blob: 927a0a8b59fd8b1347927489f5865d3145a704ce [file] [edit]
/*****************************************************************************\
* workerpool.c - worker pool
*****************************************************************************
* Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
*
* This file is part of Slurm, a resource management program.
* For details, see <https://slurm.schedmd.com/>.
* Please also read the included file: DISCLAIMER.
*
* Slurm is free software; you can redistribute it and/or modify it under
* the terms of the GNU General Public License as published by the Free
* Software Foundation; either version 2 of the License, or (at your option)
* any later version.
*
* In addition, as a special exception, the copyright holders give permission
* to link the code of portions of this program with the OpenSSL library under
* certain conditions as described in each individual source file, and
* distribute linked combinations including the two. You must obey the GNU
* General Public License in all respects for all of the code used other than
* OpenSSL. If you modify file(s) with this exception, you may extend this
* exception to your version of the file(s), but you are not obligated to do
* so. If you do not wish to do so, delete this exception statement from your
* version. If you delete this exception statement from all source files in
* the program, then also delete it here.
*
* Slurm 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 General Public License along
* with Slurm; if not, write to the Free Software Foundation, Inc.,
* 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
\*****************************************************************************/
#include "config.h"
#include "src/common/events.h"
#include "src/common/log.h"
#include "src/common/macros.h"
#include "src/common/probes.h"
#include "src/common/read_config.h"
#include "src/common/threadpool.h"
#include "src/common/workerpool.h"
#include "src/common/workq.h"
#include "src/common/xassert.h"
#include "src/common/xmalloc.h"
#include "src/common/xsched.h"
#include "src/common/xstring.h"
#include "src/conmgr/conmgr.h"
/* Limit automatically set default thread count */
#define THREAD_AUTO_MAX 32
/* Threads to create per kernel reported CPU */
#define CPU_THREAD_MULTIPLIER 2
#define CPU_THREAD_HIGH 2
#define CPU_THREAD_LOW 2
/* Count of workerpool_work_t* to preallocate */
#define ALLOC_WORK_COUNT 256
static struct {
pthread_mutex_t mutex;
/* Event when worker finishes */
event_signal_t worker_fini;
/* Event when worker initializes */
event_signal_t worker_init;
/* true if shutdown requested */
bool shutdown;
/* Number of active threads */
int thread_count;
/* Work queue shared by all threads */
workq_t *workq;
/* Allocator for workerpool_work_t* */
workq_allocator_t *alloc;
struct {
int thread_count;
} config;
} workerpool = {
.mutex = PTHREAD_MUTEX_INITIALIZER,
.worker_init = EVENT_INITIALIZER("WORKERPOOL-THREAD-INITIALIZE"),
.worker_fini = EVENT_INITIALIZER("WORKERPOOL-THREAD-FINISH"),
.shutdown = true,
.config = {
.thread_count = 0,
},
};
static void _parse_params(const int default_count, const char *params)
{
char *tmp_str = NULL, *tok = NULL, *saveptr = NULL;
if (default_count > 0)
workerpool.config.thread_count = default_count;
if (!params)
return;
tmp_str = xstrdup(params);
tok = strtok_r(tmp_str, ",", &saveptr);
while (tok) {
/*
* conmgr_threads= is treated as an alias as workerpool_threads=
*/
if (!xstrncasecmp(tok, CONMGR_PARAM_THREADS,
strlen(CONMGR_PARAM_THREADS))) {
const unsigned long count =
slurm_atoul(tok + strlen(CONMGR_PARAM_THREADS));
if (count > WORKERPOOL_THREAD_COUNT_MAX)
fatal("%s: invalid parameter %s",
__func__, tok);
workerpool.config.thread_count = count;
log_flag(THREAD, "%s: %lu threads", __func__, count);
} else if (!xstrncasecmp(tok, WORKERPOOL_PARAM_THREADS,
strlen(WORKERPOOL_PARAM_THREADS))) {
const unsigned long count =
slurm_atoul(tok +
strlen(WORKERPOOL_PARAM_THREADS));
if (count > WORKERPOOL_THREAD_COUNT_MAX)
fatal("%s: invalid parameter %s",
__func__, tok);
workerpool.config.thread_count = count;
log_flag(THREAD, "%s: %lu threads", __func__, count);
} else {
log_flag(THREAD, "%s: workerpool ignoring parameter %s",
__func__, tok);
}
tok = strtok_r(NULL, ",", &saveptr);
}
xfree(tmp_str);
}
static int _detect_cpu_count(void)
{
int count = get_assigned_cpu_count();
log_flag(THREAD, "%s: detected %d CPUs available from kernel",
__func__, count);
return count;
}
static int _resolve_thread_count(int count, const int default_count)
{
const int detected_cpus = _detect_cpu_count();
const int auto_threads_max = (detected_cpus * CPU_THREAD_MULTIPLIER);
const int auto_threads = MIN(THREAD_AUTO_MAX, auto_threads_max);
const int detected_threads_high = (detected_cpus * CPU_THREAD_HIGH);
const int detected_threads_low = (detected_cpus / CPU_THREAD_LOW);
const int warn_max_threads =
MIN(WORKERPOOL_THREAD_COUNT_MAX, detected_threads_high);
const int min_def_threads =
MIN(THREAD_AUTO_MAX,
MAX(WORKERPOOL_THREAD_COUNT_MIN, default_count));
