math.h>
#include <pthread.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/stat.h>
/* MPI headers. */
#ifdef WITH_MPI
#include <mpi.h>
#endif
/* This object's header. */
#include "scheduler.h"
/* Local headers. */
#include "atomic.h"
#include "cycle.h"
#include "engine.h"
#include "error.h"
#include "intrinsics.h"
#include "kernel_hydro.h"
#include "memuse.h"
#include "mpiuse.h"
#include "queue.h"
#include "sort_part.h"
#include "space.h"
#include "space_getsid.h"
#include "task.h"
#include "timers.h"
#include "version.h"
/**
* @brief Re-set the list of active tasks.
*/
void scheduler_clear_active(struct scheduler *s) { s->active_count = 0; }
/**
* @brief Increase the space available for unlocks. Only call when
* current index == s->size_unlock;
*/
static void scheduler_extend_unlocks(struct scheduler *s) {
/* Allocate the new buffer. */
const int size_unlocks_new = s->size_unlocks * 2; struct task **unlocks_new = (struct task **)swift_malloc(
"unlocks", sizeof(struct task *) * size_unlocks_new);
int *unlock_ind_new =
(int *)swift_malloc("unlock_ind", sizeof(int) * size_unlocks_new);
if (unlocks_new == NULL || unlock_ind_new == NULL)
error("Failed to re-allocate unlocks.");
/* Wait for all writes to the old buffer to complete. */
while (s->completed_unlock_writes < s->size_unlocks)
;
/* Copy the buffers. */
memcpy(unlocks_new, s->unlocks, sizeof(struct task *) * s->size_unlocks);
memcpy(unlock_ind_new, s->unlock_ind, sizeof(int) * s->size_unlocks);
swift_free("unlocks", s->unlocks);
swift_free("unlock_ind", s->unlock_ind);
s->unlocks = unlocks_new;
s->unlock_ind = unlock_ind_new;
/* Publish the new buffer size. */
s->size_unlocks = size_unlocks_new;
}
/**
* @brief Add an unlock_task to the given task.
*
* @param s The #scheduler.
* @param ta The unlocking #task.
* @param tb The #task that will be unlocked.
*/
void scheduler_addunlock(struct scheduler *s, struct task *ta,
struct task *tb) {
#ifdef SWIFT_DEBUG_CHECKS
if (ta == NULL) error("Unlocking task is NULL.");
if (tb == NULL) error("Unlocked task is NULL.");
#endif
/* Get an index at which to store this unlock. */
const int ind = atomic_inc(&s->nr_unlocks);
/* Does the buffer need to be grown? */
if (ind == s->size_unlocks) scheduler_extend_unlocks(s);
#ifdef SWIFT_DEBUG_CHECKS
if (ind > s->size_unlocks * 2)
message("unlocks guard enabled: %d / %d", ind, s->size_unlocks);
#endif
/* Wait for there to actually be space at my index. */
while (ind > s->size_unlocks)
;
/* Guard against case when more than (old) s->size_unlocks unlocks
* are now pending. */
if (ind == s->size_unlocks) scheduler_extend_unlocks(s);
/* Write the unlock to the scheduler. */
s->unlocks[ind] = tb;
s->unlock_ind[ind] = ta - s->tasks;
atomic_inc(&s->completed_unlock_writes);
}
/**
* @brief compute the number of similar dependencies
*
* @param s The #scheduler
* @param ta The #task
* @param tb The dependent #task
* * @return Number of dependencies
*/
int scheduler_get_number_relation(const struct scheduler *s,
const struct task *ta,
const struct task *tb) {
int count = 0;
/* loop over all tasks */
for (int i = 0; i < s->nr_tasks; i++) {
const struct task *ta_tmp = &s->tasks[i];
/* and their dependencies */
for (int j = 0; j < ta->nr_unlock_tasks; j++) {
const struct task *tb_tmp = ta->unlock_tasks[j];
if (ta->type == ta_tmp->type && ta->subtype == ta_tmp->subtype &&
tb->type == tb_tmp->type && tb->subtype == tb_tmp->subtype) {
count += 1;
}
}
}
return count;
}
/* Conservative number of dependencies per task type */
#define MAX_NUMBER_DEP 128
/**
* @brief Informations about all the task dependencies of
* a single task.
*/
struct task_dependency {
/* Main task */
/* ID of the task */
int type_in;
/* ID of the subtask */
int subtype_in;
/* Is the task implicit */
int implicit_in;
/* Dependent task */
/* ID of the dependent task */
int type_out[MAX_NUMBER_DEP];
/* ID of the dependent subtask */
int subtype_out[MAX_NUMBER_DEP];
/* Is the dependent task implicit */
int implicit_out[MAX_NUMBER_DEP];
/* Statistics */
/* number of link between the two task type */
int number_link[MAX_NUMBER_DEP];
/* number of ranks having this relation */
int number_rank[MAX_NUMBER_DEP];
};
#ifdef WITH_MPI
/**
* @brief Define the #task_dependency for MPI
*
* @param tstype The MPI_Datatype to initialize
*/
void task_dependency_define(MPI_Datatype *tstype) {
/* Define the variables */
const int count = 8; int blocklens[count];
MPI_Datatype types[count];
MPI_Aint disps[count];
/* all the type are int */
for (int i = 0; i < count; i++) {
types[i] = MPI_INT;
}
/* Task in */
disps[0] = offsetof(struct task_dependency, type_in);
blocklens[0] = 1;
disps[1] = offsetof(struct task_dependency, subtype_in);
blocklens[1] = 1;
disps[2] = offsetof(struct task_dependency, implicit_in);
blocklens[2] = 1;
/* Task out */
disps[3] = offsetof(struct task_dependency, type_out);
blocklens[3] = MAX_NUMBER_DEP;
disps[4] = offsetof(struct task_dependency, subtype_out);
blocklens[4] = MAX_NUMBER_DEP;
disps[5] = offsetof(struct task_dependency, implicit_out);
blocklens[5] = MAX_NUMBER_DEP;
/* statistics */
disps[6] = offsetof(struct task_dependency, number_link);
blocklens[6] = MAX_NUMBER_DEP;
disps[7] = offsetof(struct task_dependency, number_rank);
blocklens[7] = MAX_NUMBER_DEP;
/* define it for MPI */
MPI_Type_create_struct(count, blocklens, disps, types, tstype);
MPI_Type_commit(tstype);
}
/**
* @brief Sum operator of #task_dependency for MPI
*
* @param in_p The #task_dependency to add
* @param out_p The #task_dependency where in_p is added
* @param len The length of the arrays
* @param type The MPI datatype
*/
void task_dependency_sum(void *in_p, void *out_p, int *len,
MPI_Datatype *type) {
/* change pointer type */
struct task_dependency *in = (struct task_dependency *)in_p;
struct task_dependency *out = (struct task_dependency *)out_p;
/* Loop over all the current objects */
for (int i = 0; i < *len; i++) {
/* loop over all the object set in invals */
for (int j = 0; j < MAX_NUMBER_DEP; j++) {
/* Have we reached the end of the links? */
if (in[i].number_link[j] == -1) {
break;
}
/* get a few variables */
int tb_type = in[i].type_out[j];
int tb_subtype = in[i].subtype_out[j];
#ifdef SWIFT_DEBUG_CHECKS
/* Check tasks */
if (tb_type >= task_type_count) {
error("Unknown task type %i", tb_type);
}
if (tb_subtype >= task_subtype_count) { error("Unknown subtask type %i", tb_subtype);
}
#endif
/* find the corresponding id */
int k = 0;
while (k < MAX_NUMBER_DEP) {
/* have we reached the end of the links? */
if (out[i].number_link[k] == -1) {
/* reset the counter in order to be safe */
out[i].number_link[k] = 0;
out[i].number_rank[k] = 0;
/* set the relation */
out[i].type_in = in[i].type_in;
out[i].subtype_in = in[i].subtype_in;
out[i].implicit_in = in[i].implicit_in;
out[i].type_out[k] = in[i].type_out[j];
out[i].subtype_out[k] = in[i].subtype_out[j];
out[i].implicit_out[k] = in[i].implicit_out[j];
break;
}
/* do we have the same relation? */
if (out[i].type_out[k] == tb_type &&
out[i].subtype_out[k] == tb_subtype) {
break;
}
k++;
}
/* Check if we are still in the memory */
if (k == MAX_NUMBER_DEP) {
error("Not enough memory, please increase MAX_NUMBER_DEP");
}
#ifdef SWIFT_DEBUG_CHECKS
/* Check if correct relation */
if (out[i].type_in != in[i].type_in ||
out[i].subtype_in != in[i].subtype_in ||
out[i].implicit_in != in[i].implicit_in ||
out[i].type_out[k] != in[i].type_out[j] ||
out[i].subtype_out[k] != in[i].subtype_out[j] ||
out[i].implicit_out[k] != in[i].implicit_out[j]) {
error("Tasks do not correspond");
}
#endif
/* sum the contributions */
out[i].number_link[k] += in[i].number_link[j];
out[i].number_rank[k] += in[i].number_rank[j];
}
}
return;
}
#endif // WITH_MPI
/**
* @brief Write a dot file with the task dependencies.
