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 #if defined(HAVE_SCHED_H) && defined(HAVE_SCHED_GETAFFINITY)
41 #include <sys/syscall.h>
48 #include "thread_mpi/threads.h"
51 #include "types/commrec.h"
52 #include "types/hw_info.h"
54 #include "gmx_cpuid.h"
55 #include "gmx_omp_nthreads.h"
56 #include "md_logging.h"
57 #include "gmx_thread_affinity.h"
59 #include "gmx_fatal.h"
60 #include "gromacs/utility/gmxomp.h"
63 get_thread_affinity_layout(FILE *fplog,
65 const gmx_hw_info_t * hwinfo,
67 int pin_offset, int * pin_stride,
68 const int **locality_order)
70 int nhwthreads, npkg, ncores, nhwthreads_per_core, rc;
73 const int * hwthread_id;
74 gmx_bool bPickPinStride;
78 gmx_fatal(FARGS, "Negative thread pinning offset requested");
82 gmx_fatal(FARGS, "Negative thread pinning stride requested");
85 rc = gmx_cpuid_topology(hwinfo->cpuid_info, &nhwthreads, &npkg, &ncores,
87 &pkg_id, &core_id, &hwthread_id, locality_order);
91 /* topology information not available or invalid, ignore it */
92 nhwthreads = hwinfo->nthreads_hw_avail;
93 *locality_order = NULL;
97 /* We don't know anything about the hardware, don't pin */
98 md_print_warn(cr, fplog,
99 "NOTE: We don't know how many logical cores we have, will not pin threads");
105 if (nthreads > nhwthreads)
107 /* We are oversubscribing, don't pin */
108 md_print_warn(NULL, fplog,
109 "WARNING: Oversubscribing the CPU, will not pin threads");
114 if (pin_offset + nthreads > nhwthreads)
116 /* We are oversubscribing, don't pin */
117 md_print_warn(NULL, fplog,
118 "WARNING: The requested pin offset is too large for the available logical cores,\n"
119 " will not pin threads");
125 /* do we need to choose the pinning stride? */
126 bPickPinStride = (*pin_stride == 0);
130 if (rc == 0 && pin_offset + nthreads*nhwthreads_per_core <= nhwthreads)
132 /* Put one thread on each physical core */
133 *pin_stride = nhwthreads_per_core;
137 /* We don't know if we have SMT, and if we do, we don't know
138 * if hw threads in the same physical core are consecutive.
139 * Without SMT the pinning layout should not matter too much.
140 * so we assume a consecutive layout and maximally spread out"
141 * the threads at equal threads per core.
142 * Note that IBM is the major non-x86 case with cpuid support
143 * and probably threads are already pinned by the queuing system,
144 * so we wouldn't end up here in the first place.
146 *pin_stride = (nhwthreads - pin_offset)/nthreads;
151 /* Check the placement of the thread with the largest index to make sure
152 * that the offset & stride doesn't cause pinning beyond the last hardware thread. */
153 if (pin_offset + (nthreads-1)*(*pin_stride) >= nhwthreads)
155 /* We are oversubscribing, don't pin */
156 md_print_warn(NULL, fplog,
157 "WARNING: The requested pinning stride is too large for the available logical cores,\n"
158 " will not pin threads");
166 fprintf(fplog, "Pinning threads with a%s logical core stride of %d\n",
167 bPickPinStride ? "n auto-selected" : " user-specified",
174 /* Set CPU affinity. Can be important for performance.
175 On some systems (e.g. Cray) CPU Affinity is set by default.
176 But default assigning doesn't work (well) with only some ranks
177 having threads. This causes very low performance.
178 External tools have cumbersome syntax for setting affinity
179 in the case that only some ranks have threads.
180 Thus it is important that GROMACS sets the affinity internally
181 if only PME is using threads.
184 gmx_set_thread_affinity(FILE *fplog,
186 gmx_hw_opt_t *hw_opt,
187 const gmx_hw_info_t *hwinfo)
189 int nth_affinity_set, thread0_id_node,
190 nthread_local, nthread_node, nthread_hw_max, nphyscore;
192 const int *locality_order;
195 if (hw_opt->thread_affinity == threadaffOFF)
201 /* If the tMPI thread affinity setting is not supported encourage the user
202 * to report it as it's either a bug or an exotic platform which we might
203 * want to support. */
204 if (tMPI_Thread_setaffinity_support() != TMPI_SETAFFINITY_SUPPORT_YES)
206 /* we know Mac OS doesn't support setting thread affinity, so there's
207 no point in warning the user in that case. In any other case
208 the user might be able to do something about it. */
210 md_print_warn(NULL, fplog,
211 "Can not set thread affinities on the current platform. On NUMA systems this\n"
212 "can cause performance degradation. If you think your platform should support\n"
213 "setting affinities, contact the GROMACS developers.");
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, gmx_hostname_num(), cr->rank_intranode,
241 MPI_Scan(&nthread_local, &thread0_id_node, 1, MPI_INT, MPI_SUM, comm_intra);
242 /* MPI_Scan is inclusive, but here we need exclusive */
243 thread0_id_node -= nthread_local;
244 /* Get the total number of threads on this physical node */
245 MPI_Allreduce(&nthread_local, &nthread_node, 1, MPI_INT, MPI_SUM, comm_intra);
246 MPI_Comm_free(&comm_intra);
250 if (hw_opt->thread_affinity == threadaffAUTO &&
251 nthread_node != hwinfo->nthreads_hw_avail)
253 if (nthread_node > 1 && nthread_node < hwinfo->nthreads_hw_avail)
255 md_print_warn(cr, fplog,
256 "NOTE: The number of threads is not equal to the number of (logical) cores\n"
257 " and the -pin option is set to auto: will not pin thread to cores.\n"
258 " This can lead to significant performance degradation.\n"
259 " Consider using -pin on (and -pinoffset in case you run multiple jobs).\n");
266 if (hw_opt->core_pinning_offset != 0)
268 offset = hw_opt->core_pinning_offset;
269 md_print_info(cr, fplog, "Applying core pinning offset %d\n", offset);
272 rc = get_thread_affinity_layout(fplog, cr, hwinfo,
274 offset, &hw_opt->core_pinning_stride,
279 /* Incompatible layout, don't pin, warning was already issued */
283 /* Set the per-thread affinity. In order to be able to check the success
284 * of affinity settings, we will set nth_affinity_set to 1 on threads
285 * where the affinity setting succeded and to 0 where it failed.
