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)
40 # define _GNU_SOURCE 1
43 # include <sys/syscall.h>
50 #include "thread_mpi/threads.h"
53 #include "types/commrec.h"
54 #include "types/hw_info.h"
56 #include "gmx_cpuid.h"
57 #include "gmx_omp_nthreads.h"
58 #include "md_logging.h"
59 #include "gmx_thread_affinity.h"
61 #include "gmx_fatal.h"
62 #include "gromacs/utility/gmxomp.h"
65 get_thread_affinity_layout(FILE *fplog,
67 const gmx_hw_info_t * hwinfo,
69 int pin_offset, int * pin_stride,
70 const int **locality_order)
72 int nhwthreads, npkg, ncores, nhwthreads_per_core, rc;
75 const int * hwthread_id;
76 gmx_bool bPickPinStride;
80 gmx_fatal(FARGS, "Negative thread pinning offset requested");
84 gmx_fatal(FARGS, "Negative thread pinning stride requested");
87 rc = gmx_cpuid_topology(hwinfo->cpuid_info, &nhwthreads, &npkg, &ncores,
89 &pkg_id, &core_id, &hwthread_id, locality_order);
93 /* topology information not available or invalid, ignore it */
94 nhwthreads = hwinfo->nthreads_hw_avail;
95 *locality_order = NULL;
99 /* We don't know anything about the hardware, don't pin */
100 md_print_warn(cr, fplog,
101 "NOTE: We don't know how many logical cores we have, will not pin threads");
107 if (nthreads > nhwthreads)
109 /* We are oversubscribing, don't pin */
110 md_print_warn(NULL, fplog,
111 "WARNING: Oversubscribing the CPU, will not pin threads");
116 if (pin_offset + nthreads > nhwthreads)
118 /* We are oversubscribing, don't pin */
119 md_print_warn(NULL, fplog,
120 "WARNING: The requested pin offset is too large for the available logical cores,\n"
121 " will not pin threads");
127 /* do we need to choose the pinning stride? */
128 bPickPinStride = (*pin_stride == 0);
132 if (rc == 0 && pin_offset + nthreads*nhwthreads_per_core <= nhwthreads)
134 /* Put one thread on each physical core */
135 *pin_stride = nhwthreads_per_core;
139 /* We don't know if we have SMT, and if we do, we don't know
140 * if hw threads in the same physical core are consecutive.
141 * Without SMT the pinning layout should not matter too much.
142 * so we assume a consecutive layout and maximally spread out"
143 * the threads at equal threads per core.
144 * Note that IBM is the major non-x86 case with cpuid support
145 * and probably threads are already pinned by the queuing system,
146 * so we wouldn't end up here in the first place.
148 *pin_stride = (nhwthreads - pin_offset)/nthreads;
153 /* Check the placement of the thread with the largest index to make sure
154 * that the offset & stride doesn't cause pinning beyond the last hardware thread. */
155 if (pin_offset + (nthreads-1)*(*pin_stride) >= nhwthreads)
157 /* We are oversubscribing, don't pin */
158 md_print_warn(NULL, fplog,
159 "WARNING: The requested pinning stride is too large for the available logical cores,\n"
160 " will not pin threads");
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, nthread_hw_max, nphyscore;
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 doesn't 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. */
212 md_print_warn(NULL, fplog,
213 "Can not set thread affinities on the current platform. On NUMA systems this\n"
214 "can cause performance degradation. If you think your platform should support\n"
215 "setting affinities, contact the GROMACS developers.");
216 #endif /* __APPLE__ */
220 /* threads on this MPI process or TMPI thread */
221 if (cr->duty & DUTY_PP)
223 nthread_local = gmx_omp_nthreads_get(emntNonbonded);
227 nthread_local = gmx_omp_nthreads_get(emntPME);
230 /* map the current process to cores */
232 nthread_node = nthread_local;
234 if (PAR(cr) || MULTISIM(cr))
236 /* We need to determine a scan of the thread counts in this
241 MPI_Comm_split(MPI_COMM_WORLD,
242 gmx_physicalnode_id_hash(), cr->rank_intranode,
244 MPI_Scan(&nthread_local, &thread0_id_node, 1, MPI_INT, MPI_SUM, comm_intra);
245 /* MPI_Scan is inclusive, but here we need exclusive */
246 thread0_id_node -= nthread_local;
247 /* Get the total number of threads on this physical node */
248 MPI_Allreduce(&nthread_local, &nthread_node, 1, MPI_INT, MPI_SUM, comm_intra);
249 MPI_Comm_free(&comm_intra);
253 if (hw_opt->thread_affinity == threadaffAUTO &&
254 nthread_node != hwinfo->nthreads_hw_avail)
256 if (nthread_node > 1 && nthread_node < hwinfo->nthreads_hw_avail)
258 md_print_warn(cr, fplog,
259 "NOTE: The number of threads is not equal to the number of (logical) cores\n"
260 " and the -pin option is set to auto: will not pin thread to cores.\n"
261 " This can lead to significant performance degradation.\n"
262 " Consider using -pin on (and -pinoffset in case you run multiple jobs).\n");
269 if (hw_opt->core_pinning_offset != 0)
271 offset = hw_opt->core_pinning_offset;
272 md_print_info(cr, fplog, "Applying core pinning offset %d\n", offset);
275 rc = get_thread_affinity_layout(fplog, cr, hwinfo,
277 offset, &hw_opt->core_pinning_stride,
282 /* Incompatible layout, don't pin, warning was already issued */
286 /* Set the per-thread affinity. In order to be able to check the success
287 * of affinity settings, we will set nth_affinity_set to 1 on threads
288 * where the affinity setting succeded and to 0 where it failed.
