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.
45 #include "gromacs/utility/fatalerror.h"
47 #include "types/commrec.h"
51 #include "gmx_omp_nthreads.h"
52 #include "md_logging.h"
54 #include "gromacs/utility/gmxomp.h"
56 /** Structure with the number of threads for each OpenMP multi-threaded
57 * algorithmic module in mdrun. */
60 int gnth; /**< Global num. of threads per PP or PP+PME process/tMPI thread. */
61 int gnth_pme; /**< Global num. of threads per PME only process/tMPI thread. */
63 int nth[emntNR]; /**< Number of threads for each module, indexed with module_nth_t */
64 gmx_bool initialized; /**< TRUE if the module as been initialized. */
65 } omp_module_nthreads_t;
67 /** Names of environment variables to set the per module number of threads.
69 * Indexed with the values of module_nth_t.
71 static const char *modth_env_var[emntNR] =
73 "GMX_DEFAULT_NUM_THREADS should never be set",
74 "GMX_DOMDEC_NUM_THREADS", "GMX_PAIRSEARCH_NUM_THREADS",
75 "GMX_NONBONDED_NUM_THREADS", "GMX_BONDED_NUM_THREADS",
76 "GMX_PME_NUM_THREADS", "GMX_UPDATE_NUM_THREADS",
77 "GMX_VSITE_NUM_THREADS",
78 "GMX_LINCS_NUM_THREADS", "GMX_SETTLE_NUM_THREADS"
81 /** Names of the modules. */
82 static const char *mod_name[emntNR] =
84 "default", "domain decomposition", "pair search", "non-bonded",
85 "bonded", "PME", "update", "LINCS", "SETTLE"
88 /** Number of threads for each algorithmic module.
90 * File-scope global variable that gets set once in pick_module_nthreads()
91 * and queried via gmx_omp_nthreads_get().
93 * All fields are initialized to 0 which should result in errors if
94 * the init call is omitted.
96 static omp_module_nthreads_t modth = { 0, 0, {0, 0, 0, 0, 0, 0, 0, 0, 0}, FALSE};
99 /** Determine the number of threads for module \p mod.
101 * \p m takes values form the module_nth_t enum and maps these to the
102 * corresponding value in modth_env_var.
104 * Each number of threads per module takes the default value unless
105 * GMX_*_NUM_THERADS env var is set, case in which its value overrides
108 * The "group" scheme supports OpenMP only in PME and in thise case all but
109 * the PME nthread values default to 1.
111 static void pick_module_nthreads(FILE *fplog, int m,
113 gmx_bool bFullOmpSupport,
125 #endif /* GMX_OPENMP */
127 /* The default should never be set through a GMX_*_NUM_THREADS env var
128 * as it's always equal with gnth. */
129 if (m == emntDefault)
134 /* check the environment variable */
135 if ((env = getenv(modth_env_var[m])) != NULL)
137 sscanf(env, "%d", &nth);
141 gmx_warning("%s=%d is set, but %s is compiled without OpenMP!",
142 modth_env_var[m], nth, ShortProgram());
145 /* with the verlet codepath, when any GMX_*_NUM_THREADS env var is set,
146 * OMP_NUM_THREADS also has to be set */
147 if (bFullOmpSupport && getenv("OMP_NUM_THREADS") == NULL)
149 gmx_fatal(FARGS, "%s=%d is set, the default number of threads also "
150 "needs to be set with OMP_NUM_THREADS!",
151 modth_env_var[m], nth);
154 /* with the group scheme warn if any env var except PME is set */
155 if (!bFullOmpSupport)
159 gmx_warning("%s=%d is set, but OpenMP multithreading is not "
161 modth_env_var[m], nth, mod_name[m]);
166 /* only babble if we are really overriding with a different value */
