2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2012, by the GROMACS development team, led by
5 * David van der Spoel, Berk Hess, Erik Lindahl, and including many
6 * others, as listed in the AUTHORS file in the top-level source
7 * directory and at http://www.gromacs.org.
9 * GROMACS is free software; you can redistribute it and/or
10 * modify it under the terms of the GNU Lesser General Public License
11 * as published by the Free Software Foundation; either version 2.1
12 * of the License, or (at your option) any later version.
14 * GROMACS is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
17 * Lesser General Public License for more details.
19 * You should have received a copy of the GNU Lesser General Public
20 * License along with GROMACS; if not, see
21 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
22 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
24 * If you want to redistribute modifications to GROMACS, please
25 * consider that scientific software is very special. Version
26 * control is crucial - bugs must be traceable. We will be happy to
27 * consider code for inclusion in the official distribution, but
28 * derived work must not be called official GROMACS. Details are found
29 * in the README & COPYING files - if they are missing, get the
30 * official version at http://www.gromacs.org.
32 * To help us fund GROMACS development, we humbly ask that you cite
33 * the research papers on the package. Check out http://www.gromacs.org.
38 #if defined(HAVE_SCHED_H) && defined(HAVE_SCHED_GETAFFINITY)
41 #include <sys/syscall.h>
48 #include "types/commrec.h"
49 #include "types/hw_info.h"
50 #include "gmx_cpuid.h"
52 #include "gmx_omp_nthreads.h"
54 #include "md_logging.h"
56 #include "gmx_thread_affinity.h"
58 #include "thread_mpi/threads.h"
62 get_thread_affinity_layout(FILE *fplog,
64 const gmx_hw_info_t * hwinfo,
66 int pin_offset, int * pin_stride,
67 const int **locality_order)
69 int nhwthreads, npkg, ncores, nhwthreads_per_core, rc;
72 const int * hwthread_id;
73 gmx_bool bPickPinStride;
77 gmx_fatal(FARGS, "Negative thread pinning offset requested");
81 gmx_fatal(FARGS, "Negative thread pinning stride requested");
84 rc = gmx_cpuid_topology(hwinfo->cpuid_info, &nhwthreads, &npkg, &ncores,
86 &pkg_id, &core_id, &hwthread_id, locality_order);
90 /* topology information not available or invalid, ignore it */
91 nhwthreads = hwinfo->nthreads_hw_avail;
92 *locality_order = NULL;
96 /* We don't know anything about the hardware, don't pin */
97 md_print_warn(cr, fplog,
98 "NOTE: We don't know how many logical cores we have, will not pin threads");
104 if (nthreads > nhwthreads)
106 /* We are oversubscribing, don't pin */
107 md_print_warn(NULL, fplog,
108 "WARNING: Oversubscribing the CPU, will not pin threads");
113 if (pin_offset + nthreads > nhwthreads)
115 /* We are oversubscribing, don't pin */
116 md_print_warn(NULL, fplog,
117 "WARNING: The requested pin offset is too large for the available logical cores,\n"
118 " will not pin threads");
124 /* do we need to choose the pinning stride? */
125 bPickPinStride = (*pin_stride == 0);
129 if (rc == 0 && pin_offset + nthreads*nhwthreads_per_core <= nhwthreads)
131 /* Put one thread on each physical core */
132 *pin_stride = nhwthreads_per_core;
136 /* We don't know if we have SMT, and if we do, we don't know
137 * if hw threads in the same physical core are consecutive.
138 * Without SMT the pinning layout should not matter too much.
139 * so we assume a consecutive layout and maximally spread out"
140 * the threads at equal threads per core.
141 * Note that IBM is the major non-x86 case with cpuid support
142 * and probably threads are already pinned by the queuing system,
143 * so we wouldn't end up here in the first place.
145 *pin_stride = (nhwthreads - pin_offset)/nthreads;
150 /* Check the placement of the thread with the largest index to make sure
151 * that the offset & stride doesn't cause pinning beyond the last hardware thread. */
152 if (pin_offset + (nthreads-1)*(*pin_stride) >= nhwthreads)
154 /* We are oversubscribing, don't pin */
155 md_print_warn(NULL, fplog,
156 "WARNING: The requested pinning stride is too large for the available logical cores,\n"
157 " will not pin threads");
165 fprintf(fplog, "Pinning threads with a%s logical core stride of %d\n",
166 bPickPinStride ? "n auto-selected" : " user-specified",
173 /* Set CPU affinity. Can be important for performance.
174 On some systems (e.g. Cray) CPU Affinity is set by default.
175 But default assigning doesn't work (well) with only some ranks
176 having threads. This causes very low performance.
177 External tools have cumbersome syntax for setting affinity
178 in the case that only some ranks have threads.
179 Thus it is important that GROMACS sets the affinity internally
180 if only PME is using threads.