*
* Run plot_task_dependencies.sh for an example of how to use it
* to generate the figure.
*
* @param s The #scheduler we are working in.
* @param verbose Are we verbose about this?
*/void scheduler_write_dependencies(struct scheduler *s, int verbose) {
const ticks tic = getticks();
/* Number of possible relations between tasks */
const int nber_tasks = task_type_count * task_subtype_count;
/* To get the table for a task:
* ind = (ta * task_subtype_count + sa)
* where ta is the value of task_type and sa is the value of
* task_subtype */
struct task_dependency *task_dep = (struct task_dependency *)malloc(
nber_tasks * sizeof(struct task_dependency));
if (task_dep == NULL)
error("Error allocating memory for task-dependency graph (table).");
/* Reset counter */
for (int i = 0; i < nber_tasks; i++) {
for (int j = 0; j < MAX_NUMBER_DEP; j++) {
/* Use number_link as indicator of the existance of a relation */
task_dep[i].number_link[j] = -1;
}
}
/* loop over all tasks */
for (int i = 0; i < s->nr_tasks; i++) {
const struct task *ta = &s->tasks[i];
/* Current index */
const int ind = ta->type * task_subtype_count + ta->subtype;
struct task_dependency *cur = &task_dep[ind];
/* Set ta */
cur->type_in = ta->type;
cur->subtype_in = ta->subtype;
cur->implicit_in = ta->implicit;
/* and their dependencies */
for (int j = 0; j < ta->nr_unlock_tasks; j++) {
const struct task *tb = ta->unlock_tasks[j];
int k = 0;
while (k < MAX_NUMBER_DEP) {
/* not written yet */
if (cur->number_link[k] == -1) {
/* set tb */
cur->type_out[k] = tb->type;
cur->subtype_out[k] = tb->subtype;
cur->implicit_out[k] = tb->implicit;
/* statistics */
const int count = scheduler_get_number_relation(s, ta, tb);
cur->number_link[k] = count;
cur->number_rank[k] = 1;
break;
}
/* already written */
if (cur->type_out[k] == tb->type &&
cur->subtype_out[k] == tb->subtype) {
break;
}
k += 1;
}
/* MAX_NUMBER_DEP is too small */
if (k == MAX_NUMBER_DEP) error("Not enough memory, please increase MAX_NUMBER_DEP");
}
}
#ifdef WITH_MPI
/* create MPI operator */
MPI_Datatype data_type;
task_dependency_define(&data_type);
MPI_Op sum;
MPI_Op_create(task_dependency_sum, /* commute */ 1, &sum);
/* create recv buffer */
struct task_dependency *recv = NULL;
if (s->nodeID == 0) {
recv = (struct task_dependency *)malloc(nber_tasks *
sizeof(struct task_dependency));
/* reset counter */
for (int i = 0; i < nber_tasks; i++) {
for (int j = 0; j < MAX_NUMBER_DEP; j++) {
/* Use number_link as indicator of the existance of a relation */
recv[i].number_link[j] = -1;
}
}
}
/* Do the reduction */
int test =
MPI_Reduce(task_dep, recv, nber_tasks, data_type, sum, 0, MPI_COMM_WORLD);
if (test != MPI_SUCCESS) error("MPI reduce failed");
/* free some memory */
if (s->nodeID == 0) {
free(task_dep);
task_dep = recv;
}
#endif
if (s->nodeID == 0) {
/* Create file */
const char *filename = "dependency_graph.csv";
FILE *f = fopen(filename, "w");
if (f == NULL) error("Error opening dependency graph file.");
/* Write header */
fprintf(f, "# %s\n", git_revision());
fprintf(
f,
"task_in,task_out,implicit_in,implicit_out,mpi_in,mpi_out,cluster_in,"
"cluster_out,number_link,number_rank\n");
for (int i = 0; i < nber_tasks; i++) {
for (int j = 0; j < MAX_NUMBER_DEP; j++) {
/* Does this link exists */
if (task_dep[i].number_link[j] == -1) {
continue;
}
/* Define a few variables */
const int ta_type = task_dep[i].type_in;
const int ta_subtype = task_dep[i].subtype_in;
const int ta_implicit = task_dep[i].implicit_in;
const int tb_type = task_dep[i].type_out[j];
const int tb_subtype = task_dep[i].subtype_out[j];
const int tb_implicit = task_dep[i].implicit_out[j];
const int count = task_dep[i].number_link[j]; const int number_rank = task_dep[i].number_rank[j];
/* text to write */
char ta_name[200];
char tb_name[200];
/* construct line */
task_get_full_name(ta_type, ta_subtype, ta_name);
task_get_full_name(tb_type, tb_subtype, tb_name);
/* Check if MPI */
int ta_mpi = 0;
if (ta_type == task_type_send || ta_type == task_type_recv) ta_mpi = 1;
int tb_mpi = 0;
if (tb_type == task_type_send || tb_type == task_type_recv) tb_mpi = 1;
/* Get group name */
char ta_cluster[20];
char tb_cluster[20];
task_get_group_name(ta_type, ta_subtype, ta_cluster);
task_get_group_name(tb_type, tb_subtype, tb_cluster);
fprintf(f, "%s,%s,%d,%d,%d,%d,%s,%s,%d,%d\n", ta_name, tb_name,
ta_implicit, tb_implicit, ta_mpi, tb_mpi, ta_cluster,
tb_cluster, count, number_rank);
}
}
/* Close the file */
fclose(f);
}
/* Be clean */
free(task_dep);
if (verbose)
message("Printing task graph took %.3f %s.",
clocks_from_ticks(getticks() - tic), clocks_getunit());
}
/**
* @brief Split a hydrodynamic task if too large.
*
* @param t The #task
* @param s The #scheduler we are working in.
*/
static void scheduler_splittask_hydro(struct task *t, struct scheduler *s) {
/* Are we considering both stars and hydro when splitting? */
/* Note this is not very clean as the scheduler should not really
access the engine... */
const int with_feedback = (s->space->e->policy & engine_policy_feedback);
const int with_stars = (s->space->e->policy & engine_policy_stars);
const int with_black_holes =
(s->space->e->policy & engine_policy_black_holes);
/* Iterate on this task until we're done with it. */
int redo = 1;
while (redo) {
/* Reset the redo flag. */
redo = 0;
/* Is this a non-empty self-task? */
const int is_self =
(t->type == task_type_self) && (t->ci != NULL) &&
((t->ci->hydro.count > 0) || (with_stars && t->ci->stars.count > 0) ||
(with_black_holes && t->ci->black_holes.count > 0));
/* Is this a non-empty pair-task? */
const int is_pair = (t->type == task_type_pair) && (t->ci != NULL) &&
(t->cj != NULL) && ((t->ci->hydro.count > 0) ||
(with_feedback && t->ci->stars.count > 0) ||
(with_black_holes && t->ci->black_holes.count > 0)) &&
((t->cj->hydro.count > 0) ||
(with_feedback && t->cj->stars.count > 0) ||
(with_black_holes && t->cj->black_holes.count > 0));
/* Empty task? */
if (!is_self && !is_pair) {
t->type = task_type_none;
t->subtype = task_subtype_none;
t->ci = NULL;
t->cj = NULL;
t->skip = 1;
break;
}
/* Self-interaction? */
if (t->type == task_type_self) {
/* Get a handle on the cell involved. */
struct cell *ci = t->ci;
/* Foreign task? */
if (ci->nodeID != s->nodeID) {
t->skip = 1;
break;
}
/* Is this cell even split and the task does not violate h ? */
if (cell_can_split_self_hydro_task(ci)) {
/* Make a sub? */
if (scheduler_dosub && (ci->hydro.count < space_subsize_self_hydro) &&
(ci->stars.count < space_subsize_self_stars)) {
/* convert to a self-subtask. */
t->type = task_type_sub_self;
/* Otherwise, make tasks explicitly. */
} else {
/* Take a step back (we're going to recycle the current task)... */
redo = 1;
/* Add the self tasks. */
int first_child = 0;
while (ci->progeny[first_child] == NULL) first_child++;
t->ci = ci->progeny[first_child];
cell_set_flag(t->ci, cell_flag_has_tasks);
for (int k = first_child + 1; k < 8; k++) {
/* Do we have a non-empty progenitor? */
if (ci->progeny[k] != NULL &&
(ci->progeny[k]->hydro.count ||
(with_feedback && ci->progeny[k]->stars.count))) {
scheduler_splittask_hydro(
scheduler_addtask(s, task_type_self, t->subtype, 0, 0,
ci->progeny[k], NULL),
s);
}
}
/* Make a task for each pair of progeny */
for (int j = 0; j < 8; j++) {
/* Do we have a non-empty progenitor? */
