2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
5 * Copyright (c) 2001-2004, The GROMACS development team.
6 * Copyright (c) 2013,2014,2015,2016,2017, by the GROMACS development team, led by
7 * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
8 * and including many others, as listed in the AUTHORS file in the
9 * top-level source directory and at http://www.gromacs.org.
11 * GROMACS is free software; you can redistribute it and/or
12 * modify it under the terms of the GNU Lesser General Public License
13 * as published by the Free Software Foundation; either version 2.1
14 * of the License, or (at your option) any later version.
16 * GROMACS is distributed in the hope that it will be useful,
17 * but WITHOUT ANY WARRANTY; without even the implied warranty of
18 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
19 * Lesser General Public License for more details.
21 * You should have received a copy of the GNU Lesser General Public
22 * License along with GROMACS; if not, see
23 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
24 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
26 * If you want to redistribute modifications to GROMACS, please
27 * consider that scientific software is very special. Version
28 * control is crucial - bugs must be traceable. We will be happy to
29 * consider code for inclusion in the official distribution, but
30 * derived work must not be called official GROMACS. Details are found
31 * in the README & COPYING files - if they are missing, get the
32 * official version at http://www.gromacs.org.
34 * To help us fund GROMACS development, we humbly ask that you cite
35 * the research papers on the package. Check out http://www.gromacs.org.
48 #include "gromacs/commandline/filenm.h"
49 #include "gromacs/mdtypes/commrec.h"
50 #include "gromacs/utility/basenetwork.h"
51 #include "gromacs/utility/cstringutil.h"
52 #include "gromacs/utility/fatalerror.h"
53 #include "gromacs/utility/futil.h"
54 #include "gromacs/utility/gmxmpi.h"
55 #include "gromacs/utility/smalloc.h"
57 /* The source code in this file should be thread-safe.
58 Please keep it that way. */
60 void gmx_fill_commrec_from_mpi(t_commrec gmx_unused *cr)
63 gmx_call("gmx_fill_commrec_from_mpi");
65 if (!gmx_mpi_initialized())
67 gmx_comm("MPI has not been initialized properly");
70 cr->nnodes = gmx_node_num();
71 cr->nodeid = gmx_node_rank();
72 // TODO This communicator should be always available. Currently we
73 // make it multiple times, and keep it only when relevant. But the
74 // cost of an extra communicator is negligible in single-node
75 // cases (both thread-MPI and real MPI) case, and we need it in
76 // all multi-node MPI cases with more than one PP rank per node,
77 // with and without GPUs. By always having it available, we also
78 // don't need to protect calls to mpi_comm_physicalnode, etc.
79 if (PAR(cr) || MULTISIM(cr))
81 MPI_Comm_split(MPI_COMM_WORLD, gmx_physicalnode_id_hash(), cr->nodeid, &cr->mpi_comm_physicalnode);
83 cr->sim_nodeid = cr->nodeid;
84 cr->mpi_comm_mysim = MPI_COMM_WORLD;
85 cr->mpi_comm_mygroup = MPI_COMM_WORLD;
90 t_commrec *init_commrec()
97 gmx_fill_commrec_from_mpi(cr);
99 cr->mpi_comm_mysim = nullptr;
100 cr->mpi_comm_mygroup = nullptr;
103 cr->nodeid = cr->sim_nodeid;
106 // TODO cr->duty should not be initialized here
107 cr->duty = (DUTY_PP | DUTY_PME);
109 #if GMX_MPI && !MPI_IN_PLACE_EXISTS
110 /* initialize the MPI_IN_PLACE replacement buffers */
112 cr->mpb->ibuf = NULL;
113 cr->mpb->libuf = NULL;
114 cr->mpb->fbuf = NULL;
115 cr->mpb->dbuf = NULL;
116 cr->mpb->ibuf_alloc = 0;
117 cr->mpb->libuf_alloc = 0;
118 cr->mpb->fbuf_alloc = 0;
119 cr->mpb->dbuf_alloc = 0;
125 static void done_mpi_in_place_buf(mpi_in_place_buf_t *buf)
137 void done_commrec(t_commrec *cr)
140 if (PAR(cr) || MULTISIM(cr))
142 MPI_Comm_free(&cr->mpi_comm_physicalnode);
145 if (nullptr != cr->dd)
148 // done_domdec(cr->dd);
150 if (nullptr != cr->ms)
152 done_mpi_in_place_buf(cr->ms->mpb);
155 done_mpi_in_place_buf(cr->mpb);
159 t_commrec *reinitialize_commrec_for_this_thread(const t_commrec gmx_unused *cro)
164 /* make a thread-specific commrec */
166 /* now copy the whole thing, so settings like the number of PME nodes
170 /* and we start setting our own thread-specific values for things */
171 gmx_fill_commrec_from_mpi(cr);
173 // TODO cr->duty should not be initialized here
174 cr->duty = (DUTY_PP | DUTY_PME);
182 void gmx_setup_nodecomm(FILE gmx_unused *fplog, t_commrec *cr)
186 /* Many MPI implementations do not optimize MPI_Allreduce
187 * (and probably also other global communication calls)
188 * for multi-core nodes connected by a network.