289 * Reducing these 0/1 values over the threads will give the total number
290 * of threads on which we succeeded.
292 nth_affinity_set = 0;
293 #pragma omp parallel num_threads(nthread_local) reduction(+:nth_affinity_set)
295 int thread_id, thread_id_node;
297 gmx_bool setaffinity_ret;
299 thread_id = gmx_omp_get_thread_num();
300 thread_id_node = thread0_id_node + thread_id;
301 index = offset + thread_id_node*hw_opt->core_pinning_stride;
302 if (locality_order != NULL)
304 core = locality_order[index];
311 setaffinity_ret = tMPI_Thread_setaffinity_single(tMPI_Thread_self(), core);
313 /* store the per-thread success-values of the setaffinity */
314 nth_affinity_set = (setaffinity_ret == 0);
318 fprintf(debug, "On rank %2d, thread %2d, index %2d, core %2d the affinity setting returned %d\n",
319 cr->nodeid, gmx_omp_get_thread_num(), index, core, setaffinity_ret);
323 if (nth_affinity_set > nthread_local)
327 sprintf(msg, "Looks like we have set affinity for more threads than "
328 "we have (%d > %d)!\n", nth_affinity_set, nthread_local);
333 /* check & warn if some threads failed to set their affinities */
334 if (nth_affinity_set != nthread_local)
336 char sbuf1[STRLEN], sbuf2[STRLEN];
338 /* sbuf1 contains rank info, while sbuf2 OpenMP thread info */
339 sbuf1[0] = sbuf2[0] = '\0';
340 /* Only add rank info if we have more than one rank. */
344 #ifdef GMX_THREAD_MPI
345 sprintf(sbuf1, "In tMPI thread #%d: ", cr->nodeid);
346 #else /* GMX_LIB_MPI */
347 sprintf(sbuf1, "In MPI process #%d: ", cr->nodeid);
352 if (nthread_local > 1)
354 sprintf(sbuf2, "for %d/%d thread%s ",
355 nthread_local - nth_affinity_set, nthread_local,
356 nthread_local > 1 ? "s" : "");
359 md_print_warn(NULL, fplog,
360 "WARNING: %sAffinity setting %sfailed.\n"
361 " This can cause performance degradation! If you think your setting are\n"
362 " correct, contact the GROMACS developers.",
369 /* Check the process affinity mask and if it is found to be non-zero,
370 * will honor it and disable mdrun internal affinity setting.
371 * Note that this will only work on Linux as we use a GNU feature.
374 gmx_check_thread_affinity_set(FILE gmx_unused *fplog,
375 const t_commrec gmx_unused *cr,
376 gmx_hw_opt_t gmx_unused *hw_opt,
377 int gmx_unused nthreads_hw_avail,
378 gmx_bool gmx_unused bAfterOpenmpInit)
380 #ifdef HAVE_SCHED_AFFINITY
381 cpu_set_t mask_current;
382 int i, ret, cpu_count, cpu_set;
386 if (hw_opt->thread_affinity == threadaffOFF)
388 /* internal affinity setting is off, don't bother checking process affinity */
392 CPU_ZERO(&mask_current);
393 if ((ret = sched_getaffinity(0, sizeof(cpu_set_t), &mask_current)) != 0)
395 /* failed to query affinity mask, will just return */
398 fprintf(debug, "Failed to query affinity mask (error %d)", ret);
403 /* Before proceeding with the actual check, make sure that the number of
404 * detected CPUs is >= the CPUs in the current set.
405 * We need to check for CPU_COUNT as it was added only in glibc 2.6. */
407 if (nthreads_hw_avail < CPU_COUNT(&mask_current))
411 fprintf(debug, "%d hardware threads detected, but %d was returned by CPU_COUNT",
412 nthreads_hw_avail, CPU_COUNT(&mask_current));
416 #endif /* CPU_COUNT */
419 for (i = 0; (i < nthreads_hw_avail && i < CPU_SETSIZE); i++)
421 bAllSet = bAllSet && (CPU_ISSET(i, &mask_current) != 0);
426 if (hw_opt->thread_affinity == threadaffAUTO)
428 if (!bAfterOpenmpInit)
430 md_print_warn(cr, fplog,
431 "Non-default thread affinity set, disabling internal thread affinity");
435 md_print_warn(cr, fplog,
436 "Non-default thread affinity set probably by the OpenMP library,\n"
437 "disabling internal thread affinity");
439 hw_opt->thread_affinity = threadaffOFF;
443 /* Only warn once, at the last check (bAfterOpenmpInit==TRUE) */
444 if (bAfterOpenmpInit)
446 md_print_warn(cr, fplog,
447 "Overriding thread affinity set outside %s\n",
454 fprintf(debug, "Non-default affinity mask found\n");
461 fprintf(debug, "Default affinity mask found\n");
464 #endif /* HAVE_SCHED_AFFINITY */