if (ci->progeny[j] != NULL &&
(ci->progeny[j]->hydro.count ||
(with_feedback && ci->progeny[j]->stars.count))) {
for (int k = j + 1; k < 8; k++) {
/* Do we have a second non-empty progenitor? */
if (ci->progeny[k] != NULL &&
(ci->progeny[k]->hydro.count || (with_feedback && ci->progeny[k]->stars.count))) {
scheduler_splittask_hydro(
scheduler_addtask(s, task_type_pair, t->subtype,
sub_sid_flag[j][k], 0, ci->progeny[j],
ci->progeny[k]),
s);
}
}
}
}
}
} /* Cell is split */
} /* Self interaction */
/* Pair interaction? */
else if (t->type == task_type_pair) {
/* Get a handle on the cells involved. */
struct cell *ci = t->ci;
struct cell *cj = t->cj;
/* Foreign task? */
if (ci->nodeID != s->nodeID && cj->nodeID != s->nodeID) {
t->skip = 1;
break;
}
/* Get the sort ID, use space_getsid and not t->flags
to make sure we get ci and cj swapped if needed. */
double shift[3];
const int sid = space_getsid(s->space, &ci, &cj, shift);
#ifdef SWIFT_DEBUG_CHECKS
if (sid != t->flags)
error("Got pair task with incorrect flags: sid=%d flags=%lld", sid,
t->flags);
#endif
/* Should this task be split-up? */
if (cell_can_split_pair_hydro_task(ci) &&
cell_can_split_pair_hydro_task(cj)) {
const int h_count_i = ci->hydro.count;
const int h_count_j = cj->hydro.count;
const int s_count_i = ci->stars.count;
const int s_count_j = cj->stars.count;
int do_sub_hydro = 1;
int do_sub_stars_i = 1;
int do_sub_stars_j = 1;
if (h_count_i > 0 && h_count_j > 0) {
/* Note: Use division to avoid integer overflow. */
do_sub_hydro =
h_count_i * sid_scale[sid] < space_subsize_pair_hydro / h_count_j;
}
if (s_count_i > 0 && h_count_j > 0) {
/* Note: Use division to avoid integer overflow. */
do_sub_stars_i =
s_count_i * sid_scale[sid] < space_subsize_pair_stars / h_count_j;
}
if (s_count_j > 0 && h_count_i > 0) {
/* Note: Use division to avoid integer overflow. */
do_sub_stars_j =
s_count_j * sid_scale[sid] < space_subsize_pair_stars / h_count_i;
}
/* Replace by a single sub-task? */
if (scheduler_dosub &&
(do_sub_hydro && do_sub_stars_i && do_sub_stars_j) &&
!sort_is_corner(sid)) {
/* Make this task a sub task. */
t->type = task_type_sub_pair;
/* Otherwise, split it. */
} else {
/* Take a step back (we're going to recycle the current task)... */
redo = 1;
/* Loop over the sub-cell pairs for the current sid and add new tasks
* for them. */
struct cell_split_pair *csp = &cell_split_pairs[sid];
t->ci = ci->progeny[csp->pairs[0].pid];
t->cj = cj->progeny[csp->pairs[0].pjd];
cell_set_flag(t->ci, cell_flag_has_tasks);
if (t->cj != NULL)cell_set_flag(t->cj, cell_flag_has_tasks);
t->flags = csp->pairs[0].sid;
for (int k = 1; k < csp->count; k++) {
scheduler_splittask_hydro(
scheduler_addtask(s, task_type_pair, t->subtype,
csp->pairs[k].sid, 0,
ci->progeny[csp->pairs[k].pid],
cj->progeny[csp->pairs[k].pjd]),
s);
}
}
/* Otherwise, break it up if it is too large? */
} else if (scheduler_doforcesplit && ci->split && cj->split &&
(ci->hydro.count > space_maxsize / cj->hydro.count)) {
// message( "force splitting pair with %i and %i parts." ,
// ci->hydro.count , cj->hydro.count );
/* Replace the current task. */
t->type = task_type_none;
for (int j = 0; j < 8; j++)
if (ci->progeny[j] != NULL && ci->progeny[j]->hydro.count)
for (int k = 0; k < 8; k++)
if (cj->progeny[k] != NULL && cj->progeny[k]->hydro.count) {
struct task *tl =
scheduler_addtask(s, task_type_pair, t->subtype, 0, 0,
ci->progeny[j], cj->progeny[k]);
scheduler_splittask_hydro(tl, s);
tl->flags = space_getsid(s->space, &t->ci, &t->cj, shift);
}
}
} /* pair interaction? */
} /* iterate over the current task. */
}
/**
* @brief Split a gravity task if too large.
*
* @param t The #task
* @param s The #scheduler we are working in.
*/
static void scheduler_splittask_gravity(struct task *t, struct scheduler *s) {
const struct space *sp = s->space;
struct engine *e = sp->e;
/* Iterate on this task until we're done with it. */
int redo = 1; while (redo) {
/* Reset the redo flag. */
redo = 0;
/* Non-splittable task? */
if ((t->ci == NULL) || (t->type == task_type_pair && t->cj == NULL)) {
t->type = task_type_none;
t->subtype = task_subtype_none;
t->ci = NULL;
t->cj = NULL;
t->skip = 1;
break;
}
/* Self-interaction? */
if (t->type == task_type_self) {
/* Get a handle on the cell involved. */
const struct cell *ci = t->ci;
/* Foreign task? */
if (ci->nodeID != s->nodeID) {
t->skip = 1;
break;
}
/* Should we split this task? */
if (cell_can_split_self_gravity_task(ci)) {
if (scheduler_dosub && ci->grav.count < space_subsize_self_grav) {
/* Otherwise, split it. */
} else {
/* Take a step back (we're going to recycle the current task)... */
redo = 1;
/* Add the self tasks. */
int first_child = 0;
while (ci->progeny[first_child] == NULL) first_child++;
t->ci = ci->progeny[first_child];
cell_set_flag(t->ci, cell_flag_has_tasks);
for (int k = first_child + 1; k < 8; k++)
if (ci->progeny[k] != NULL)
scheduler_splittask_gravity(
scheduler_addtask(s, task_type_self, t->subtype, 0, 0,
ci->progeny[k], NULL),
s);
/* Make a task for each pair of progeny */
if (t->subtype != task_subtype_external_grav) {
for (int j = 0; j < 8; j++)
if (ci->progeny[j] != NULL)
for (int k = j + 1; k < 8; k++)
if (ci->progeny[k] != NULL)
scheduler_splittask_gravity(
scheduler_addtask(s, task_type_pair, t->subtype,
sub_sid_flag[j][k], 0, ci->progeny[j],
ci->progeny[k]),
s);
} /* Self-gravity only */
} /* Make tasks explicitly */
} /* Cell is split */
} /* Self interaction */
/* Pair interaction? */
else if (t->type == task_type_pair) {
/* Get a handle on the cells involved. */
struct cell *ci = t->ci;
struct cell *cj = t->cj;
/* Foreign task? */
if (ci->nodeID != s->nodeID && cj->nodeID != s->nodeID) {
t->skip = 1;
break;
}
/* Should this task be split-up? */
if (cell_can_split_pair_gravity_task(ci) &&
cell_can_split_pair_gravity_task(cj)) {
const long long gcount_i = ci->grav.count;
const long long gcount_j = cj->grav.count;
/* Replace by a single sub-task? */
if (scheduler_dosub &&
gcount_i * gcount_j < ((long long)space_subsize_pair_grav)) {
/* Otherwise, split it. */
} else {
/* Turn the task into a M-M task that will take care of all the
* progeny pairs */
t->type = task_type_grav_mm;
t->subtype = task_subtype_none;
t->flags = 0;
/* Make a task for every other pair of progeny */
for (int i = 0; i < 8; i++) {
if (ci->progeny[i] != NULL) {
for (int j = 0; j < 8; j++) {
if (cj->progeny[j] != NULL) {
/* Can we use a M-M interaction here? */
if (cell_can_use_pair_mm_rebuild(ci->progeny[i],
cj->progeny[j], e, sp)) {
/* Flag this pair as being treated by the M-M task.
* We use the 64 bits in the task->flags field to store
* this information. The corresponding taks will unpack
* the information and operate according to the choices
* made here. */
const int flag = i * 8 + j;
t->flags |= (1ULL << flag);
} else {
/* Ok, we actually have to create a task */
scheduler_splittask_gravity(
scheduler_addtask(s, task_type_pair, task_subtype_grav,
0, 0, ci->progeny[i], cj->progeny[j]),
s);
}
}
}
}
}
/* Can none of the progenies use M-M calculations? */
if (t->flags == 0) {
t->type = task_type_none;
t->subtype = task_subtype_none;
t->ci = NULL;
t->cj = NULL;
t->skip = 1;
}
} /* Split the pair */
}
} /* pair interaction? */
} /* iterate over the current task. */
}
/**
* @brief Split a FOF task if too large.