189 * We can optimize such communication by using one MPI call
190 * within each node and one between the nodes.
191 * For MVAPICH2 and Intel MPI this reduces the time for
192 * the global_stat communication by 25%
193 * for 2x2-core 3 GHz Woodcrest connected by mixed DDR/SDR Infiniband.
194 * B. Hess, November 2007
204 MPI_Comm_size(cr->mpi_comm_mygroup, &n);
205 MPI_Comm_rank(cr->mpi_comm_mygroup, &rank);
207 int nodehash = gmx_physicalnode_id_hash();
211 fprintf(debug, "In gmx_setup_nodecomm: splitting communicator of size %d\n", n);
215 /* The intra-node communicator, split on node number */
216 MPI_Comm_split(cr->mpi_comm_mygroup, nodehash, rank, &nc->comm_intra);
217 MPI_Comm_rank(nc->comm_intra, &nc->rank_intra);
220 fprintf(debug, "In gmx_setup_nodecomm: node ID %d rank within node %d\n",
221 rank, nc->rank_intra);
223 /* The inter-node communicator, split on rank_intra.
224 * We actually only need the one for rank=0,
225 * but it is easier to create them all.
227 MPI_Comm_split(cr->mpi_comm_mygroup, nc->rank_intra, rank, &nc->comm_inter);
228 /* Check if this really created two step communication */
231 MPI_Comm_size(nc->comm_inter, &ng);
232 MPI_Comm_size(nc->comm_intra, &ni);
235 fprintf(debug, "In gmx_setup_nodecomm: groups %d, my group size %d\n",
239 if (getenv("GMX_NO_NODECOMM") == nullptr &&
240 ((ng > 1 && ng < n) || (ni > 1 && ni < n)))
245 fprintf(fplog, "Using two step summing over %d groups of on average %.1f ranks\n\n",
246 ng, (real)n/(real)ng);
248 if (nc->rank_intra > 0)
250 MPI_Comm_free(&nc->comm_inter);
255 /* One group or all processes in a separate group, use normal summing */
256 MPI_Comm_free(&nc->comm_inter);
257 MPI_Comm_free(&nc->comm_intra);
260 fprintf(debug, "In gmx_setup_nodecomm: not unsing separate inter- and intra-node communicators.\n");
265 /* tMPI runs only on a single node so just use the nodeid */
266 nc->rank_intra = cr->nodeid;
270 void gmx_init_intranode_counters(t_commrec *cr)
272 /* counters for PP+PME and PP-only processes on my physical node */
273 int nrank_intranode, rank_intranode;
274 int nrank_pp_intranode, rank_pp_intranode;
275 /* thread-MPI is not initialized when not running in parallel */
276 #if GMX_MPI && !GMX_THREAD_MPI
277 int nrank_world, rank_world;
278 int i, myhash, *hash, *hash_s, *hash_pp, *hash_pp_s;
280 MPI_Comm_size(MPI_COMM_WORLD, &nrank_world);
281 MPI_Comm_rank(MPI_COMM_WORLD, &rank_world);
283 /* Get a (hopefully unique) hash that identifies our physical node */
284 myhash = gmx_physicalnode_id_hash();
286 /* We can't rely on MPI_IN_PLACE, so we need send and receive buffers */