167 if ((bSepPME && m == emntPME && nth != modth.gnth_pme) || (nth != modth.gnth))
169 sprintf(sbuf, "%s=%d set, overriding the default number of %s threads",
170 modth_env_var[m], nth, mod_name[m]);
173 fprintf(stderr, "\n%s\n", sbuf);
177 fprintf(fplog, "%s\n", sbuf);
183 /* pick the global PME node nthreads if we are setting the number
184 * of threads in separate PME nodes */
185 nth = (bSepPME && m == emntPME) ? modth.gnth_pme : modth.gnth;
188 gmx_omp_nthreads_set(m, nth);
191 void gmx_omp_nthreads_read_env(int *nthreads_omp,
192 gmx_bool bIsSimMaster)
195 gmx_bool bCommandLineSetNthreadsOMP = *nthreads_omp > 0;
198 assert(nthreads_omp);
200 if ((env = getenv("OMP_NUM_THREADS")) != NULL)
204 sscanf(env, "%d", &nt_omp);
207 gmx_fatal(FARGS, "OMP_NUM_THREADS is invalid: '%s'", env);
210 if (bCommandLineSetNthreadsOMP && nt_omp != *nthreads_omp)
212 gmx_fatal(FARGS, "Environment variable OMP_NUM_THREADS (%d) and the number of threads requested on the command line (%d) have different values. Either omit one, or set them both to the same value.", nt_omp, *nthreads_omp);
215 /* Setting the number of OpenMP threads. */
216 *nthreads_omp = nt_omp;
218 /* Output the results */
220 "The number of OpenMP threads was set by environment variable OMP_NUM_THREADS to %d%s\n",
222 bCommandLineSetNthreadsOMP ? " (and the command-line setting agreed with that)" : "");
225 /* This prints once per simulation for multi-simulations,
226 * which might help diagnose issues with inhomogenous
228 fputs(buffer, stderr);
232 /* This prints once per process for real MPI (i.e. once
233 * per debug file), and once per simulation for thread MPI
234 * (because of logic in the calling function). */
235 fputs(buffer, debug);
240 void gmx_omp_nthreads_init(FILE *fplog, t_commrec *cr,
241 int nthreads_hw_avail,
242 int omp_nthreads_req,
243 int omp_nthreads_pme_req,
244 gmx_bool gmx_unused bThisNodePMEOnly,
245 gmx_bool bFullOmpSupport)
247 int nth, nth_pmeonly, gmx_maxth, nppn;
249 gmx_bool bSepPME, bOMP;
255 #endif /* GMX_OPENMP */
257 /* number of MPI processes/threads per physical node */
258 nppn = cr->nrank_intranode;
260 bSepPME = ( (cr->duty & DUTY_PP) && !(cr->duty & DUTY_PME)) ||
261 (!(cr->duty & DUTY_PP) && (cr->duty & DUTY_PME));
263 #ifdef GMX_THREAD_MPI
264 /* modth is shared among tMPI threads, so for thread safety do the
265 * detection is done on the master only. It is not thread-safe with
266 * multiple simulations, but that's anyway not supported by tMPI. */
270 /* just return if the initialization has already been done */
271 if (modth.initialized)
276 /* With full OpenMP support (verlet scheme) set the number of threads
277 * per process / default:
278 * - 1 if not compiled with OpenMP or
279 * - OMP_NUM_THREADS if the env. var is set, or
280 * - omp_nthreads_req = #of threads requested by the user on the mdrun
281 * command line, otherwise
282 * - take the max number of available threads and distribute them
283 * on the processes/tMPI threads.
284 * ~ The GMX_*_NUM_THREADS env var overrides the number of threads of
285 * the respective module and it has to be used in conjunction with
288 * With the group scheme OpenMP multithreading is only supported in PME,
289 * for all other modules nthreads is set to 1.