*
* @param t The #task * @param s The #scheduler we are working in.
*/
static void scheduler_splittask_fof(struct task *t, struct scheduler *s) {
/* Iterate on this task until we're done with it. */
int redo = 1;
while (redo) {
/* Reset the redo flag. */
redo = 0;
/* Non-splittable task? */
if ((t->ci == NULL) || (t->type == task_type_fof_pair && t->cj == NULL) ||
t->ci->grav.count == 0 || (t->cj != NULL && t->cj->grav.count == 0)) {
t->type = task_type_none;
t->subtype = task_subtype_none;
t->ci = NULL;
t->cj = NULL;
t->skip = 1;
break;
}
/* Self-interaction? */
if (t->type == task_type_fof_self) {
/* Get a handle on the cell involved. */
struct cell *ci = t->ci;
/* Foreign task? */
if (ci->nodeID != s->nodeID) {
t->skip = 1;
break;
}
/* Is this cell even split? */
if (cell_can_split_self_fof_task(ci)) {
/* Take a step back (we're going to recycle the current task)... */
redo = 1;
/* Add the self tasks. */
int first_child = 0;
while (ci->progeny[first_child] == NULL) first_child++;
t->ci = ci->progeny[first_child];
for (int k = first_child + 1; k < 8; k++)
if (ci->progeny[k] != NULL && ci->progeny[k]->grav.count)
scheduler_splittask_fof(
scheduler_addtask(s, task_type_fof_self, t->subtype, 0, 0,
ci->progeny[k], NULL),
s);
/* Make a task for each pair of progeny */
for (int j = 0; j < 8; j++)
if (ci->progeny[j] != NULL && ci->progeny[j]->grav.count)
for (int k = j + 1; k < 8; k++)
if (ci->progeny[k] != NULL && ci->progeny[k]->grav.count)
scheduler_splittask_fof(
scheduler_addtask(s, task_type_fof_pair, t->subtype, 0, 0,
ci->progeny[j], ci->progeny[k]),
s);
} /* Cell is split */
} /* Self interaction */
} /* iterate over the current task. */
}
/**
* @brief Mapper function to split FOF tasks that may be too large.
* * @param map_data the tasks to process
* @param num_elements the number of tasks.
* @param extra_data The #scheduler we are working in.
*/
void scheduler_splittasks_fof_mapper(void *map_data, int num_elements,
void *extra_data) {
/* Extract the parameters. */
struct scheduler *s = (struct scheduler *)extra_data;
struct task *tasks = (struct task *)map_data;
for (int ind = 0; ind < num_elements; ind++) {
struct task *t = &tasks[ind];
/* Invoke the correct splitting strategy */
if (t->type == task_type_fof_self || t->type == task_type_fof_pair) {
scheduler_splittask_fof(t, s);
}
}
}
/**
* @brief Mapper function to split non-FOF tasks that may be too large.
*
* @param map_data the tasks to process
* @param num_elements the number of tasks.
* @param extra_data The #scheduler we are working in.
*/
void scheduler_splittasks_mapper(void *map_data, int num_elements,
void *extra_data) {
/* Extract the parameters. */
struct scheduler *s = (struct scheduler *)extra_data;
struct task *tasks = (struct task *)map_data;
for (int ind = 0; ind < num_elements; ind++) {
struct task *t = &tasks[ind];
/* Invoke the correct splitting strategy */
if (t->subtype == task_subtype_density) {
scheduler_splittask_hydro(t, s);
} else if (t->subtype == task_subtype_external_grav) {
scheduler_splittask_gravity(t, s);
} else if (t->subtype == task_subtype_grav) {
scheduler_splittask_gravity(t, s);
} else if (t->type == task_type_grav_mesh) {
/* For future use */
} else {
#ifdef SWIFT_DEBUG_CHECKS
error("Unexpected task sub-type %s/%s", taskID_names[t->type],
subtaskID_names[t->subtype]);
#endif
}
}
}
/**
* @brief Splits all the tasks in the scheduler that are too large.
*
* @param s The #scheduler.
* @param fof_tasks Are we splitting the FOF tasks (1)? Or the regular tasks
* (0)?
* @param verbose Are we talkative?
*/
void scheduler_splittasks(struct scheduler *s, const int fof_tasks,
const int verbose) {
if (verbose) {
message("space_subsize_self_hydro= %d", space_subsize_self_hydro);
message("space_subsize_pair_hydro= %d", space_subsize_pair_hydro);
message("space_subsize_self_stars= %d", space_subsize_self_stars);
message("space_subsize_pair_stars= %d", space_subsize_pair_stars); message("space_subsize_self_grav= %d", space_subsize_self_grav);
message("space_subsize_pair_grav= %d", space_subsize_pair_grav);
}
if (fof_tasks) {
/* Call the mapper on each current task. */
threadpool_map(s->threadpool, scheduler_splittasks_fof_mapper, s->tasks,
s->nr_tasks, sizeof(struct task), 0, s);
} else {
/* Call the mapper on each current task. */
threadpool_map(s->threadpool, scheduler_splittasks_mapper, s->tasks,
s->nr_tasks, sizeof(struct task), 0, s);
}
}
/**
* @brief Add a #task to the #scheduler.
*
* @param s The #scheduler we are working in.
* @param type The type of the task.
* @param subtype The sub-type of the task.
* @param flags The flags of the task.
* @param implicit If true, only use this task to unlock dependencies, i.e.
* this task is never enqueued.
* @param ci The first cell to interact.
* @param cj The second cell to interact.
*/
struct task *scheduler_addtask(struct scheduler *s, enum task_types type,
enum task_subtypes subtype, int flags,
int implicit, struct cell *ci, struct cell *cj) {
/* Get the next free task. */
const int ind = atomic_inc(&s->tasks_next);
/* Overflow? */
if (ind >= s->size)
error(
"Task list overflow (%d). Need to increase "
"Scheduler:tasks_per_cell.",
ind);
/* Get a pointer to the new task. */
struct task *t = &s->tasks[ind];
/* Copy the data. */
t->type = type;
t->subtype = subtype;
t->flags = flags;
t->wait = 0;
t->ci = ci;
t->cj = cj;
t->skip = 1; /* Mark tasks as skip by default. */
t->implicit = implicit;
t->weight = 0;
t->rank = 0;
t->nr_unlock_tasks = 0;
#ifdef SWIFT_DEBUG_TASKS
t->rid = -1;
#endif
t->tic = 0;
t->toc = 0;
if (ci != NULL) cell_set_flag(ci, cell_flag_has_tasks);
if (cj != NULL) cell_set_flag(cj, cell_flag_has_tasks);
/* Add an index for it. */
// lock_lock( &s->lock );
s->tasks_ind[atomic_inc(&s->nr_tasks)] = ind;
// lock_unlock_blind( &s->lock );
/* Return a pointer to the new task. */
return t;
}
/**
* @brief Set the unlock pointers in each task.
*
* @param s The #scheduler.
*/
void scheduler_set_unlocks(struct scheduler *s) {
/* Store the counts for each task. */
short int *counts;
if ((counts = (short int *)swift_malloc(
"counts", sizeof(short int) * s->nr_tasks)) == NULL)
error("Failed to allocate temporary counts array.");
bzero(counts, sizeof(short int) * s->nr_tasks);
for (int k = 0; k < s->nr_unlocks; k++) {
counts[s->unlock_ind[k]] += 1;
/* Check that we are not overflowing */
if (counts[s->unlock_ind[k]] < 0)
error(
"Task (type=%s/%s) unlocking more than %lld other tasks!\n"
"This likely a result of having tasks at vastly different levels"
"in the tree.\nYou may want to play with the 'Scheduler' "
"parameters to modify the task splitting strategy and reduce"
"the difference in task depths.",
taskID_names[s->tasks[s->unlock_ind[k]].type],
subtaskID_names[s->tasks[s->unlock_ind[k]].subtype],
(1LL << (8 * sizeof(short int) - 1)) - 1);
}
/* Compute the offset for each unlock block. */
int *offsets;
if ((offsets = (int *)swift_malloc("offsets",
sizeof(int) * (s->nr_tasks + 1))) == NULL)
error("Failed to allocate temporary offsets array.");
offsets[0] = 0;
for (int k = 0; k < s->nr_tasks; k++) {
offsets[k + 1] = offsets[k] + counts[k];
#ifdef SWIFT_DEBUG_CHECKS
/* Check that we are not overflowing */
if (offsets[k + 1] < 0) error("Task unlock offset array overflowing");
#endif
}
/* Create and fill a temporary array with the sorted unlocks. */
struct task **unlocks;
if ((unlocks = (struct task **)swift_malloc(
"unlocks", sizeof(struct task *) * s->size_unlocks)) == NULL)
error("Failed to allocate temporary unlocks array.");
for (int k = 0; k < s->nr_unlocks; k++) {
const int ind = s->unlock_ind[k];
unlocks[offsets[ind]] = s->unlocks[k];
offsets[ind] += 1;
}
/* Swap the unlocks. */
swift_free("unlocks", s->unlocks);
s->unlocks = unlocks;
/* Re-set the offsets. */
offsets[0] = 0;
for (int k = 1; k < s->nr_tasks; k++)
offsets[k] = offsets[k - 1] + counts[k - 1];
/* Set the unlocks in the tasks. */
for (int k = 0; k < s->nr_tasks; k++) {
struct task *t = &s->tasks[k]; t->nr_unlock_tasks = counts[k];
t->unlock_tasks = &s->unlocks[offsets[k]];
}
#ifdef SWIFT_DEBUG_CHECKS
/* Verify that there are no duplicate unlocks. */
for (int k = 0; k < s->nr_tasks; k++) {
struct task *t = &s->tasks[k];
for (int i = 0; i < t->nr_unlock_tasks; i++) {
for (int j = i + 1; j < t->nr_unlock_tasks; j++) {
if (t->unlock_tasks[i] == t->unlock_tasks[j])
error("duplicate unlock! t->type=%s/%s unlocking type=%s/%s",
taskID_names[t->type], subtaskID_names[t->subtype],
taskID_names[t->unlock_tasks[i]->type],
subtaskID_names[t->unlock_tasks[i]->subtype]);
}
}
}
#endif
/* Clean up. */
swift_free("counts", counts);
swift_free("offsets", offsets);
}
/**
* @brief Sort the tasks in topological order over all queues.