287 snew(hash, nrank_world);
288 snew(hash_s, nrank_world);
289 snew(hash_pp, nrank_world);
290 snew(hash_pp_s, nrank_world);
292 hash_s[rank_world] = myhash;
293 hash_pp_s[rank_world] = thisRankHasDuty(cr, DUTY_PP) ? myhash : -1;
295 MPI_Allreduce(hash_s, hash, nrank_world, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
296 MPI_Allreduce(hash_pp_s, hash_pp, nrank_world, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
300 nrank_pp_intranode = 0;
301 rank_pp_intranode = 0;
302 for (i = 0; i < nrank_world; i++)
304 if (hash[i] == myhash)
312 if (hash_pp[i] == myhash)
314 nrank_pp_intranode++;
315 if (thisRankHasDuty(cr, DUTY_PP) && i < rank_world)
326 /* Serial or thread-MPI code: we run within a single physical node */
327 nrank_intranode = cr->nnodes;
328 rank_intranode = cr->sim_nodeid;
329 nrank_pp_intranode = cr->nnodes - cr->npmenodes;
330 rank_pp_intranode = cr->nodeid;
336 if (thisRankHasDuty(cr, DUTY_PP) && thisRankHasDuty(cr, DUTY_PME))
338 sprintf(sbuf, "PP+PME");
342 sprintf(sbuf, "%s", thisRankHasDuty(cr, DUTY_PP) ? "PP" : "PME");
344 fprintf(debug, "On %3s rank %d: nrank_intranode=%d, rank_intranode=%d, "
345 "nrank_pp_intranode=%d, rank_pp_intranode=%d\n",
346 sbuf, cr->sim_nodeid,
347 nrank_intranode, rank_intranode,
348 nrank_pp_intranode, rank_pp_intranode);
351 cr->nrank_intranode = nrank_intranode;
352 cr->rank_intranode = rank_intranode;
353 cr->nrank_pp_intranode = nrank_pp_intranode;
354 cr->rank_pp_intranode = rank_pp_intranode;
358 void gmx_barrier(const t_commrec gmx_unused *cr)
361 gmx_call("gmx_barrier");
363 MPI_Barrier(cr->mpi_comm_mygroup);
367 void gmx_barrier_physical_node(const t_commrec gmx_unused *cr)
370 gmx_call("gmx_barrier_physical_node");
372 MPI_Barrier(cr->mpi_comm_physicalnode);
376 void gmx_bcast(int gmx_unused nbytes, void gmx_unused *b, const t_commrec gmx_unused *cr)
379 gmx_call("gmx_bast");
381 MPI_Bcast(b, nbytes, MPI_BYTE, MASTERRANK(cr), cr->mpi_comm_mygroup);
385 void gmx_bcast_sim(int gmx_unused nbytes, void gmx_unused *b, const t_commrec gmx_unused *cr)
388 gmx_call("gmx_bast");
390 MPI_Bcast(b, nbytes, MPI_BYTE, MASTERRANK(cr), cr->mpi_comm_mysim);
394 void gmx_sumd(int gmx_unused nr, double gmx_unused r[], const t_commrec gmx_unused *cr)
397 gmx_call("gmx_sumd");
399 #if MPI_IN_PLACE_EXISTS
402 if (cr->nc.rank_intra == 0)
404 /* Use two step summing. */
405 MPI_Reduce(MPI_IN_PLACE, r, nr, MPI_DOUBLE, MPI_SUM, 0,
407 /* Sum the roots of the internal (intra) buffers. */
408 MPI_Allreduce(MPI_IN_PLACE, r, nr, MPI_DOUBLE, MPI_SUM,