290 * The number of PME threads is equal to:
291 * - 1 if not compiled with OpenMP or
292 * - GMX_PME_NUM_THREADS if defined, otherwise
293 * - OMP_NUM_THREADS if defined, otherwise
297 if ((env = getenv("OMP_NUM_THREADS")) != NULL)
299 if (!bOMP && (strncmp(env, "1", 1) != 0))
301 gmx_warning("OMP_NUM_THREADS is set, but %s was compiled without OpenMP support!",
306 nth = gmx_omp_get_max_threads();
309 else if (omp_nthreads_req > 0)
311 nth = omp_nthreads_req;
313 else if (bFullOmpSupport && bOMP)
315 /* max available threads per node */
316 nth = nthreads_hw_avail;
318 /* divide the threads among the MPI processes/tMPI threads */
329 /* now we have the global values, set them:
330 * - 1 if not compiled with OpenMP and for the group scheme
331 * - nth for the verlet scheme when compiled with OpenMP
333 if (bFullOmpSupport && bOMP)
344 if (omp_nthreads_pme_req > 0)
346 modth.gnth_pme = omp_nthreads_pme_req;
350 modth.gnth_pme = nth;
358 /* now set the per-module values */
359 modth.nth[emntDefault] = modth.gnth;
360 pick_module_nthreads(fplog, emntDomdec, SIMMASTER(cr), bFullOmpSupport, bSepPME);
361 pick_module_nthreads(fplog, emntPairsearch, SIMMASTER(cr), bFullOmpSupport, bSepPME);
362 pick_module_nthreads(fplog, emntNonbonded, SIMMASTER(cr), bFullOmpSupport, bSepPME);
363 pick_module_nthreads(fplog, emntBonded, SIMMASTER(cr), bFullOmpSupport, bSepPME);
364 pick_module_nthreads(fplog, emntPME, SIMMASTER(cr), bFullOmpSupport, bSepPME);
365 pick_module_nthreads(fplog, emntUpdate, SIMMASTER(cr), bFullOmpSupport, bSepPME);
366 pick_module_nthreads(fplog, emntVSITE, SIMMASTER(cr), bFullOmpSupport, bSepPME);
367 pick_module_nthreads(fplog, emntLINCS, SIMMASTER(cr), bFullOmpSupport, bSepPME);
368 pick_module_nthreads(fplog, emntSETTLE, SIMMASTER(cr), bFullOmpSupport, bSepPME);
370 /* set the number of threads globally */
373 #ifndef GMX_THREAD_MPI
374 if (bThisNodePMEOnly)
376 gmx_omp_set_num_threads(modth.gnth_pme);
379 #endif /* GMX_THREAD_MPI */
383 gmx_omp_set_num_threads(nth);
387 gmx_omp_set_num_threads(1);
392 modth.initialized = TRUE;
394 #ifdef GMX_THREAD_MPI
395 /* Non-master threads have to wait for the detection to be done. */
398 MPI_Barrier(cr->mpi_comm_mysim);
402 /* inform the user about the settings */
405 #ifdef GMX_THREAD_MPI
406 const char *mpi_str = "per tMPI thread";
408 const char *mpi_str = "per MPI process";
411 /* for group scheme we print PME threads info only */
414 md_print_info(cr, fplog, "Using %d OpenMP thread%s %s\n",
415 modth.gnth, modth.gnth > 1 ? "s" : "",
416 cr->nnodes > 1 ? mpi_str : "");
418 if (bSepPME && modth.gnth_pme != modth.gnth)
420 md_print_info(cr, fplog, "Using %d OpenMP thread%s %s for PME\n",
421 modth.gnth_pme, modth.gnth_pme > 1 ? "s" : "",
422 cr->nnodes > 1 ? mpi_str : "");
426 /* detect and warn about oversubscription
427 * TODO: enable this for separate PME nodes as well! */
428 if (!bSepPME && cr->rank_pp_intranode == 0)
430 char sbuf[STRLEN], sbuf1[STRLEN], sbuf2[STRLEN];
432 if (modth.gnth*nppn > nthreads_hw_avail)
434 sprintf(sbuf, "threads");
440 #ifdef GMX_THREAD_MPI
441 sprintf(sbuf, "thread-MPI threads");
443 sprintf(sbuf, "MPI processes");
444 sprintf(sbuf1, " per node");
445 sprintf(sbuf2, "On node %d: o", cr->sim_nodeid);
449 md_print_warn(cr, fplog,
450 "WARNING: %sversubscribing the available %d logical CPU cores%s with %d %s.\n"
451 " This will cause considerable performance loss!",
452 sbuf2, nthreads_hw_avail, sbuf1, nppn*modth.gnth, sbuf);
457 int gmx_omp_nthreads_get(int mod)
459 if (mod < 0 || mod >= emntNR)
461 /* invalid module queried */
466 return modth.nth[mod];
471 gmx_omp_nthreads_set(int mod, int nthreads)
473 /* Catch an attempt to set the number of threads on an invalid
475 assert(mod >= 0 && mod < emntNR);
477 modth.nth[mod] = nthreads;