*
* @param s The #scheduler.
*/
void scheduler_ranktasks(struct scheduler *s) {
struct task *tasks = s->tasks;
int *tid = s->tasks_ind;
const int nr_tasks = s->nr_tasks;
/* Run through the tasks and get all the waits right. */
for (int i = 0; i < nr_tasks; i++) {
struct task *t = &tasks[i];
// Increment the waits of the dependances
for (int k = 0; k < t->nr_unlock_tasks; k++) {
t->unlock_tasks[k]->wait++;
}
}
/* Load the tids of tasks with no waits. */
int left = 0;
for (int k = 0; k < nr_tasks; k++)
if (tasks[k].wait == 0) {
tid[left] = k;
left += 1;
}
/* Main loop. */
for (int j = 0, rank = 0; j < nr_tasks; rank++) {
/* Did we get anything? */
if (j == left) error("Unsatisfiable task dependencies detected.");
/* Unlock the next layer of tasks. */
const int left_old = left;
for (; j < left_old; j++) {
struct task *t = &tasks[tid[j]];
t->rank = rank;
/* message( "task %i of type %s has rank %i." , i ,
(t->type == task_type_self) ? "self" : (t->type == task_type_pair) ?
"pair" : "sort" , rank ); */
for (int k = 0; k < t->nr_unlock_tasks; k++) {
struct task *u = t->unlock_tasks[k];
if (--u->wait == 0) {
tid[left] = u - tasks; left += 1;
}
}
}
/* Move back to the old left (like Sanders!). */
j = left_old;
}
#ifdef SWIFT_DEBUG_CHECKS
/* Verify that the tasks were ranked correctly. */
for (int k = 1; k < s->nr_tasks; k++)
if (tasks[tid[k - 1]].rank > tasks[tid[k]].rank)
error("Task ranking failed.");
#endif
}
/**
* @brief (Re)allocate the task arrays.
*
* @param s The #scheduler.
* @param size The maximum number of tasks in the #scheduler.
*/
void scheduler_reset(struct scheduler *s, int size) {
/* Do we need to re-allocate? */
if (size > s->size) {
/* Free existing task lists if necessary. */
scheduler_free_tasks(s);
/* Allocate the new lists. */
if (swift_memalign("tasks", (void **)&s->tasks, task_align,
size * sizeof(struct task)) != 0)
error("Failed to allocate task array.");
if ((s->tasks_ind = (int *)swift_malloc("tasks_ind", sizeof(int) * size)) ==
NULL)
error("Failed to allocate task lists.");
if ((s->tid_active =
(int *)swift_malloc("tid_active", sizeof(int) * size)) == NULL)
error("Failed to allocate aactive task lists.");
}
/* Reset the counters. */
s->size = size;
s->nr_tasks = 0;
s->tasks_next = 0;
s->waiting = 0;
s->nr_unlocks = 0;
s->completed_unlock_writes = 0;
s->active_count = 0;
/* Set the task pointers in the queues. */
for (int k = 0; k < s->nr_queues; k++) s->queues[k].tasks = s->tasks;
}
/**
* @brief Compute the task weights
*
* @param s The #scheduler.
* @param verbose Are we talkative?
*/
void scheduler_reweight(struct scheduler *s, int verbose) {
const int nr_tasks = s->nr_tasks;
int *tid = s->tasks_ind;
struct task *tasks = s->tasks;
const int nodeID = s->nodeID;
const float wscale = 0.001f;
const ticks tic = getticks();
/* Run through the tasks backwards and set their weights. */
for (int k = nr_tasks - 1; k >= 0; k--) {
struct task *t = &tasks[tid[k]];
float cost = 0.f;
t->weight = 0.f;
for (int j = 0; j < t->nr_unlock_tasks; j++)
if (t->unlock_tasks[j]->weight > t->weight)
t->weight = t->unlock_tasks[j]->weight;
const float count_i = (t->ci != NULL) ? t->ci->hydro.count : 0.f;
const float count_j = (t->cj != NULL) ? t->cj->hydro.count : 0.f;
const float gcount_i = (t->ci != NULL) ? t->ci->grav.count : 0.f;
const float gcount_j = (t->cj != NULL) ? t->cj->grav.count : 0.f;
const float scount_i = (t->ci != NULL) ? t->ci->stars.count : 0.f;
const float scount_j = (t->cj != NULL) ? t->cj->stars.count : 0.f;
const float bcount_i = (t->ci != NULL) ? t->ci->black_holes.count : 0.f;
const float bcount_j = (t->cj != NULL) ? t->cj->black_holes.count : 0.f;
switch (t->type) {
case task_type_sort:
cost = wscale * intrinsics_popcount(t->flags) * count_i *
(sizeof(int) * 8 - intrinsics_clz(t->ci->hydro.count));
break;
case task_type_stars_sort:
cost = wscale * intrinsics_popcount(t->flags) * scount_i *
(sizeof(int) * 8 - intrinsics_clz(t->ci->stars.count));
break;
case task_type_self:
if (t->subtype == task_subtype_grav) {
cost = 1.f * (wscale * gcount_i) * gcount_i;
} else if (t->subtype == task_subtype_external_grav)
cost = 1.f * wscale * gcount_i;
else if (t->subtype == task_subtype_stars_density)
cost = 1.f * wscale * scount_i * count_i;
else if (t->subtype == task_subtype_stars_feedback)
cost = 1.f * wscale * scount_i * count_i;
else if (t->subtype == task_subtype_bh_density)
cost = 1.f * wscale * bcount_i * count_i;
else if (t->subtype == task_subtype_bh_feedback)
cost = 1.f * wscale * bcount_i * count_i;
else // hydro loops
cost = 1.f * (wscale * count_i) * count_i;
break;
case task_type_pair:
if (t->subtype == task_subtype_grav) {
if (t->ci->nodeID != nodeID || t->cj->nodeID != nodeID)
cost = 3.f * (wscale * gcount_i) * gcount_j;
else
cost = 2.f * (wscale * gcount_i) * gcount_j;
} else if (t->subtype == task_subtype_stars_density ||
t->subtype == task_subtype_stars_feedback) {
if (t->ci->nodeID != nodeID)
cost = 3.f * wscale * count_i * scount_j * sid_scale[t->flags];
else if (t->cj->nodeID != nodeID)
cost = 3.f * wscale * scount_i * count_j * sid_scale[t->flags];
else
cost = 2.f * wscale * (scount_i * count_j + scount_j * count_i) *
sid_scale[t->flags];
} else if (t->subtype == task_subtype_bh_density ||
t->subtype == task_subtype_bh_feedback) {
if (t->ci->nodeID != nodeID)
cost = 3.f * wscale * count_i * bcount_j * sid_scale[t->flags];
else if (t->cj->nodeID != nodeID) cost = 3.f * wscale * bcount_i * count_j * sid_scale[t->flags];
else
cost = 2.f * wscale * (bcount_i * count_j + bcount_j * count_i) *
sid_scale[t->flags];
} else { // hydro loops
if (t->ci->nodeID != nodeID || t->cj->nodeID != nodeID)
cost = 3.f * (wscale * count_i) * count_j * sid_scale[t->flags];
else
cost = 2.f * (wscale * count_i) * count_j * sid_scale[t->flags];
}
break;
case task_type_sub_pair:
#ifdef SWIFT_DEBUG_CHECKS
if (t->flags < 0) error("Negative flag value!");