413 /* This is here because of the silly MPI specification
414 that MPI_IN_PLACE should be put in sendbuf instead of recvbuf */
415 MPI_Reduce(r, nullptr, nr, MPI_DOUBLE, MPI_SUM, 0, cr->nc.comm_intra);
417 MPI_Bcast(r, nr, MPI_DOUBLE, 0, cr->nc.comm_intra);
421 MPI_Allreduce(MPI_IN_PLACE, r, nr, MPI_DOUBLE, MPI_SUM,
422 cr->mpi_comm_mygroup);
427 if (nr > cr->mpb->dbuf_alloc)
429 cr->mpb->dbuf_alloc = nr;
430 srenew(cr->mpb->dbuf, cr->mpb->dbuf_alloc);
434 /* Use two step summing */
435 MPI_Allreduce(r, cr->mpb->dbuf, nr, MPI_DOUBLE, MPI_SUM, cr->nc.comm_intra);
436 if (cr->nc.rank_intra == 0)
438 /* Sum with the buffers reversed */
439 MPI_Allreduce(cr->mpb->dbuf, r, nr, MPI_DOUBLE, MPI_SUM,
442 MPI_Bcast(r, nr, MPI_DOUBLE, 0, cr->nc.comm_intra);
446 MPI_Allreduce(r, cr->mpb->dbuf, nr, MPI_DOUBLE, MPI_SUM,
447 cr->mpi_comm_mygroup);
448 for (i = 0; i < nr; i++)
450 r[i] = cr->mpb->dbuf[i];
457 void gmx_sumf(int gmx_unused nr, float gmx_unused r[], const t_commrec gmx_unused *cr)
460 gmx_call("gmx_sumf");
462 #if MPI_IN_PLACE_EXISTS
465 /* Use two step summing. */
466 if (cr->nc.rank_intra == 0)
468 MPI_Reduce(MPI_IN_PLACE, r, nr, MPI_FLOAT, MPI_SUM, 0,
470 /* Sum the roots of the internal (intra) buffers */
471 MPI_Allreduce(MPI_IN_PLACE, r, nr, MPI_FLOAT, MPI_SUM,
476 /* This is here because of the silly MPI specification
477 that MPI_IN_PLACE should be put in sendbuf instead of recvbuf */
478 MPI_Reduce(r, nullptr, nr, MPI_FLOAT, MPI_SUM, 0, cr->nc.comm_intra);
480 MPI_Bcast(r, nr, MPI_FLOAT, 0, cr->nc.comm_intra);
484 MPI_Allreduce(MPI_IN_PLACE, r, nr, MPI_FLOAT, MPI_SUM, cr->mpi_comm_mygroup);
489 if (nr > cr->mpb->fbuf_alloc)
491 cr->mpb->fbuf_alloc = nr;
492 srenew(cr->mpb->fbuf, cr->mpb->fbuf_alloc);
496 /* Use two step summing */
497 MPI_Allreduce(r, cr->mpb->fbuf, nr, MPI_FLOAT, MPI_SUM, cr->nc.comm_intra);
498 if (cr->nc.rank_intra == 0)
500 /* Sum with the buffers reversed */
501 MPI_Allreduce(cr->mpb->fbuf, r, nr, MPI_FLOAT, MPI_SUM,
504 MPI_Bcast(r, nr, MPI_FLOAT, 0, cr->nc.comm_intra);
508 MPI_Allreduce(r, cr->mpb->fbuf, nr, MPI_FLOAT, MPI_SUM,
509 cr->mpi_comm_mygroup);
510 for (i = 0; i < nr; i++)
512 r[i] = cr->mpb->fbuf[i];
519 void gmx_sumi(int gmx_unused nr, int gmx_unused r[], const t_commrec gmx_unused *cr)
522 gmx_call("gmx_sumi");
524 #if MPI_IN_PLACE_EXISTS
527 /* Use two step summing */
528 if (cr->nc.rank_intra == 0)
530 MPI_Reduce(MPI_IN_PLACE, r, nr, MPI_INT, MPI_SUM, 0, cr->nc.comm_intra);
531 /* Sum with the buffers reversed */
532 MPI_Allreduce(MPI_IN_PLACE, r, nr, MPI_INT, MPI_SUM, cr->nc.comm_inter);