286 * Reducing these 0/1 values over the threads will give the total number
287 * of threads on which we succeeded.
289 nth_affinity_set = 0;
290 #pragma omp parallel num_threads(nthread_local) reduction(+:nth_affinity_set)
292 int thread_id, thread_id_node;
294 gmx_bool setaffinity_ret;
296 thread_id = gmx_omp_get_thread_num();
297 thread_id_node = thread0_id_node + thread_id;
298 index = offset + thread_id_node*hw_opt->core_pinning_stride;
299 if (locality_order != NULL)
301 core = locality_order[index];
308 setaffinity_ret = tMPI_Thread_setaffinity_single(tMPI_Thread_self(), core);
310 /* store the per-thread success-values of the setaffinity */
311 nth_affinity_set = (setaffinity_ret == 0);
315 fprintf(debug, "On rank %2d, thread %2d, index %2d, core %2d the affinity setting returned %d\n",
316 cr->nodeid, gmx_omp_get_thread_num(), index, core, setaffinity_ret);
320 if (nth_affinity_set > nthread_local)
324 sprintf(msg, "Looks like we have set affinity for more threads than "
325 "we have (%d > %d)!\n", nth_affinity_set, nthread_local);
330 /* check & warn if some threads failed to set their affinities */
331 if (nth_affinity_set != nthread_local)
333 char sbuf1[STRLEN], sbuf2[STRLEN];
335 /* sbuf1 contains rank info, while sbuf2 OpenMP thread info */
336 sbuf1[0] = sbuf2[0] = '\0';
337 /* Only add rank info if we have more than one rank. */
341 #ifdef GMX_THREAD_MPI
342 sprintf(sbuf1, "In tMPI thread #%d: ", cr->nodeid);
343 #else /* GMX_LIB_MPI */
344 sprintf(sbuf1, "In MPI process #%d: ", cr->nodeid);
349 if (nthread_local > 1)
351 sprintf(sbuf2, "for %d/%d thread%s ",
352 nthread_local - nth_affinity_set, nthread_local,
353 nthread_local > 1 ? "s" : "");
356 md_print_warn(NULL, fplog,
357 "WARNING: %sAffinity setting %sfailed.\n"
358 " This can cause performance degradation! If you think your setting are\n"
359 " correct, contact the GROMACS developers.",
366 /* Check the process affinity mask and if it is found to be non-zero,
367 * will honor it and disable mdrun internal affinity setting.
368 * Note that this will only work on Linux as we use a GNU feature.
371 gmx_check_thread_affinity_set(FILE gmx_unused *fplog,
372 const t_commrec gmx_unused *cr,
373 gmx_hw_opt_t gmx_unused *hw_opt,
374 int gmx_unused ncpus,
375 gmx_bool gmx_unused bAfterOpenmpInit)
377 #ifdef HAVE_SCHED_GETAFFINITY
378 cpu_set_t mask_current;
379 int i, ret, cpu_count, cpu_set;
383 if (hw_opt->thread_affinity == threadaffOFF)
385 /* internal affinity setting is off, don't bother checking process affinity */
389 CPU_ZERO(&mask_current);
390 if ((ret = sched_getaffinity(0, sizeof(cpu_set_t), &mask_current)) != 0)
392 /* failed to query affinity mask, will just return */
395 fprintf(debug, "Failed to query affinity mask (error %d)", ret);
400 /* Before proceeding with the actual check, make sure that the number of
401 * detected CPUs is >= the CPUs in the current set.
402 * We need to check for CPU_COUNT as it was added only in glibc 2.6. */
404 if (ncpus < CPU_COUNT(&mask_current))
408 fprintf(debug, "%d CPUs detected, but %d was returned by CPU_COUNT",
409 ncpus, CPU_COUNT(&mask_current));
413 #endif /* CPU_COUNT */
416 for (i = 0; (i < ncpus && i < CPU_SETSIZE); i++)
418 bAllSet = bAllSet && (CPU_ISSET(i, &mask_current) != 0);
423 if (hw_opt->thread_affinity == threadaffAUTO)
425 if (!bAfterOpenmpInit)
427 md_print_warn(cr, fplog,
428 "Non-default thread affinity set, disabling internal thread affinity");
432 md_print_warn(cr, fplog,
433 "Non-default thread affinity set probably by the OpenMP library,\n"
434 "disabling internal thread affinity");
436 hw_opt->thread_affinity = threadaffOFF;
440 /* Only warn once, at the last check (bAfterOpenmpInit==TRUE) */
441 if (bAfterOpenmpInit)
443 md_print_warn(cr, fplog,
444 "Overriding thread affinity set outside %s\n",
451 fprintf(debug, "Non-default affinity mask found\n");
458 fprintf(debug, "Default affinity mask found\n");
461 #endif /* HAVE_SCHED_GETAFFINITY */