const int warn_min_threads = MIN(detected_threads_low, min_def_threads);
if (!count && (workerpool.config.thread_count > 0)) {
count = workerpool.config.thread_count;
log_flag(THREAD, "%s: Setting thread count to %s%d threads",
__func__, WORKERPOOL_PARAM_THREADS,
workerpool.config.thread_count);
}
if (!count) {
if (default_count > 0) {
count = default_count;
log_flag(THREAD, "%s: Setting thread count to default %d threads",
__func__, default_count);
} else {
count = auto_threads;
log_flag(THREAD, "%s: Setting thread count to %d/%d for %d available CPUs",
__func__, auto_threads, auto_threads_max,
detected_cpus);
}
} else if ((count > warn_max_threads) || (count < warn_min_threads)) {
warning("%s%d is configured outside of the suggested range of [%d, %d] for %d CPUs. Performance will be negatively impacted, potentially causing difficult to debug hangs. Please keep within the suggested range or use the automatically detected thread count of %d threads.",
WORKERPOOL_PARAM_THREADS, count, warn_min_threads,
warn_max_threads, detected_cpus, auto_threads);
}
if (count < WORKERPOOL_THREAD_COUNT_MIN) {
error("%s: %s%d too low, increasing to %d",
__func__, WORKERPOOL_PARAM_THREADS, count,
WORKERPOOL_THREAD_COUNT_MIN);
count = WORKERPOOL_THREAD_COUNT_MIN;
} else if (count > WORKERPOOL_THREAD_COUNT_MAX) {
error("%s: %s%d too high, decreasing to %d",
__func__, WORKERPOOL_PARAM_THREADS, count,
WORKERPOOL_THREAD_COUNT_MAX);
count = WORKERPOOL_THREAD_COUNT_MAX;
}
return count;
}
static void *_worker(void *arg)
{
workq_t *workq = NULL;
slurm_mutex_lock(&workerpool.mutex);
workerpool.thread_count++;
xassert(workerpool.thread_count > 0);
workq = workerpool.workq;
EVENT_BROADCAST(&workerpool.worker_init);
slurm_mutex_unlock(&workerpool.mutex);
workq_run(workq, true);
slurm_mutex_lock(&workerpool.mutex);
xassert(workq == workerpool.workq);
workerpool.thread_count--;
xassert(workerpool.thread_count >= 0);
EVENT_BROADCAST(&workerpool.worker_fini);
slurm_mutex_unlock(&workerpool.mutex);
return NULL;
}
static void _probe_verbose(probe_log_t *log)
{
probe_log(log, "state: shutdown=%c thread_count=%d",
BOOL_CHARIFY(workerpool.shutdown), workerpool.thread_count);
}
static probe_status_t _probe(probe_log_t *log, void *arg)
{
probe_status_t status = PROBE_RC_UNKNOWN;
slurm_mutex_lock(&workerpool.mutex);
if (log)
_probe_verbose(log);
if (workerpool.shutdown)
status = PROBE_RC_ONLINE;
else
status = PROBE_RC_READY;
slurm_mutex_unlock(&workerpool.mutex);
return status;
}
extern void workerpool_init(const int thread_count,
const int default_thread_count, const char *params)
{
int count = -1;
slurm_mutex_lock(&workerpool.mutex);
if (!workerpool.shutdown) {
slurm_mutex_unlock(&workerpool.mutex);
return;
}
xassert(workerpool.thread_count <= 0);
if (!workerpool.workq)
probe_register("workerpool", _probe, NULL);
_parse_params(default_thread_count, params);
workerpool.shutdown = false;
count = _resolve_thread_count(thread_count, default_thread_count);
workerpool.workq = workq_init(workerpool.workq, "workerpool");
workerpool.alloc = workq_allocator(workerpool.workq, ALLOC_WORK_COUNT,
"workerpool");
for (int i = 0; i < count; i++)
slurm_thread_create_detached(NULL, _worker, NULL);
/* Wait for all threads to start up */
while (workerpool.thread_count < count) {
log_flag(THREAD, "%s: waiting for workers %d/%d to initialize",
__func__, workerpool.thread_count, count);
EVENT_WAIT(&workerpool.worker_init, &workerpool.mutex);
}
log_flag(THREAD, "%s: started %d worker threads", __func__, count);
slurm_mutex_unlock(&workerpool.mutex);
}
extern void workerpool_fini(void)
{
slurm_mutex_lock(&workerpool.mutex);
if (workerpool.shutdown) {
slurm_mutex_unlock(&workerpool.mutex);
return;
}
log_flag(THREAD, "%s: shutting down %d worker threads",
__func__, workerpool.thread_count);
workerpool.shutdown = true;
/*
* workq_fini() tells the workq to shutdown and waits for every worker
* to leave workq_run(). The workers then reacquire workerpool.mutex
* (held here) and assert workq == workerpool.workq in _worker(), so
* workerpool.workq must stay valid until they have fully exited.
*
* The probe-named workq is intentionally left allocated by workq_fini()
* (the probe still references it), so keep the pointer here. A later
* workerpool_init() reuses it in place instead of leaking it and
* registering a duplicate "workerpool" probe.
*/
workq_fini(workerpool.workq);
/* Wait for all threads to finish */
while (workerpool.thread_count > 0) {
log_flag(THREAD, "%s: waiting for workers %d to finish",
__func__, workerpool.thread_count);
EVENT_WAIT(&workerpool.worker_fini, &workerpool.mutex);
}
slurm_mutex_unlock(&workerpool.mutex);
}
extern void workerpool_enqueue(workq_allocator_t *alloc,
workq_priority_t priority, work_func_t func,
const char *func_name, void *arg,
const char *caller)
{
slurm_mutex_lock(&workerpool.mutex);
xassert(!workerpool.shutdown);
if (!alloc)
alloc = workerpool.alloc;
workq_enqueue(workerpool.workq, alloc, priority, func, func_name,
caller, arg);
slurm_mutex_unlock(&workerpool.mutex);
}