2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2012,2013,2014,2015, by the GROMACS development team, led by
5 * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
6 * and including many others, as listed in the AUTHORS file in the
7 * top-level source 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.
37 #include "gromacs/legacyheaders/gmx_thread_affinity.h"
45 #ifdef HAVE_SCHED_AFFINITY
47 # include <sys/syscall.h>
50 #include "thread_mpi/threads.h"
52 #include "gromacs/legacyheaders/copyrite.h"
53 #include "gromacs/legacyheaders/gmx_cpuid.h"
54 #include "gromacs/legacyheaders/gmx_omp_nthreads.h"
55 #include "gromacs/legacyheaders/md_logging.h"
56 #include "gromacs/legacyheaders/typedefs.h"
57 #include "gromacs/legacyheaders/types/commrec.h"
58 #include "gromacs/legacyheaders/types/hw_info.h"
59 #include "gromacs/utility/basenetwork.h"
60 #include "gromacs/utility/cstringutil.h"
61 #include "gromacs/utility/fatalerror.h"
62 #include "gromacs/utility/gmxassert.h"
63 #include "gromacs/utility/gmxomp.h"
64 #include "gromacs/utility/smalloc.h"
67 get_thread_affinity_layout(FILE *fplog,
69 const gmx_hw_info_t * hwinfo,
71 int pin_offset, int * pin_stride,
72 const int **locality_order)
74 int nhwthreads, npkg, ncores, nhwthreads_per_core, rc;
77 const int * hwthread_id;
78 gmx_bool bPickPinStride;
82 gmx_fatal(FARGS, "Negative thread pinning offset requested");
86 gmx_fatal(FARGS, "Negative thread pinning stride requested");
89 rc = gmx_cpuid_topology(hwinfo->cpuid_info, &nhwthreads, &npkg, &ncores,
91 &pkg_id, &core_id, &hwthread_id, locality_order);
95 /* topology information not available or invalid, ignore it */
96 nhwthreads = hwinfo->nthreads_hw_avail;
97 *locality_order = NULL;
101 /* We don't know anything about the hardware, don't pin */
102 md_print_warn(cr, fplog,
103 "NOTE: No information on available cores, thread pinning disabled.");
109 if (nthreads > nhwthreads)
111 /* We are oversubscribing, don't pin */
112 md_print_warn(NULL, fplog,
113 "NOTE: Oversubscribing the CPU, will not pin threads");
118 if (pin_offset + nthreads > nhwthreads)
120 /* We are oversubscribing, don't pin */
121 md_print_warn(NULL, fplog,
122 "WARNING: Requested offset too large for available cores, thread pinning disabled.");
128 /* do we need to choose the pinning stride? */
129 bPickPinStride = (*pin_stride == 0);
133 if (rc == 0 && pin_offset + nthreads*nhwthreads_per_core <= nhwthreads)
135 /* Put one thread on each physical core */
136 *pin_stride = nhwthreads_per_core;
140 /* We don't know if we have SMT, and if we do, we don't know
141 * if hw threads in the same physical core are consecutive.
142 * Without SMT the pinning layout should not matter too much.
143 * so we assume a consecutive layout and maximally spread out"
144 * the threads at equal threads per core.
145 * Note that IBM is the major non-x86 case with cpuid support
146 * and probably threads are already pinned by the queuing system,
147 * so we wouldn't end up here in the first place.
149 *pin_stride = (nhwthreads - pin_offset)/nthreads;
154 /* Check the placement of the thread with the largest index to make sure
155 * that the offset & stride doesn't cause pinning beyond the last hardware thread. */
156 if (pin_offset + (nthreads-1)*(*pin_stride) >= nhwthreads)
158 /* We are oversubscribing, don't pin */
159 md_print_warn(NULL, fplog,
160 "WARNING: Requested stride too large for available cores, thread pinning disabled.");
168 fprintf(fplog, "Pinning threads with a%s logical core stride of %d\n",
169 bPickPinStride ? "n auto-selected" : " user-specified",
176 /* Set CPU affinity. Can be important for performance.
177 On some systems (e.g. Cray) CPU Affinity is set by default.
178 But default assigning doesn't work (well) with only some ranks
179 having threads. This causes very low performance.
180 External tools have cumbersome syntax for setting affinity
181 in the case that only some ranks have threads.
182 Thus it is important that GROMACS sets the affinity internally
183 if only PME is using threads.