#endif
if (t->subtype == task_subtype_stars_density ||
t->subtype == task_subtype_stars_feedback) {
if (t->ci->nodeID != nodeID) {
cost = 3.f * (wscale * count_i) * scount_j * sid_scale[t->flags];
} else if (t->cj->nodeID != nodeID) {
cost = 3.f * (wscale * scount_i) * count_j * sid_scale[t->flags];
} else {
cost = 2.f * wscale * (scount_i * count_j + scount_j * count_i) *
sid_scale[t->flags];
}
} else if (t->subtype == task_subtype_bh_density ||
t->subtype == task_subtype_bh_feedback) {
if (t->ci->nodeID != nodeID) {
cost = 3.f * (wscale * count_i) * bcount_j * sid_scale[t->flags];
} else if (t->cj->nodeID != nodeID) {
cost = 3.f * (wscale * bcount_i) * count_j * sid_scale[t->flags];
} else {
cost = 2.f * wscale * (bcount_i * count_j + bcount_j * count_i) *
sid_scale[t->flags];
}
} else { // hydro loops
if (t->ci->nodeID != nodeID || t->cj->nodeID != nodeID) {
cost = 3.f * (wscale * count_i) * count_j * sid_scale[t->flags];
} else {
cost = 2.f * (wscale * count_i) * count_j * sid_scale[t->flags];
}
}
break;
case task_type_sub_self:
if (t->subtype == task_subtype_stars_density) {
cost = 1.f * (wscale * scount_i) * count_i;
} else if (t->subtype == task_subtype_stars_feedback) {
cost = 1.f * (wscale * scount_i) * count_i;
} else if (t->subtype == task_subtype_bh_density) {
cost = 1.f * (wscale * bcount_i) * count_i;
} else if (t->subtype == task_subtype_bh_feedback) {
cost = 1.f * (wscale * bcount_i) * count_i;
} else {
cost = 1.f * (wscale * count_i) * count_i;
}
break;
case task_type_ghost:
if (t->ci == t->ci->hydro.super) cost = wscale * count_i;
break;
case task_type_extra_ghost:
if (t->ci == t->ci->hydro.super) cost = wscale * count_i;
break;
case task_type_stars_ghost:
if (t->ci == t->ci->hydro.super) cost = wscale * scount_i;
break; case task_type_bh_density_ghost:
if (t->ci == t->ci->hydro.super) cost = wscale * bcount_i;
break;
case task_type_drift_part:
cost = wscale * count_i;
break;
case task_type_drift_gpart:
cost = wscale * gcount_i;
break;
case task_type_drift_spart:
cost = wscale * scount_i;
break;
case task_type_drift_bpart:
cost = wscale * bcount_i;
break;
case task_type_init_grav:
cost = wscale * gcount_i;
break;
case task_type_grav_down:
cost = wscale * gcount_i;
break;
case task_type_grav_long_range:
cost = wscale * gcount_i;
break;
case task_type_grav_mm:
cost = wscale * (gcount_i + gcount_j);
break;
case task_type_end_hydro_force:
cost = wscale * count_i;
break;
case task_type_end_grav_force:
cost = wscale * gcount_i;
break;
case task_type_cooling:
cost = wscale * count_i;
break;
case task_type_star_formation:
cost = wscale * (count_i + scount_i);
break;
case task_type_kick1:
cost = wscale * (count_i + gcount_i + scount_i + bcount_i);
break;
case task_type_kick2:
cost = wscale * (count_i + gcount_i + scount_i + bcount_i);
break;
case task_type_timestep:
cost = wscale * (count_i + gcount_i + scount_i + bcount_i);
break;
case task_type_send:
if (count_i < 1e5)
cost = 10.f * (wscale * count_i) * count_i;
else
cost = 2e9;
break;
case task_type_recv:
if (count_i < 1e5)
cost = 5.f * (wscale * count_i) * count_i;
else
cost = 1e9;
break;
default:
cost = 0;
break;
}
t->weight += cost;
}
if (verbose)
message("took %.3f %s.", clocks_from_ticks(getticks() - tic),
clocks_getunit());
/* int min = tasks[0].weight, max = tasks[0].weight;
for ( int k = 1 ; k < nr_tasks ; k++ )
if ( tasks[k].weight < min )
min = tasks[k].weight;
else if ( tasks[k].weight > max )
max = tasks[k].weight;
message( "task weights are in [ %i , %i ]." , min , max ); */
}
/**
* @brief #threadpool_map function which runs through the task
* graph and re-computes the task wait counters.
*/
void scheduler_rewait_mapper(void *map_data, int num_elements,
void *extra_data) {
struct scheduler *s = (struct scheduler *)extra_data;
const int *tid = (int *)map_data;
for (int ind = 0; ind < num_elements; ind++) {
struct task *t = &s->tasks[tid[ind]];
/* Ignore skipped tasks. */
if (t->skip) continue;
/* Increment the task's own wait counter for the enqueueing. */
atomic_inc(&t->wait);
#ifdef SWIFT_DEBUG_CHECKS
/* Check that we don't have more waits that what can be stored. */
if (t->wait < 0)
error("Task (type=%s/%s) unlocked by more than %lld tasks!",
taskID_names[t->type], subtaskID_names[t->subtype],
(1LL << (8 * sizeof(t->wait) - 1)) - 1);
#endif
/* Sets the waits of the dependances */
for (int k = 0; k < t->nr_unlock_tasks; k++) {
struct task *u = t->unlock_tasks[k];
atomic_inc(&u->wait);
}
}
}
void scheduler_enqueue_mapper(void *map_data, int num_elements,
void *extra_data) {
struct scheduler *s = (struct scheduler *)extra_data;
const int *tid = (int *)map_data;
struct task *tasks = s->tasks;
for (int ind = 0; ind < num_elements; ind++) {
struct task *t = &tasks[tid[ind]];
if (atomic_dec(&t->wait) == 1 && !t->skip) {
scheduler_enqueue(s, t);
}
}
pthread_cond_broadcast(&s->sleep_cond);
}
/**
* @brief Start the scheduler, i.e. fill the queues with ready tasks.
*
* @param s The #scheduler.
*/
void scheduler_start(struct scheduler *s) {
/* Re-wait the tasks. */
if (s->active_count > 1000) {
threadpool_map(s->threadpool, scheduler_rewait_mapper, s->tid_active,
s->active_count, sizeof(int), 0, s);
} else { scheduler_rewait_mapper(s->tid_active, s->active_count, s);
}
/* Loop over the tasks and enqueue whoever is ready. */
if (s->active_count > 1000) {
threadpool_map(s->threadpool, scheduler_enqueue_mapper, s->tid_active,
s->active_count, sizeof(int), 0, s);
} else {
scheduler_enqueue_mapper(s->tid_active, s->active_count, s);
}
/* Clear the list of active tasks. */
s->active_count = 0;
/* To be safe, fire of one last sleep_cond in a safe way. */
pthread_mutex_lock(&s->sleep_mutex);
pthread_cond_broadcast(&s->sleep_cond);
pthread_mutex_unlock(&s->sleep_mutex);
}
/**
* @brief Put a task on one of the queues.
*
* @param s The #scheduler.
* @param t The #task.
