2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2012,2013,2014, 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.
38 #ifdef HAVE_SCHED_AFFINITY
40 # define _GNU_SOURCE 1
43 # include <sys/syscall.h>
50 #include "thread_mpi/threads.h"
52 #include "gromacs/legacyheaders/typedefs.h"
53 #include "gromacs/legacyheaders/types/commrec.h"
54 #include "gromacs/legacyheaders/types/hw_info.h"
55 #include "gromacs/legacyheaders/copyrite.h"
56 #include "gromacs/legacyheaders/gmx_cpuid.h"
57 #include "gromacs/legacyheaders/gmx_omp_nthreads.h"
58 #include "gromacs/legacyheaders/md_logging.h"
59 #include "gromacs/legacyheaders/gmx_thread_affinity.h"
61 #include "gromacs/utility/basenetwork.h"
62 #include "gromacs/utility/cstringutil.h"
63 #include "gromacs/utility/fatalerror.h"
64 #include "gromacs/utility/gmxomp.h"
65 #include "gromacs/utility/smalloc.h"
68 get_thread_affinity_layout(FILE *fplog,
70 const gmx_hw_info_t * hwinfo,
72 int pin_offset, int * pin_stride,
73 const int **locality_order)
75 int nhwthreads, npkg, ncores, nhwthreads_per_core, rc;
78 const int * hwthread_id;
79 gmx_bool bPickPinStride;
83 gmx_fatal(FARGS, "Negative thread pinning offset requested");
87 gmx_fatal(FARGS, "Negative thread pinning stride requested");
90 rc = gmx_cpuid_topology(hwinfo->cpuid_info, &nhwthreads, &npkg, &ncores,
92 &pkg_id, &core_id, &hwthread_id, locality_order);
96 /* topology information not available or invalid, ignore it */
97 nhwthreads = hwinfo->nthreads_hw_avail;
98 *locality_order = NULL;
102 /* We don't know anything about the hardware, don't pin */
103 md_print_warn(cr, fplog,
104 "NOTE: We don't know how many logical cores we have, will not pin threads");
110 if (nthreads > nhwthreads)
112 /* We are oversubscribing, don't pin */
113 md_print_warn(NULL, fplog,
114 "WARNING: Oversubscribing the CPU, will not pin threads");
119 if (pin_offset + nthreads > nhwthreads)
121 /* We are oversubscribing, don't pin */
122 md_print_warn(NULL, fplog,
123 "WARNING: The requested pin offset is too large for the available logical cores,\n"
124 " will not pin threads");
130 /* do we need to choose the pinning stride? */
131 bPickPinStride = (*pin_stride == 0);
135 if (rc == 0 && pin_offset + nthreads*nhwthreads_per_core <= nhwthreads)
137 /* Put one thread on each physical core */
138 *pin_stride = nhwthreads_per_core;
142 /* We don't know if we have SMT, and if we do, we don't know
143 * if hw threads in the same physical core are consecutive.
144 * Without SMT the pinning layout should not matter too much.
145 * so we assume a consecutive layout and maximally spread out"
146 * the threads at equal threads per core.
147 * Note that IBM is the major non-x86 case with cpuid support
148 * and probably threads are already pinned by the queuing system,
149 * so we wouldn't end up here in the first place.
151 *pin_stride = (nhwthreads - pin_offset)/nthreads;
156 /* Check the placement of the thread with the largest index to make sure
157 * that the offset & stride doesn't cause pinning beyond the last hardware thread. */
158 if (pin_offset + (nthreads-1)*(*pin_stride) >= nhwthreads)
160 /* We are oversubscribing, don't pin */
161 md_print_warn(NULL, fplog,
162 "WARNING: The requested pinning stride is too large for the available logical cores,\n"
163 " will not pin threads");
171 fprintf(fplog, "Pinning threads with a%s logical core stride of %d\n",
172 bPickPinStride ? "n auto-selected" : " user-specified",
179 /* Set CPU affinity. Can be important for performance.
180 On some systems (e.g. Cray) CPU Affinity is set by default.
181 But default assigning doesn't work (well) with only some ranks
182 having threads. This causes very low performance.
183 External tools have cumbersome syntax for setting affinity
184 in the case that only some ranks have threads.
185 Thus it is important that GROMACS sets the affinity internally
186 if only PME is using threads.