183 gmx_set_thread_affinity(FILE *fplog,
185 gmx_hw_opt_t *hw_opt,
187 const gmx_hw_info_t *hwinfo,
188 const t_inputrec *inputrec)
190 int nth_affinity_set, thread_id_node, thread_id,
191 nthread_local, nthread_node, nthread_hw_max, nphyscore;
193 /* these are inherently global properties that are shared among all threads
195 static const int *locality_order;
197 static gmx_bool have_locality_order = FALSE;
198 static tMPI_Thread_mutex_t locality_order_mtx =
199 TMPI_THREAD_MUTEX_INITIALIZER;
200 static tMPI_Thread_cond_t locality_order_cond =
201 TMPI_THREAD_COND_INITIALIZER;
203 if (hw_opt->thread_affinity == threadaffOFF)
209 /* If the tMPI thread affinity setting is not supported encourage the user
210 * to report it as it's either a bug or an exotic platform which we might
211 * want to support. */
212 if (tMPI_Thread_setaffinity_support() != TMPI_SETAFFINITY_SUPPORT_YES)
214 /* we know Mac OS doesn't support setting thread affinity, so there's
215 no point in warning the user in that case. In any other case
216 the user might be able to do something about it. */
218 md_print_warn(NULL, fplog,
219 "Can not set thread affinities on the current platform. On NUMA systems this\n"
220 "can cause performance degradation. If you think your platform should support\n"
221 "setting affinities, contact the GROMACS developers.");
222 #endif /* __APPLE__ */
226 /* threads on this MPI process or TMPI thread */
227 if (cr->duty & DUTY_PP)
229 nthread_local = gmx_omp_nthreads_get(emntNonbonded);
233 nthread_local = gmx_omp_nthreads_get(emntPME);
236 /* map the current process to cores */
238 nthread_node = nthread_local;
240 if (PAR(cr) || MULTISIM(cr))
242 /* We need to determine a scan of the thread counts in this
247 MPI_Comm_split(MPI_COMM_WORLD, gmx_hostname_num(), cr->rank_intranode,
249 MPI_Scan(&nthread_local, &thread_id_node, 1, MPI_INT, MPI_SUM, comm_intra);
250 /* MPI_Scan is inclusive, but here we need exclusive */
251 thread_id_node -= nthread_local;
252 /* Get the total number of threads on this physical node */
253 MPI_Allreduce(&nthread_local, &nthread_node, 1, MPI_INT, MPI_SUM, comm_intra);
254 MPI_Comm_free(&comm_intra);
258 if (hw_opt->thread_affinity == threadaffAUTO &&
259 nthread_node != hwinfo->nthreads_hw_avail)
261 if (nthread_node > 1 && nthread_node < hwinfo->nthreads_hw_avail)
263 md_print_warn(cr, fplog,
264 "NOTE: The number of threads is not equal to the number of (logical) cores\n"
265 " and the -pin option is set to auto: will not pin thread to cores.\n"
266 " This can lead to significant performance degradation.\n"
267 " Consider using -pin on (and -pinoffset in case you run multiple jobs).\n");
274 if (hw_opt->core_pinning_offset != 0)
276 offset = hw_opt->core_pinning_offset;
277 md_print_info(cr, fplog, "Applying core pinning offset %d\n", offset);
280 /* hw_opt is shared among tMPI threads, so for thread safety we need to do
281 * the layout detection only on master as core_pinning_stride is an in-out
282 * parameter and gets auto-set depending on its initial value.
284 * This is not thread-safe with multi-simulations, but that's anyway not
285 * supported by tMPI. */
291 ret = tMPI_Thread_mutex_lock(&locality_order_mtx);
294 goto locality_order_err;
296 rc = get_thread_affinity_layout(fplog, cr, hwinfo,
298 offset, &hw_opt->core_pinning_stride,
300 have_locality_order = TRUE;
301 ret = tMPI_Thread_cond_broadcast(&locality_order_cond);
304 tMPI_Thread_mutex_unlock(&locality_order_mtx);
305 goto locality_order_err;
307 ret = tMPI_Thread_mutex_unlock(&locality_order_mtx);
310 goto locality_order_err;
316 /* all other threads wait for the locality order data. */
317 ret = tMPI_Thread_mutex_lock(&locality_order_mtx);
320 goto locality_order_err;
323 while (!have_locality_order)
325 ret = tMPI_Thread_cond_wait(&locality_order_cond,
326 &locality_order_mtx);
329 tMPI_Thread_mutex_unlock(&locality_order_mtx);
330 goto locality_order_err;
333 ret = tMPI_Thread_mutex_unlock(&locality_order_mtx);
336 goto locality_order_err;
342 /* Incompatible layout, don't pin, warning was already issued */
346 /* Set the per-thread affinity. In order to be able to check the success
347 * of affinity settings, we will set nth_affinity_set to 1 on threads
348 * where the affinity setting succeded and to 0 where it failed.
