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) 2013,2014,2015,2016,2017,2018, 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 *cr,
61 const gmx_multisim_t *ms)
64 gmx_call("gmx_fill_commrec_from_mpi");
68 if (!gmx_mpi_initialized())
70 gmx_comm("MPI has not been initialized properly");
73 cr->nnodes = gmx_node_num();
74 cr->nodeid = gmx_node_rank();
75 // TODO This communicator should be always available. Currently we
76 // make it multiple times, and keep it only when relevant. But the
77 // cost of an extra communicator is negligible in single-node
78 // cases (both thread-MPI and real MPI) case, and we need it in
79 // all multi-node MPI cases with more than one PP rank per node,
80 // with and without GPUs. By always having it available, we also
81 // don't need to protect calls to mpi_comm_physicalnode, etc.
82 if (PAR(cr) || isMultiSim(ms))
84 MPI_Comm_split(MPI_COMM_WORLD, gmx_physicalnode_id_hash(), cr->nodeid, &cr->mpi_comm_physicalnode);
86 cr->sim_nodeid = cr->nodeid;
87 cr->mpi_comm_mysim = MPI_COMM_WORLD;
88 cr->mpi_comm_mygroup = MPI_COMM_WORLD;
93 t_commrec *init_commrec()
99 cr->mpi_comm_physicalnode = MPI_COMM_NULL;
101 gmx_fill_commrec_from_mpi(cr, nullptr);
103 cr->mpi_comm_mysim = MPI_COMM_NULL;
104 cr->mpi_comm_mygroup = MPI_COMM_NULL;
107 cr->nodeid = cr->sim_nodeid;
110 // TODO cr->duty should not be initialized here
111 cr->duty = (DUTY_PP | DUTY_PME);
113 #if GMX_MPI && !MPI_IN_PLACE_EXISTS
114 /* initialize the MPI_IN_PLACE replacement buffers */
116 cr->mpb->ibuf = NULL;
117 cr->mpb->libuf = NULL;
118 cr->mpb->fbuf = NULL;
119 cr->mpb->dbuf = NULL;
120 cr->mpb->ibuf_alloc = 0;
121 cr->mpb->libuf_alloc = 0;
122 cr->mpb->fbuf_alloc = 0;
123 cr->mpb->dbuf_alloc = 0;
129 void done_mpi_in_place_buf(mpi_in_place_buf_t *buf)
141 void done_commrec(t_commrec *cr)
144 if (cr->mpi_comm_physicalnode != MPI_COMM_NULL)
146 MPI_Comm_free(&cr->mpi_comm_physicalnode);
149 if (nullptr != cr->dd)
152 // done_domdec(cr->dd);
154 done_mpi_in_place_buf(cr->mpb);
158 t_commrec *reinitialize_commrec_for_this_thread(const t_commrec *cro,
159 const gmx_multisim_t *ms)
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, ms);
173 // TODO cr->duty should not be initialized here
174 cr->duty = (DUTY_PP | DUTY_PME);
178 GMX_UNUSED_VALUE(cro);
179 GMX_UNUSED_VALUE(ms);
184 void gmx_setup_nodecomm(FILE gmx_unused *fplog, t_commrec *cr)
188 /* Many MPI implementations do not optimize MPI_Allreduce
189 * (and probably also other global communication calls)
190 * for multi-core nodes connected by a network.
191 * We can optimize such communication by using one MPI call
192 * within each node and one between the nodes.
193 * For MVAPICH2 and Intel MPI this reduces the time for
194 * the global_stat communication by 25%
195 * for 2x2-core 3 GHz Woodcrest connected by mixed DDR/SDR Infiniband.
196 * B. Hess, November 2007
206 MPI_Comm_size(cr->mpi_comm_mygroup, &n);
207 MPI_Comm_rank(cr->mpi_comm_mygroup, &rank);
209 int nodehash = gmx_physicalnode_id_hash();
213 fprintf(debug, "In gmx_setup_nodecomm: splitting communicator of size %d\n", n);
217 /* The intra-node communicator, split on node number */
218 MPI_Comm_split(cr->mpi_comm_mygroup, nodehash, rank, &nc->comm_intra);
219 MPI_Comm_rank(nc->comm_intra, &nc->rank_intra);
222 fprintf(debug, "In gmx_setup_nodecomm: node ID %d rank within node %d\n",
223 rank, nc->rank_intra);
225 /* The inter-node communicator, split on rank_intra.