186 gmx_set_thread_affinity(FILE *fplog,
188 gmx_hw_opt_t *hw_opt,
189 const gmx_hw_info_t *hwinfo)
191 int nth_affinity_set, thread0_id_node,
192 nthread_local, nthread_node;
194 const int *locality_order;
197 if (hw_opt->thread_affinity == threadaffOFF)
203 /* If the tMPI thread affinity setting is not supported encourage the user
204 * to report it as it's either a bug or an exotic platform which we might
205 * want to support. */
206 if (tMPI_Thread_setaffinity_support() != TMPI_SETAFFINITY_SUPPORT_YES)
208 /* we know Mac OS & BlueGene do not support setting thread affinity, so there's
209 no point in warning the user in that case. In any other case
210 the user might be able to do something about it. */
211 #if !defined(__APPLE__) && !defined(__bg__)
212 md_print_warn(NULL, fplog,
213 "NOTE: Cannot set thread affinities on the current platform.");
214 #endif /* __APPLE__ */
218 /* threads on this MPI process or TMPI thread */
219 if (cr->duty & DUTY_PP)
221 nthread_local = gmx_omp_nthreads_get(emntNonbonded);
225 nthread_local = gmx_omp_nthreads_get(emntPME);
228 /* map the current process to cores */
230 nthread_node = nthread_local;
232 if (PAR(cr) || MULTISIM(cr))
234 /* We need to determine a scan of the thread counts in this
239 MPI_Comm_split(MPI_COMM_WORLD,
240 gmx_physicalnode_id_hash(), cr->rank_intranode,
242 MPI_Scan(&nthread_local, &thread0_id_node, 1, MPI_INT, MPI_SUM, comm_intra);
243 /* MPI_Scan is inclusive, but here we need exclusive */
244 thread0_id_node -= nthread_local;
245 /* Get the total number of threads on this physical node */
246 MPI_Allreduce(&nthread_local, &nthread_node, 1, MPI_INT, MPI_SUM, comm_intra);
247 MPI_Comm_free(&comm_intra);
251 if (hw_opt->thread_affinity == threadaffAUTO &&
252 nthread_node != hwinfo->nthreads_hw_avail)
254 if (nthread_node > 1 && nthread_node < hwinfo->nthreads_hw_avail)
256 md_print_warn(cr, fplog,
257 "NOTE: The number of threads is not equal to the number of (logical) cores\n"
258 " and the -pin option is set to auto: will not pin thread to cores.\n"
259 " This can lead to significant performance degradation.\n"
260 " Consider using -pin on (and -pinoffset in case you run multiple jobs).\n");
267 if (hw_opt->core_pinning_offset != 0)
269 offset = hw_opt->core_pinning_offset;
270 md_print_info(cr, fplog, "Applying core pinning offset %d\n", offset);
273 rc = get_thread_affinity_layout(fplog, cr, hwinfo,
275 offset, &hw_opt->core_pinning_stride,
280 /* Incompatible layout, don't pin, warning was already issued */
284 /* Set the per-thread affinity. In order to be able to check the success
285 * of affinity settings, we will set nth_affinity_set to 1 on threads
286 * where the affinity setting succeded and to 0 where it failed.
287 * Reducing these 0/1 values over the threads will give the total number
288 * of threads on which we succeeded.
291 // To avoid warnings from the static analyzer we initialize nth_affinity_set
292 // to zero outside the OpenMP block, and then add to it inside the block.
293 // The value will still always be 0 or 1 from each thread.
294 nth_affinity_set = 0;
295 #pragma omp parallel num_threads(nthread_local) reduction(+:nth_affinity_set)
297 int thread_id, thread_id_node;
299 gmx_bool setaffinity_ret;
301 thread_id = gmx_omp_get_thread_num();
302 thread_id_node = thread0_id_node + thread_id;
303 index = offset + thread_id_node*hw_opt->core_pinning_stride;
304 if (locality_order != NULL)
306 core = locality_order[index];
313 setaffinity_ret = tMPI_Thread_setaffinity_single(tMPI_Thread_self(), core);
315 /* store the per-thread success-values of the setaffinity */
316 nth_affinity_set += (setaffinity_ret == 0);
320 fprintf(debug, "On rank %2d, thread %2d, index %2d, core %2d the affinity setting returned %d\n",
321 cr->nodeid, gmx_omp_get_thread_num(), index, core, setaffinity_ret);
325 if (nth_affinity_set > nthread_local)
329 sprintf(msg, "Looks like we have set affinity for more threads than "
330 "we have (%d > %d)!\n", nth_affinity_set, nthread_local);
335 /* check & warn if some threads failed to set their affinities */
336 if (nth_affinity_set != nthread_local)
338 char sbuf1[STRLEN], sbuf2[STRLEN];
340 /* sbuf1 contains rank info, while sbuf2 OpenMP thread info */
341 sbuf1[0] = sbuf2[0] = '\0';
342 /* Only add rank info if we have more than one rank. */
346 #ifdef GMX_THREAD_MPI
347 sprintf(sbuf1, "In tMPI thread #%d: ", cr->nodeid);
348 #else /* GMX_LIB_MPI */
349 sprintf(sbuf1, "In MPI process #%d: ", cr->nodeid);
354 if (nthread_local > 1)
356 sprintf(sbuf2, "for %d/%d thread%s ",
357 nthread_local - nth_affinity_set, nthread_local,
358 nthread_local > 1 ? "s" : "");
361 md_print_warn(NULL, fplog,
362 "NOTE: %sAffinity setting %sfailed. This can cause performance degradation.\n"
363 " If you think your settings are correct, ask on the gmx-users list.",
370 /* Check the process affinity mask and if it is found to be non-zero,
371 * will honor it and disable mdrun internal affinity setting.