189 gmx_set_thread_affinity(FILE *fplog,
191 gmx_hw_opt_t *hw_opt,
192 const gmx_hw_info_t *hwinfo)
194 int nth_affinity_set, thread0_id_node,
195 nthread_local, nthread_node, nthread_hw_max, nphyscore;
197 const int *locality_order;
200 if (hw_opt->thread_affinity == threadaffOFF)
206 /* If the tMPI thread affinity setting is not supported encourage the user
207 * to report it as it's either a bug or an exotic platform which we might
208 * want to support. */
209 if (tMPI_Thread_setaffinity_support() != TMPI_SETAFFINITY_SUPPORT_YES)
211 /* we know Mac OS doesn't support setting thread affinity, so there's
212 no point in warning the user in that case. In any other case
213 the user might be able to do something about it. */
215 md_print_warn(NULL, fplog,
216 "Can not set thread affinities on the current platform. On NUMA systems this\n"
217 "can cause performance degradation. If you think your platform should support\n"
218 "setting affinities, contact the GROMACS developers.");
219 #endif /* __APPLE__ */
223 /* threads on this MPI process or TMPI thread */
224 if (cr->duty & DUTY_PP)
226 nthread_local = gmx_omp_nthreads_get(emntNonbonded);
230 nthread_local = gmx_omp_nthreads_get(emntPME);
233 /* map the current process to cores */
235 nthread_node = nthread_local;
237 if (PAR(cr) || MULTISIM(cr))
239 /* We need to determine a scan of the thread counts in this
244 MPI_Comm_split(MPI_COMM_WORLD,
245 gmx_physicalnode_id_hash(), cr->rank_intranode,
247 MPI_Scan(&nthread_local, &thread0_id_node, 1, MPI_INT, MPI_SUM, comm_intra);
248 /* MPI_Scan is inclusive, but here we need exclusive */
249 thread0_id_node -= nthread_local;
250 /* Get the total number of threads on this physical node */
251 MPI_Allreduce(&nthread_local, &nthread_node, 1, MPI_INT, MPI_SUM, comm_intra);
252 MPI_Comm_free(&comm_intra);
256 if (hw_opt->thread_affinity == threadaffAUTO &&
257 nthread_node != hwinfo->nthreads_hw_avail)
259 if (nthread_node > 1 && nthread_node < hwinfo->nthreads_hw_avail)
261 md_print_warn(cr, fplog,
262 "NOTE: The number of threads is not equal to the number of (logical) cores\n"
263 " and the -pin option is set to auto: will not pin thread to cores.\n"
264 " This can lead to significant performance degradation.\n"
265 " Consider using -pin on (and -pinoffset in case you run multiple jobs).\n");
272 if (hw_opt->core_pinning_offset != 0)
274 offset = hw_opt->core_pinning_offset;
275 md_print_info(cr, fplog, "Applying core pinning offset %d\n", offset);
278 rc = get_thread_affinity_layout(fplog, cr, hwinfo,
280 offset, &hw_opt->core_pinning_stride,
285 /* Incompatible layout, don't pin, warning was already issued */
289 /* Set the per-thread affinity. In order to be able to check the success
290 * of affinity settings, we will set nth_affinity_set to 1 on threads
291 * where the affinity setting succeded and to 0 where it failed.
292 * Reducing these 0/1 values over the threads will give the total number
293 * of threads on which we succeeded.
295 nth_affinity_set = 0;
296 #pragma omp parallel num_threads(nthread_local) reduction(+:nth_affinity_set)
298 int thread_id, thread_id_node;
300 gmx_bool setaffinity_ret;
302 thread_id = gmx_omp_get_thread_num();
303 thread_id_node = thread0_id_node + thread_id;
304 index = offset + thread_id_node*hw_opt->core_pinning_stride;
305 if (locality_order != NULL)
307 core = locality_order[index];
314 setaffinity_ret = tMPI_Thread_setaffinity_single(tMPI_Thread_self(), core);
316 /* store the per-thread success-values of the setaffinity */
317 nth_affinity_set = (setaffinity_ret == 0);
321 fprintf(debug, "On rank %2d, thread %2d, index %2d, core %2d the affinity setting returned %d\n",
322 cr->nodeid, gmx_omp_get_thread_num(), index, core, setaffinity_ret);
326 if (nth_affinity_set > nthread_local)
330 sprintf(msg, "Looks like we have set affinity for more threads than "
331 "we have (%d > %d)!\n", nth_affinity_set, nthread_local);
336 /* check & warn if some threads failed to set their affinities */
337 if (nth_affinity_set != nthread_local)
339 char sbuf1[STRLEN], sbuf2[STRLEN];
341 /* sbuf1 contains rank info, while sbuf2 OpenMP thread info */
342 sbuf1[0] = sbuf2[0] = '\0';
343 /* Only add rank info if we have more than one rank. */
347 #ifdef GMX_THREAD_MPI
348 sprintf(sbuf1, "In tMPI thread #%d: ", cr->nodeid);
349 #else /* GMX_LIB_MPI */
350 sprintf(sbuf1, "In MPI process #%d: ", cr->nodeid);
355 if (nthread_local > 1)
357 sprintf(sbuf2, "for %d/%d thread%s ",
358 nthread_local - nth_affinity_set, nthread_local,
359 nthread_local > 1 ? "s" : "");
362 md_print_warn(NULL, fplog,
363 "WARNING: %sAffinity setting %sfailed.\n"
364 " This can cause performance degradation! If you think your setting are\n"
365 " correct, contact the GROMACS developers.",
372 /* Check the process affinity mask and if it is found to be non-zero,
373 * will honor it and disable mdrun internal affinity setting.