349 * Reducing these 0/1 values over the threads will give the total number
350 * of threads on which we succeeded.
352 nth_affinity_set = 0;
353 #pragma omp parallel firstprivate(thread_id_node) num_threads(nthread_local) \
354 reduction(+:nth_affinity_set)
357 gmx_bool setaffinity_ret;
359 thread_id = gmx_omp_get_thread_num();
360 thread_id_node += thread_id;
361 index = offset + thread_id_node*hw_opt->core_pinning_stride;
362 if (locality_order != NULL)
364 core = locality_order[index];
371 setaffinity_ret = tMPI_Thread_setaffinity_single(tMPI_Thread_self(), core);
373 /* store the per-thread success-values of the setaffinity */
374 nth_affinity_set = (setaffinity_ret == 0);
378 fprintf(debug, "On rank %2d, thread %2d, core %2d the affinity setting returned %d\n",
379 cr->nodeid, gmx_omp_get_thread_num(), core, setaffinity_ret);
383 if (nth_affinity_set > nthread_local)
387 sprintf(msg, "Looks like we have set affinity for more threads than "
388 "we have (%d > %d)!\n", nth_affinity_set, nthread_local);
393 /* check & warn if some threads failed to set their affinities */
394 if (nth_affinity_set != nthread_local)
396 char sbuf1[STRLEN], sbuf2[STRLEN];
398 /* sbuf1 contains rank info, while sbuf2 OpenMP thread info */
399 sbuf1[0] = sbuf2[0] = '\0';
400 /* Only add rank info if we have more than one rank. */
404 #ifdef GMX_THREAD_MPI
405 sprintf(sbuf1, "In tMPI thread #%d: ", cr->nodeid);
406 #else /* GMX_LIB_MPI */
407 sprintf(sbuf1, "In MPI process #%d: ", cr->nodeid);
412 if (nthread_local > 1)
414 sprintf(sbuf2, "for %d/%d thread%s ",
415 nthread_local - nth_affinity_set, nthread_local,
416 nthread_local > 1 ? "s" : "");
419 md_print_warn(NULL, fplog,
420 "WARNING: %sAffinity setting %sfailed.\n"
421 " This can cause performance degradation! If you think your setting are\n"
422 " correct, contact the GROMACS developers.",
429 /* any error in affinity setting shouldn't be fatal, but should generate
431 md_print_warn(NULL, fplog,
432 "WARNING: Obtaining affinity information failed due to a basic system error: %s.\n"
433 " This can cause performance degradation! ",
438 /* Check the process affinity mask and if it is found to be non-zero,
439 * will honor it and disable mdrun internal affinity setting.
440 * Note that this will only work on Linux as we use a GNU feature.
443 gmx_check_thread_affinity_set(FILE *fplog, const t_commrec *cr,
444 gmx_hw_opt_t *hw_opt, int ncpus,
445 gmx_bool bAfterOpenmpInit)
447 #ifdef HAVE_SCHED_GETAFFINITY
448 cpu_set_t mask_current;
449 int i, ret, cpu_count, cpu_set;
453 if (hw_opt->thread_affinity == threadaffOFF)
455 /* internal affinity setting is off, don't bother checking process affinity */
459 CPU_ZERO(&mask_current);
460 if ((ret = sched_getaffinity(0, sizeof(cpu_set_t), &mask_current)) != 0)
462 /* failed to query affinity mask, will just return */
465 fprintf(debug, "Failed to query affinity mask (error %d)", ret);
470 /* Before proceeding with the actual check, make sure that the number of
471 * detected CPUs is >= the CPUs in the current set.
472 * We need to check for CPU_COUNT as it was added only in glibc 2.6. */
474 if (ncpus < CPU_COUNT(&mask_current))
478 fprintf(debug, "%d CPUs detected, but %d was returned by CPU_COUNT",
479 ncpus, CPU_COUNT(&mask_current));
483 #endif /* CPU_COUNT */
486 for (i = 0; (i < ncpus && i < CPU_SETSIZE); i++)
488 bAllSet = bAllSet && (CPU_ISSET(i, &mask_current) != 0);
493 if (hw_opt->thread_affinity == threadaffAUTO)
495 if (!bAfterOpenmpInit)
497 md_print_warn(cr, fplog,
498 "Non-default thread affinity set, disabling internal thread affinity");
502 md_print_warn(cr, fplog,
503 "Non-default thread affinity set probably by the OpenMP library,\n"
504 "disabling internal thread affinity");
506 hw_opt->thread_affinity = threadaffOFF;
510 /* Only warn once, at the last check (bAfterOpenmpInit==TRUE) */
511 if (bAfterOpenmpInit)
513 md_print_warn(cr, fplog,
514 "Overriding thread affinity set outside %s\n",
521 fprintf(debug, "Non-default affinity mask found\n");
528 fprintf(debug, "Default affinity mask found\n");
531 #endif /* HAVE_SCHED_GETAFFINITY */