*/
void scheduler_enqueue(struct scheduler *s, struct task *t) {
/* The target queue for this task. */
int qid = -1;
/* Ignore skipped tasks */
if (t->skip) return;
/* If this is an implicit task, just pretend it's done. */
if (t->implicit) {
#ifdef SWIFT_DEBUG_CHECKS
t->ti_run = s->space->e->ti_current;
#endif
t->skip = 1;
for (int j = 0; j < t->nr_unlock_tasks; j++) {
struct task *t2 = t->unlock_tasks[j];
if (atomic_dec(&t2->wait) == 1) scheduler_enqueue(s, t2);
}
}
/* Otherwise, look for a suitable queue. */
else {
#ifdef WITH_MPI
int err = MPI_SUCCESS;
#endif
/* Find the previous owner for each task type, and do
any pre-processing needed. */
switch (t->type) {
case task_type_self:
case task_type_sub_self:
if (t->subtype == task_subtype_grav ||
t->subtype == task_subtype_external_grav)
qid = t->ci->grav.super->owner;
else
qid = t->ci->hydro.super->owner;
break;
case task_type_sort:
case task_type_ghost:
case task_type_drift_part:
qid = t->ci->hydro.super->owner;
break;
case task_type_drift_gpart:
qid = t->ci->grav.super->owner;
break; case task_type_kick1:
case task_type_kick2:
case task_type_stars_ghost:
case task_type_logger:
case task_type_stars_sort:
case task_type_timestep:
qid = t->ci->super->owner;
break;
case task_type_pair:
case task_type_sub_pair:
qid = t->ci->super->owner;
if (qid < 0 ||
s->queues[qid].count > s->queues[t->cj->super->owner].count)
qid = t->cj->super->owner;
break;
case task_type_recv:
#ifdef WITH_MPI
{
size_t size = 0; /* Size in bytes. */
size_t count = 0; /* Number of elements to receive */
MPI_Datatype type = MPI_BYTE; /* Type of the elements */
void *buff = NULL; /* Buffer to accept elements */
if (t->subtype == task_subtype_tend_part) {
count = size =
t->ci->mpi.pcell_size * sizeof(struct pcell_step_hydro);
buff = t->buff = malloc(count);
} else if (t->subtype == task_subtype_tend_gpart) {
count = size = t->ci->mpi.pcell_size * sizeof(struct pcell_step_grav);
buff = t->buff = malloc(count);
} else if (t->subtype == task_subtype_tend_spart) {
count = size =
t->ci->mpi.pcell_size * sizeof(struct pcell_step_stars);
buff = t->buff = malloc(count);
} else if (t->subtype == task_subtype_tend_bpart) {
count = size =
t->ci->mpi.pcell_size * sizeof(struct pcell_step_black_holes);
buff = t->buff = malloc(count);
} else if (t->subtype == task_subtype_part_swallow) {
count = size =
t->ci->hydro.count * sizeof(struct black_holes_part_data);
buff = t->buff = malloc(count);
} else if (t->subtype == task_subtype_bpart_merger) {
count = size =
sizeof(struct black_holes_bpart_data) * t->ci->black_holes.count;
buff = t->buff = malloc(count);
} else if (t->subtype == task_subtype_xv ||
t->subtype == task_subtype_rho ||
t->subtype == task_subtype_gradient) {
count = t->ci->hydro.count;
size = count * sizeof(struct part);
type = part_mpi_type;
buff = t->ci->hydro.parts;
} else if (t->subtype == task_subtype_gpart) {
count = t->ci->grav.count;
size = count * sizeof(struct gpart); type = gpart_mpi_type;
buff = t->ci->grav.parts;
} else if (t->subtype == task_subtype_spart) {
count = t->ci->stars.count;
size = count * sizeof(struct spart);
type = spart_mpi_type;
buff = t->ci->stars.parts;
} else if (t->subtype == task_subtype_bpart_rho ||
t->subtype == task_subtype_bpart_swallow ||
t->subtype == task_subtype_bpart_feedback) {
count = t->ci->black_holes.count;
size = count * sizeof(struct bpart);
type = bpart_mpi_type;
buff = t->ci->black_holes.parts;
} else if (t->subtype == task_subtype_multipole) {
count = t->ci->mpi.pcell_size;
size = count * sizeof(struct gravity_tensors);
type = multipole_mpi_type;
buff = t->buff = malloc(size);
} else if (t->subtype == task_subtype_sf_counts) {
count = size = t->ci->mpi.pcell_size * sizeof(struct pcell_sf);
buff = t->buff = malloc(count);
} else {
error("Unknown communication sub-type");
}
err = MPI_Irecv(buff, count, type, t->ci->nodeID, t->flags,
subtaskMPI_comms[t->subtype], &t->req);
if (err != MPI_SUCCESS) {
mpi_error(err, "Failed to emit irecv for particle data.");
}
/* And log, if logging enabled. */
mpiuse_log_allocation(t->type, t->subtype, &t->req, 1, size,
t->ci->nodeID, t->flags);
qid = 1 % s->nr_queues;
}
#else
error("SWIFT was not compiled with MPI support.");
#endif
break;
case task_type_send:
#ifdef WITH_MPI
{
size_t size = 0; /* Size in bytes. */
size_t count = 0; /* Number of elements to send */
MPI_Datatype type = MPI_BYTE; /* Type of the elements */
void *buff = NULL; /* Buffer to send */
if (t->subtype == task_subtype_tend_part) {
size = count =
t->ci->mpi.pcell_size * sizeof(struct pcell_step_hydro);
buff = t->buff = malloc(size);
cell_pack_end_step_hydro(t->ci, (struct pcell_step_hydro *)buff);
} else if (t->subtype == task_subtype_tend_gpart) {
size = count = t->ci->mpi.pcell_size * sizeof(struct pcell_step_grav); buff = t->buff = malloc(size);
cell_pack_end_step_grav(t->ci, (struct pcell_step_grav *)buff);
} else if (t->subtype == task_subtype_tend_spart) {
size = count =
t->ci->mpi.pcell_size * sizeof(struct pcell_step_stars);
buff = t->buff = malloc(size);
cell_pack_end_step_stars(t->ci, (struct pcell_step_stars *)buff);
} else if (t->subtype == task_subtype_tend_bpart) {
size = count =
t->ci->mpi.pcell_size * sizeof(struct pcell_step_black_holes);
buff = t->buff = malloc(size);
cell_pack_end_step_black_holes(t->ci,
(struct pcell_step_black_holes *)buff);
} else if (t->subtype == task_subtype_part_swallow) {
size = count =
t->ci->hydro.count * sizeof(struct black_holes_part_data);
buff = t->buff = malloc(size);
cell_pack_part_swallow(t->ci, (struct black_holes_part_data *)buff);
} else if (t->subtype == task_subtype_bpart_merger) {
size = count =
sizeof(struct black_holes_bpart_data) * t->ci->black_holes.count;
buff = t->buff = malloc(size);
cell_pack_bpart_swallow(t->ci,
(struct black_holes_bpart_data *)t->buff);
} else if (t->subtype == task_subtype_xv ||
t->subtype == task_subtype_rho ||
t->subtype == task_subtype_gradient) {
count = t->ci->hydro.count;
size = count * sizeof(struct part);
type = part_mpi_type;
buff = t->ci->hydro.parts;
} else if (t->subtype == task_subtype_gpart) {
count = t->ci->grav.count;
size = count * sizeof(struct gpart);
type = gpart_mpi_type;
buff = t->ci->grav.parts;
} else if (t->subtype == task_subtype_spart) {
count = t->ci->stars.count;
size = count * sizeof(struct spart);
type = spart_mpi_type;
buff = t->ci->stars.parts;
} else if (t->subtype == task_subtype_bpart_rho ||
t->subtype == task_subtype_bpart_swallow ||
t->subtype == task_subtype_bpart_feedback) {
count = t->ci->black_holes.count;
size = count * sizeof(struct bpart);
type = bpart_mpi_type;
buff = t->ci->black_holes.parts;
} else if (t->subtype == task_subtype_multipole) {
count = t->ci->mpi.pcell_size;
size = count * sizeof(struct gravity_tensors);
type = multipole_mpi_type; buff = t->buff = malloc(size);
cell_pack_multipoles(t->ci, (struct gravity_tensors *)buff);
} else if (t->subtype == task_subtype_sf_counts) {
size = count = t->ci->mpi.pcell_size * sizeof(struct pcell_sf);
buff = t->buff = malloc(size);
cell_pack_sf_counts(t->ci, (struct pcell_sf *)t->buff);
} else {
error("Unknown communication sub-type");
}
if (size > s->mpi_message_limit) {
err = MPI_Isend(buff, count, type, t->cj->nodeID, t->flags,
subtaskMPI_comms[t->subtype], &t->req);
} else {
err = MPI_Issend(buff, count, type, t->cj->nodeID, t->flags,
subtaskMPI_comms[t->subtype], &t->req);
}
if (err != MPI_SUCCESS) {
mpi_error(err, "Failed to emit isend for particle data.");
}
/* And log, if logging enabled. */
mpiuse_log_allocation(t->type, t->subtype, &t->req, 1, size,
t->cj->nodeID, t->flags);
qid = 0;
}
#else
error("SWIFT was not compiled with MPI support.");
#endif
break;
default:
qid = -1;
}
if (qid >= s->nr_queues) error("Bad computed qid.");
/* If no previous owner, pick a random queue. */
/* Note that getticks() is random enough */
if (qid < 0) qid = getticks() % s->nr_queues;
/* Increase the waiting counter. */
atomic_inc(&s->waiting);
/* Insert the task into that queue. */
queue_insert(&s->queues[qid], t);
}
}
/**
* @brief Take care of a tasks dependencies.
*
* @param s The #scheduler.
* @param t The finished #task.
*
* @return A pointer to the next task, if a suitable one has
* been identified.
*/
struct task *scheduler_done(struct scheduler *s, struct task *t) {
/* Release whatever locks this task held. */
if (!t->implicit) task_unlock(t);
/* Loop through the dependencies and add them to a queue if
they are ready. */
for (int k = 0; k < t->nr_unlock_tasks; k++) {
struct task *t2 = t->unlock_tasks[k]; if (t2->skip) continue;
const int res = atomic_dec(&t2->wait);
if (res < 1) {
error("Negative wait!");
} else if (res == 1) {
scheduler_enqueue(s, t2);
}
}
/* Task definitely done, signal any sleeping runners. */
if (!t->implicit) {
t->toc = getticks();
pthread_mutex_lock(&s->sleep_mutex);
atomic_dec(&s->waiting);
pthread_cond_broadcast(&s->sleep_cond);
pthread_mutex_unlock(&s->sleep_mutex);
}
/* Mark the task as skip. */
t->skip = 1;
/* Return the next best task. Note that we currently do not
implement anything that does this, as getting it to respect
priorities is too tricky and currently unnecessary. */
return NULL;
}
/**
* @brief Resolve a single dependency by hand.