374 * Note that this will only work on Linux as we use a GNU feature.
377 gmx_check_thread_affinity_set(FILE *fplog,
379 gmx_hw_opt_t *hw_opt,
380 int gmx_unused nthreads_hw_avail,
381 gmx_bool bAfterOpenmpInit)
383 #ifdef HAVE_SCHED_AFFINITY
384 cpu_set_t mask_current;
385 int i, ret, cpu_count, cpu_set;
390 if (!bAfterOpenmpInit)
392 /* Check for externally set OpenMP affinity and turn off internal
393 * pinning if any is found. We need to do this check early to tell
394 * thread-MPI whether it should do pinning when spawning threads.
395 * TODO: the above no longer holds, we should move these checks later
397 if (hw_opt->thread_affinity != threadaffOFF)
400 if (!gmx_omp_check_thread_affinity(&message))
402 /* TODO: with -pin auto we should only warn when using all cores */
403 md_print_warn(cr, fplog, "%s", message);
405 hw_opt->thread_affinity = threadaffOFF;
409 /* With thread-MPI this is needed as pinning might get turned off,
410 * which needs to be known before starting thread-MPI.
411 * With thread-MPI hw_opt is processed here on the master rank
412 * and passed to the other ranks later, so we only do this on master.
418 #ifndef GMX_THREAD_MPI
423 #ifdef HAVE_SCHED_GETAFFINITY
424 if (hw_opt->thread_affinity == threadaffOFF)
426 /* internal affinity setting is off, don't bother checking process affinity */
430 CPU_ZERO(&mask_current);
431 if ((ret = sched_getaffinity(0, sizeof(cpu_set_t), &mask_current)) != 0)
433 /* failed to query affinity mask, will just return */
436 fprintf(debug, "Failed to query affinity mask (error %d)", ret);
441 /* Before proceeding with the actual check, make sure that the number of
442 * detected CPUs is >= the CPUs in the current set.
443 * We need to check for CPU_COUNT as it was added only in glibc 2.6. */
445 if (nthreads_hw_avail < CPU_COUNT(&mask_current))
449 fprintf(debug, "%d hardware threads detected, but %d was returned by CPU_COUNT",
450 nthreads_hw_avail, CPU_COUNT(&mask_current));
454 #endif /* CPU_COUNT */
457 for (i = 0; (i < nthreads_hw_avail && i < CPU_SETSIZE); i++)
459 bAllSet = bAllSet && (CPU_ISSET(i, &mask_current) != 0);
464 if (hw_opt->thread_affinity == threadaffAUTO)
466 if (!bAfterOpenmpInit)
468 md_print_warn(cr, fplog,
469 "Non-default thread affinity set, disabling internal thread affinity");
473 md_print_warn(cr, fplog,
474 "Non-default thread affinity set probably by the OpenMP library,\n"
475 "disabling internal thread affinity");
477 hw_opt->thread_affinity = threadaffOFF;
481 /* Only warn once, at the last check (bAfterOpenmpInit==TRUE) */
482 if (bAfterOpenmpInit)
484 md_print_warn(cr, fplog,
485 "Overriding thread affinity set outside %s\n",
492 fprintf(debug, "Non-default affinity mask found\n");
499 fprintf(debug, "Default affinity mask found\n");
502 #endif /* HAVE_SCHED_AFFINITY */