372 * Note that this will only work on Linux as we use a GNU feature.
375 gmx_check_thread_affinity_set(FILE *fplog,
377 gmx_hw_opt_t *hw_opt,
378 int gmx_unused nthreads_hw_avail,
379 gmx_bool bAfterOpenmpInit)
381 GMX_RELEASE_ASSERT(hw_opt, "hw_opt must be a non-NULL pointer");
383 if (!bAfterOpenmpInit)
385 /* Check for externally set OpenMP affinity and turn off internal
386 * pinning if any is found. We need to do this check early to tell
387 * thread-MPI whether it should do pinning when spawning threads.
388 * TODO: the above no longer holds, we should move these checks later
390 if (hw_opt->thread_affinity != threadaffOFF)
393 if (!gmx_omp_check_thread_affinity(&message))
395 /* TODO: with -pin auto we should only warn when using all cores */
396 md_print_warn(cr, fplog, "%s", message);
398 hw_opt->thread_affinity = threadaffOFF;
402 /* With thread-MPI this is needed as pinning might get turned off,
403 * which needs to be known before starting thread-MPI.
404 * With thread-MPI hw_opt is processed here on the master rank
405 * and passed to the other ranks later, so we only do this on master.
411 #ifndef GMX_THREAD_MPI
416 #ifdef HAVE_SCHED_GETAFFINITY
418 cpu_set_t mask_current;
420 if (hw_opt->thread_affinity == threadaffOFF)
422 /* internal affinity setting is off, don't bother checking process affinity */
426 CPU_ZERO(&mask_current);
427 if ((ret = sched_getaffinity(0, sizeof(cpu_set_t), &mask_current)) != 0)
429 /* failed to query affinity mask, will just return */
432 fprintf(debug, "Failed to query affinity mask (error %d)", ret);
437 /* Before proceeding with the actual check, make sure that the number of
438 * detected CPUs is >= the CPUs in the current set.
439 * We need to check for CPU_COUNT as it was added only in glibc 2.6. */
441 if (nthreads_hw_avail < CPU_COUNT(&mask_current))
445 fprintf(debug, "%d hardware threads detected, but %d was returned by CPU_COUNT",
446 nthreads_hw_avail, CPU_COUNT(&mask_current));
450 #endif /* CPU_COUNT */
452 gmx_bool bAllSet = TRUE;
453 for (int i = 0; (i < nthreads_hw_avail && i < CPU_SETSIZE); i++)
455 bAllSet = bAllSet && (CPU_ISSET(i, &mask_current) != 0);
459 gmx_bool bAllSet_All;
461 MPI_Allreduce(&bAllSet, &bAllSet_All, 1, MPI_INT, MPI_LAND, MPI_COMM_WORLD);
462 bAllSet = bAllSet_All;
467 if (hw_opt->thread_affinity == threadaffAUTO)
469 if (!bAfterOpenmpInit)
471 md_print_warn(cr, fplog,
472 "Non-default thread affinity set, disabling internal thread affinity");
476 md_print_warn(cr, fplog,
477 "Non-default thread affinity set probably by the OpenMP library,\n"
478 "disabling internal thread affinity");
480 hw_opt->thread_affinity = threadaffOFF;
484 /* Only warn once, at the last check (bAfterOpenmpInit==TRUE) */
485 if (bAfterOpenmpInit)
487 md_print_warn(cr, fplog,
488 "Overriding thread affinity set outside %s\n",
495 fprintf(debug, "Non-default affinity mask found\n");
502 fprintf(debug, "Default affinity mask found\n");
505 #endif /* HAVE_SCHED_AFFINITY */