*
* @param s The #scheduler.
* @param t The dependent #task.
*
* @return A pointer to the next task, if a suitable one has
* been identified.
*/
struct task *scheduler_unlock(struct scheduler *s, struct task *t) {
/* Loop through the dependencies and add them to a queue if
they are ready. */
for (int k = 0; k < t->nr_unlock_tasks; k++) {
struct task *t2 = t->unlock_tasks[k];
const int res = atomic_dec(&t2->wait);
if (res < 1) {
error("Negative wait!");
} else if (res == 1) {
scheduler_enqueue(s, t2);
}
}
/* Task definitely done. */
if (!t->implicit) {
t->toc = getticks();
pthread_mutex_lock(&s->sleep_mutex);
atomic_dec(&s->waiting);
pthread_cond_broadcast(&s->sleep_cond);
pthread_mutex_unlock(&s->sleep_mutex);
}
/* Return the next best task. Note that we currently do not
implement anything that does this, as getting it to respect
priorities is too tricky and currently unnecessary. */
return NULL;
}
/**
* @brief Get a task, preferably from the given queue.
*
* @param s The #scheduler.
* @param qid The ID of the preferred #queue. * @param prev the previous task that was run.
*
* @return A pointer to a #task or @c NULL if there are no available tasks.
*/
struct task *scheduler_gettask(struct scheduler *s, int qid,
const struct task *prev) {
struct task *res = NULL;
const int nr_queues = s->nr_queues;
unsigned int seed = qid;
/* Check qid. */
if (qid >= nr_queues || qid < 0) error("Bad queue ID.");
/* Loop as long as there are tasks... */
while (s->waiting > 0 && res == NULL) {
/* Try more than once before sleeping. */
for (int tries = 0; res == NULL && s->waiting && tries < scheduler_maxtries;
tries++) {
/* Try to get a task from the suggested queue. */
if (s->queues[qid].count > 0 || s->queues[qid].count_incoming > 0) {
TIMER_TIC
res = queue_gettask(&s->queues[qid], prev, 0);
TIMER_TOC(timer_qget);
if (res != NULL) break;
}
/* If unsuccessful, try stealing from the other queues. */
if (s->flags & scheduler_flag_steal) {
int count = 0, qids[nr_queues];
for (int k = 0; k < nr_queues; k++)
if (s->queues[k].count > 0 || s->queues[k].count_incoming > 0) {
qids[count++] = k;
}
for (int k = 0; k < scheduler_maxsteal && count > 0; k++) {
const int ind = rand_r(&seed) % count;
TIMER_TIC
res = queue_gettask(&s->queues[qids[ind]], prev, 0);
TIMER_TOC(timer_qsteal);
if (res != NULL)
break;
else
qids[ind] = qids[--count];
}
if (res != NULL) break;
}
}
/* If we failed, take a short nap. */
#ifdef WITH_MPI
if (res == NULL && qid > 1)
#else
if (res == NULL)
#endif
{
pthread_mutex_lock(&s->sleep_mutex);
res = queue_gettask(&s->queues[qid], prev, 1);
if (res == NULL && s->waiting > 0) {
pthread_cond_wait(&s->sleep_cond, &s->sleep_mutex);
}
pthread_mutex_unlock(&s->sleep_mutex);
}
}
/* Start the timer on this task, if we got one. */
if (res != NULL) {
res->tic = getticks();
#ifdef SWIFT_DEBUG_TASKS
res->rid = qid;
#endif
}
/* No milk today. */
return res;
}
/**
* @brief Initialize the #scheduler.
*
* @param s The #scheduler.
* @param space The #space we are working with
* @param nr_tasks The number of tasks to allocate initially.
* @param nr_queues The number of queues in this scheduler.
* @param flags The #scheduler flags.
* @param nodeID The MPI rank
* @param tp Parallel processing threadpool.
*/
void scheduler_init(struct scheduler *s, struct space *space, int nr_tasks,
int nr_queues, unsigned int flags, int nodeID,
struct threadpool *tp) {
/* Init the lock. */
lock_init(&s->lock);
/* Allocate the queues. */
if (swift_memalign("queues", (void **)&s->queues, queue_struct_align,
sizeof(struct queue) * nr_queues) != 0)
error("Failed to allocate queues.");
/* Initialize each queue. */
for (int k = 0; k < nr_queues; k++) queue_init(&s->queues[k], NULL);
/* Init the sleep mutex and cond. */
if (pthread_cond_init(&s->sleep_cond, NULL) != 0 ||
pthread_mutex_init(&s->sleep_mutex, NULL) != 0)
error("Failed to initialize sleep barrier.");
/* Init the unlocks. */
if ((s->unlocks = (struct task **)swift_malloc(
"unlocks", sizeof(struct task *) * scheduler_init_nr_unlocks)) ==
NULL ||
(s->unlock_ind = (int *)swift_malloc(
"unlock_ind", sizeof(int) * scheduler_init_nr_unlocks)) == NULL)
error("Failed to allocate unlocks.");
s->nr_unlocks = 0;
s->size_unlocks = scheduler_init_nr_unlocks;
/* Set the scheduler variables. */
s->nr_queues = nr_queues;
s->flags = flags;
s->space = space;
s->nodeID = nodeID;
s->threadpool = tp;
/* Init the tasks array. */
s->size = 0;
s->tasks = NULL;
s->tasks_ind = NULL;
pthread_key_create(&s->local_seed_pointer, NULL);
scheduler_reset(s, nr_tasks);
}
/**
* @brief Prints the list of tasks to a file
*
* @param s The #scheduler
* @param fileName Name of the file to write to
*/
void scheduler_print_tasks(const struct scheduler *s, const char *fileName) {
const int nr_tasks = s->nr_tasks, *tid = s->tasks_ind;
struct task *t, *tasks = s->tasks;
FILE *file = fopen(fileName, "w");
fprintf(file, "# Rank Name Subname unlocks waits\n");
for (int k = nr_tasks - 1; k >= 0; k--) {
t = &tasks[tid[k]];
if (t->skip) continue;
fprintf(file, "%d %s %s %d %d\n", k, taskID_names[t->type],
subtaskID_names[t->subtype], t->nr_unlock_tasks, t->wait);
}
fclose(file);
}
/**
* @brief Frees up the memory allocated for this #scheduler
*/
void scheduler_clean(struct scheduler *s) {
scheduler_free_tasks(s);
swift_free("unlocks", s->unlocks);
swift_free("unlock_ind", s->unlock_ind);
for (int i = 0; i < s->nr_queues; ++i) queue_clean(&s->queues[i]);
swift_free("queues", s->queues);
}
/**
* @brief Free the task arrays allocated by this #scheduler.
*/
void scheduler_free_tasks(struct scheduler *s) {
if (s->tasks != NULL) {
swift_free("tasks", s->tasks);
s->tasks = NULL;
}
if (s->tasks_ind != NULL) {
swift_free("tasks_ind", s->tasks_ind);
s->tasks_ind = NULL;
}
if (s->tid_active != NULL) {
swift_free("tid_active", s->tid_active);
s->tid_active = NULL;
}
s->size = 0;
}
/**
* @brief write down each task level
*/
void scheduler_write_task_level(const struct scheduler *s) {
/* init */
const int max_depth = 30;
const struct task *tasks = s->tasks;
int nr_tasks = s->nr_tasks;
/* Init counter */
int size = task_type_count * task_subtype_count * max_depth;
int *count = (int *)malloc(size * sizeof(int));
if (count == NULL) error("Failed to allocate memory");
for (int i = 0; i < size; i++) count[i] = 0;
/* Count tasks */
for (int i = 0; i < nr_tasks; i++) {
const struct task *t = &tasks[i];
if (t->ci) {
if ((int)t->ci->depth >= max_depth)
error("Cell is too deep, you need to increase max_depth");
int ind = t->type * task_subtype_count * max_depth;
ind += t->subtype * max_depth;
ind += (int)t->ci->depth;
count[ind] += 1;
}
}
/* Open file */
char filename[200] = "task_level.txt";
FILE *f = fopen(filename, "w");
if (f == NULL) error("Error opening task level file.");
/* Print header */
fprintf(f, "# task_type, task_subtype, depth, count\n");
/* Print tasks level */
for (int i = 0; i < size; i++) {
if (count[i] == 0) continue;
int type = i / (task_subtype_count * max_depth);
int subtype = i - task_subtype_count * max_depth * type;
subtype /= max_depth;
int depth = i - task_subtype_count * max_depth * type;
depth -= subtype * max_depth;
fprintf(f, "%s %s %i %i\n", taskID_names[type], subtaskID_names[subtype],
depth, count[i]);
}
/* clean up */
fclose(f);
free(count);
}