226 * We actually only need the one for rank=0,
227 * but it is easier to create them all.
229 MPI_Comm_split(cr->mpi_comm_mygroup, nc->rank_intra, rank, &nc->comm_inter);
230 /* Check if this really created two step communication */
233 MPI_Comm_size(nc->comm_inter, &ng);
234 MPI_Comm_size(nc->comm_intra, &ni);
237 fprintf(debug, "In gmx_setup_nodecomm: groups %d, my group size %d\n",
241 if (getenv("GMX_NO_NODECOMM") == nullptr &&
242 ((ng > 1 && ng < n) || (ni > 1 && ni < n)))
247 fprintf(fplog, "Using two step summing over %d groups of on average %.1f ranks\n\n",
248 ng, (real)n/(real)ng);
250 if (nc->rank_intra > 0)
252 MPI_Comm_free(&nc->comm_inter);
257 /* One group or all processes in a separate group, use normal summing */
258 MPI_Comm_free(&nc->comm_inter);
259 MPI_Comm_free(&nc->comm_intra);
262 fprintf(debug, "In gmx_setup_nodecomm: not unsing separate inter- and intra-node communicators.\n");
267 /* tMPI runs only on a single node so just use the nodeid */
268 nc->rank_intra = cr->nodeid;
272 void gmx_init_intranode_counters(t_commrec *cr)
274 /* counters for PP+PME and PP-only processes on my physical node */
275 int nrank_intranode, rank_intranode;
276 /* thread-MPI is not initialized when not running in parallel */
277 #if GMX_MPI && !GMX_THREAD_MPI
278 int nrank_world, rank_world;
279 int i, myhash, *hash, *hash_s, *hash_pp, *hash_pp_s;
281 MPI_Comm_size(MPI_COMM_WORLD, &nrank_world);
282 MPI_Comm_rank(MPI_COMM_WORLD, &rank_world);
284 /* Get a (hopefully unique) hash that identifies our physical node */
285 myhash = gmx_physicalnode_id_hash();
287 /* We can't rely on MPI_IN_PLACE, so we need send and receive buffers */
288 snew(hash, nrank_world);
289 snew(hash_s, nrank_world);
290 snew(hash_pp, nrank_world);
291 snew(hash_pp_s, nrank_world);
293 hash_s[rank_world] = myhash;
294 hash_pp_s[rank_world] = thisRankHasDuty(cr, DUTY_PP) ? myhash : -1;
296 MPI_Allreduce(hash_s, hash, nrank_world, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
297 MPI_Allreduce(hash_pp_s, hash_pp, nrank_world, MPI_INT, MPI_SUM, MPI_COMM_WORLD);
301 for (i = 0; i < nrank_world; i++)
303 if (hash[i] == myhash)
317 /* Serial or thread-MPI code: we run within a single physical node */
318 nrank_intranode = cr->nnodes;
319 rank_intranode = cr->sim_nodeid;
325 if (thisRankHasDuty(cr, DUTY_PP) && thisRankHasDuty(cr, DUTY_PME))
327 sprintf(sbuf, "PP+PME");
331 sprintf(sbuf, "%s", thisRankHasDuty(cr, DUTY_PP) ? "PP" : "PME");
333 fprintf(debug, "On %3s rank %d: nrank_intranode=%d, rank_intranode=%d\n",
334 sbuf, cr->sim_nodeid,
335 nrank_intranode, rank_intranode);
338 cr->nrank_intranode = nrank_intranode;
339 cr->rank_intranode = rank_intranode;
343 void gmx_barrier(const t_commrec gmx_unused *cr)
346 gmx_call("gmx_barrier");