536 /* This is here because of the silly MPI specification
537 that MPI_IN_PLACE should be put in sendbuf instead of recvbuf */
538 MPI_Reduce(r, nullptr, nr, MPI_INT, MPI_SUM, 0, cr->nc.comm_intra);
540 MPI_Bcast(r, nr, MPI_INT, 0, cr->nc.comm_intra);
544 MPI_Allreduce(MPI_IN_PLACE, r, nr, MPI_INT, MPI_SUM, cr->mpi_comm_mygroup);
549 if (nr > cr->mpb->ibuf_alloc)
551 cr->mpb->ibuf_alloc = nr;
552 srenew(cr->mpb->ibuf, cr->mpb->ibuf_alloc);
556 /* Use two step summing */
557 MPI_Allreduce(r, cr->mpb->ibuf, nr, MPI_INT, MPI_SUM, cr->nc.comm_intra);
558 if (cr->nc.rank_intra == 0)
560 /* Sum with the buffers reversed */
561 MPI_Allreduce(cr->mpb->ibuf, r, nr, MPI_INT, MPI_SUM, cr->nc.comm_inter);
563 MPI_Bcast(r, nr, MPI_INT, 0, cr->nc.comm_intra);
567 MPI_Allreduce(r, cr->mpb->ibuf, nr, MPI_INT, MPI_SUM, cr->mpi_comm_mygroup);
568 for (i = 0; i < nr; i++)
570 r[i] = cr->mpb->ibuf[i];
577 void gmx_sumli(int gmx_unused nr, gmx_int64_t gmx_unused r[], const t_commrec gmx_unused *cr)
580 gmx_call("gmx_sumli");
582 #if MPI_IN_PLACE_EXISTS
585 /* Use two step summing */
586 if (cr->nc.rank_intra == 0)
588 MPI_Reduce(MPI_IN_PLACE, r, nr, MPI_INT64_T, MPI_SUM, 0,
590 /* Sum with the buffers reversed */
591 MPI_Allreduce(MPI_IN_PLACE, r, nr, MPI_INT64_T, MPI_SUM,
596 /* This is here because of the silly MPI specification
597 that MPI_IN_PLACE should be put in sendbuf instead of recvbuf */
598 MPI_Reduce(r, nullptr, nr, MPI_INT64_T, MPI_SUM, 0, cr->nc.comm_intra);
600 MPI_Bcast(r, nr, MPI_INT64_T, 0, cr->nc.comm_intra);
604 MPI_Allreduce(MPI_IN_PLACE, r, nr, MPI_INT64_T, MPI_SUM, cr->mpi_comm_mygroup);
609 if (nr > cr->mpb->libuf_alloc)
611 cr->mpb->libuf_alloc = nr;
612 srenew(cr->mpb->libuf, cr->mpb->libuf_alloc);
616 /* Use two step summing */
617 MPI_Allreduce(r, cr->mpb->libuf, nr, MPI_INT64_T, MPI_SUM,
619 if (cr->nc.rank_intra == 0)
621 /* Sum with the buffers reversed */
622 MPI_Allreduce(cr->mpb->libuf, r, nr, MPI_INT64_T, MPI_SUM,
625 MPI_Bcast(r, nr, MPI_INT64_T, 0, cr->nc.comm_intra);
629 MPI_Allreduce(r, cr->mpb->libuf, nr, MPI_INT64_T, MPI_SUM,
630 cr->mpi_comm_mygroup);
631 for (i = 0; i < nr; i++)
633 r[i] = cr->mpb->libuf[i];
643 static void gmx_sumd_comm(int nr, double r[], MPI_Comm mpi_comm)
645 #if MPI_IN_PLACE_EXISTS
646 MPI_Allreduce(MPI_IN_PLACE, r, nr, MPI_DOUBLE, MPI_SUM, mpi_comm);
648 /* this function is only used in code that is not performance critical,
649 (during setup, when comm_rec is not the appropriate communication
650 structure), so this isn't as bad as it looks. */
655 MPI_Allreduce(r, buf, nr, MPI_DOUBLE, MPI_SUM, mpi_comm);