348 MPI_Barrier(cr->mpi_comm_mygroup);
352 void gmx_barrier_physical_node(const t_commrec gmx_unused *cr)
355 gmx_call("gmx_barrier_physical_node");
357 MPI_Barrier(cr->mpi_comm_physicalnode);
361 void gmx_bcast(int gmx_unused nbytes, void gmx_unused *b, const t_commrec gmx_unused *cr)
364 gmx_call("gmx_bast");
366 MPI_Bcast(b, nbytes, MPI_BYTE, MASTERRANK(cr), cr->mpi_comm_mygroup);
370 void gmx_bcast_sim(int gmx_unused nbytes, void gmx_unused *b, const t_commrec gmx_unused *cr)
373 gmx_call("gmx_bast");
375 MPI_Bcast(b, nbytes, MPI_BYTE, MASTERRANK(cr), cr->mpi_comm_mysim);
379 void gmx_sumd(int gmx_unused nr, double gmx_unused r[], const t_commrec gmx_unused *cr)
382 gmx_call("gmx_sumd");
384 #if MPI_IN_PLACE_EXISTS
387 if (cr->nc.rank_intra == 0)
389 /* Use two step summing. */
390 MPI_Reduce(MPI_IN_PLACE, r, nr, MPI_DOUBLE, MPI_SUM, 0,
392 /* Sum the roots of the internal (intra) buffers. */
393 MPI_Allreduce(MPI_IN_PLACE, r, nr, MPI_DOUBLE, MPI_SUM,
398 /* This is here because of the silly MPI specification
399 that MPI_IN_PLACE should be put in sendbuf instead of recvbuf */
400 MPI_Reduce(r, nullptr, nr, MPI_DOUBLE, MPI_SUM, 0, cr->nc.comm_intra);
402 MPI_Bcast(r, nr, MPI_DOUBLE, 0, cr->nc.comm_intra);
406 MPI_Allreduce(MPI_IN_PLACE, r, nr, MPI_DOUBLE, MPI_SUM,
407 cr->mpi_comm_mygroup);
412 if (nr > cr->mpb->dbuf_alloc)
414 cr->mpb->dbuf_alloc = nr;
415 srenew(cr->mpb->dbuf, cr->mpb->dbuf_alloc);
419 /* Use two step summing */
420 MPI_Allreduce(r, cr->mpb->dbuf, nr, MPI_DOUBLE, MPI_SUM, cr->nc.comm_intra);
421 if (cr->nc.rank_intra == 0)
423 /* Sum with the buffers reversed */
424 MPI_Allreduce(cr->mpb->dbuf, r, nr, MPI_DOUBLE, MPI_SUM,
427 MPI_Bcast(r, nr, MPI_DOUBLE, 0, cr->nc.comm_intra);
431 MPI_Allreduce(r, cr->mpb->dbuf, nr, MPI_DOUBLE, MPI_SUM,
432 cr->mpi_comm_mygroup);
433 for (i = 0; i < nr; i++)
435 r[i] = cr->mpb->dbuf[i];
442 void gmx_sumf(int gmx_unused nr, float gmx_unused r[], const t_commrec gmx_unused *cr)
445 gmx_call("gmx_sumf");
447 #if MPI_IN_PLACE_EXISTS
450 /* Use two step summing. */
451 if (cr->nc.rank_intra == 0)
453 MPI_Reduce(MPI_IN_PLACE, r, nr, MPI_FLOAT, MPI_SUM, 0,
455 /* Sum the roots of the internal (intra) buffers */
456 MPI_Allreduce(MPI_IN_PLACE, r, nr, MPI_FLOAT, MPI_SUM,
461 /* This is here because of the silly MPI specification
462 that MPI_IN_PLACE should be put in sendbuf instead of recvbuf */
463 MPI_Reduce(r, nullptr, nr, MPI_FLOAT, MPI_SUM, 0, cr->nc.comm_intra);
465 MPI_Bcast(r, nr, MPI_FLOAT, 0, cr->nc.comm_intra);
469 MPI_Allreduce(MPI_IN_PLACE, r, nr, MPI_FLOAT, MPI_SUM, cr->mpi_comm_mygroup);
474 if (nr > cr->mpb->fbuf_alloc)
476 cr->mpb->fbuf_alloc = nr;
477 srenew(cr->mpb->fbuf, cr->mpb->fbuf_alloc);
481 /* Use two step summing */