656 for (i = 0; i < nr; i++)
666 static void gmx_sumf_comm(int nr, float r[], MPI_Comm mpi_comm)
668 #if MPI_IN_PLACE_EXISTS
669 MPI_Allreduce(MPI_IN_PLACE, r, nr, MPI_FLOAT, MPI_SUM, mpi_comm);
671 /* this function is only used in code that is not performance critical,
672 (during setup, when comm_rec is not the appropriate communication
673 structure), so this isn't as bad as it looks. */
678 MPI_Allreduce(r, buf, nr, MPI_FLOAT, MPI_SUM, mpi_comm);
679 for (i = 0; i < nr; i++)
688 void gmx_sumd_sim(int gmx_unused nr, double gmx_unused r[], const gmx_multisim_t gmx_unused *ms)
691 gmx_call("gmx_sumd_sim");
693 gmx_sumd_comm(nr, r, ms->mpi_comm_masters);
697 void gmx_sumf_sim(int gmx_unused nr, float gmx_unused r[], const gmx_multisim_t gmx_unused *ms)
700 gmx_call("gmx_sumf_sim");
702 gmx_sumf_comm(nr, r, ms->mpi_comm_masters);
706 void gmx_sumi_sim(int gmx_unused nr, int gmx_unused r[], const gmx_multisim_t gmx_unused *ms)
709 gmx_call("gmx_sumi_sim");
711 #if MPI_IN_PLACE_EXISTS
712 MPI_Allreduce(MPI_IN_PLACE, r, nr, MPI_INT, MPI_SUM, ms->mpi_comm_masters);
714 /* this is thread-unsafe, but it will do for now: */
717 if (nr > ms->mpb->ibuf_alloc)
719 ms->mpb->ibuf_alloc = nr;
720 srenew(ms->mpb->ibuf, ms->mpb->ibuf_alloc);
722 MPI_Allreduce(r, ms->mpb->ibuf, nr, MPI_INT, MPI_SUM, ms->mpi_comm_masters);
723 for (i = 0; i < nr; i++)
725 r[i] = ms->mpb->ibuf[i];
731 void gmx_sumli_sim(int gmx_unused nr, gmx_int64_t gmx_unused r[], const gmx_multisim_t gmx_unused *ms)
734 gmx_call("gmx_sumli_sim");
736 #if MPI_IN_PLACE_EXISTS
737 MPI_Allreduce(MPI_IN_PLACE, r, nr, MPI_INT64_T, MPI_SUM,
738 ms->mpi_comm_masters);
740 /* this is thread-unsafe, but it will do for now: */
743 if (nr > ms->mpb->libuf_alloc)
745 ms->mpb->libuf_alloc = nr;
746 srenew(ms->mpb->libuf, ms->mpb->libuf_alloc);
748 MPI_Allreduce(r, ms->mpb->libuf, nr, MPI_INT64_T, MPI_SUM,
749 ms->mpi_comm_masters);
750 for (i = 0; i < nr; i++)
752 r[i] = ms->mpb->libuf[i];
758 const char *opt2fn_master(const char *opt, int nfile, const t_filenm fnm[],
761 return SIMMASTER(cr) ? opt2fn(opt, nfile, fnm) : nullptr;
764 void gmx_fatal_collective(int f_errno, const char *file, int line,
765 MPI_Comm comm, gmx_bool bMaster,
766 const char *fmt, ...)
772 /* Check if we are calling on all processes in MPI_COMM_WORLD */
773 MPI_Comm_compare(comm, MPI_COMM_WORLD, &result);
774 /* Any result except MPI_UNEQUAL allows us to call MPI_Finalize */
775 bFinalize = (result != MPI_UNEQUAL);
777 GMX_UNUSED_VALUE(comm);
782 gmx_fatal_mpi_va(f_errno, file, line, bMaster, bFinalize, fmt, ap);