482 MPI_Allreduce(r, cr->mpb->fbuf, nr, MPI_FLOAT, MPI_SUM, cr->nc.comm_intra);
483 if (cr->nc.rank_intra == 0)
485 /* Sum with the buffers reversed */
486 MPI_Allreduce(cr->mpb->fbuf, r, nr, MPI_FLOAT, MPI_SUM,
489 MPI_Bcast(r, nr, MPI_FLOAT, 0, cr->nc.comm_intra);
493 MPI_Allreduce(r, cr->mpb->fbuf, nr, MPI_FLOAT, MPI_SUM,
494 cr->mpi_comm_mygroup);
495 for (i = 0; i < nr; i++)
497 r[i] = cr->mpb->fbuf[i];
504 void gmx_sumi(int gmx_unused nr, int gmx_unused r[], const t_commrec gmx_unused *cr)
507 gmx_call("gmx_sumi");
509 #if MPI_IN_PLACE_EXISTS
512 /* Use two step summing */
513 if (cr->nc.rank_intra == 0)
515 MPI_Reduce(MPI_IN_PLACE, r, nr, MPI_INT, MPI_SUM, 0, cr->nc.comm_intra);
516 /* Sum with the buffers reversed */
517 MPI_Allreduce(MPI_IN_PLACE, r, nr, MPI_INT, MPI_SUM, cr->nc.comm_inter);
521 /* This is here because of the silly MPI specification
522 that MPI_IN_PLACE should be put in sendbuf instead of recvbuf */
523 MPI_Reduce(r, nullptr, nr, MPI_INT, MPI_SUM, 0, cr->nc.comm_intra);
525 MPI_Bcast(r, nr, MPI_INT, 0, cr->nc.comm_intra);
529 MPI_Allreduce(MPI_IN_PLACE, r, nr, MPI_INT, MPI_SUM, cr->mpi_comm_mygroup);
534 if (nr > cr->mpb->ibuf_alloc)
536 cr->mpb->ibuf_alloc = nr;
537 srenew(cr->mpb->ibuf, cr->mpb->ibuf_alloc);
541 /* Use two step summing */
542 MPI_Allreduce(r, cr->mpb->ibuf, nr, MPI_INT, MPI_SUM, cr->nc.comm_intra);
543 if (cr->nc.rank_intra == 0)
545 /* Sum with the buffers reversed */
546 MPI_Allreduce(cr->mpb->ibuf, r, nr, MPI_INT, MPI_SUM, cr->nc.comm_inter);
548 MPI_Bcast(r, nr, MPI_INT, 0, cr->nc.comm_intra);
552 MPI_Allreduce(r, cr->mpb->ibuf, nr, MPI_INT, MPI_SUM, cr->mpi_comm_mygroup);
553 for (i = 0; i < nr; i++)
555 r[i] = cr->mpb->ibuf[i];
562 void gmx_sumli(int gmx_unused nr, gmx_int64_t gmx_unused r[], const t_commrec gmx_unused *cr)
565 gmx_call("gmx_sumli");
567 #if MPI_IN_PLACE_EXISTS
570 /* Use two step summing */
571 if (cr->nc.rank_intra == 0)
573 MPI_Reduce(MPI_IN_PLACE, r, nr, MPI_INT64_T, MPI_SUM, 0,
575 /* Sum with the buffers reversed */
576 MPI_Allreduce(MPI_IN_PLACE, r, nr, MPI_INT64_T, MPI_SUM,
581 /* This is here because of the silly MPI specification
582 that MPI_IN_PLACE should be put in sendbuf instead of recvbuf */
583 MPI_Reduce(r, nullptr, nr, MPI_INT64_T, MPI_SUM, 0, cr->nc.comm_intra);
585 MPI_Bcast(r, nr, MPI_INT64_T, 0, cr->nc.comm_intra);
589 MPI_Allreduce(MPI_IN_PLACE, r, nr, MPI_INT64_T, MPI_SUM, cr->mpi_comm_mygroup);
594 if (nr > cr->mpb->libuf_alloc)
596 cr->mpb->libuf_alloc = nr;
597 srenew(cr->mpb->libuf, cr->mpb->libuf_alloc);
601 /* Use two step summing */
602 MPI_Allreduce(r, cr->mpb->libuf, nr, MPI_INT64_T, MPI_SUM,
604 if (cr->nc.rank_intra == 0)
606 /* Sum with the buffers reversed */
607 MPI_Allreduce(cr->mpb->libuf, r, nr, MPI_INT64_T, MPI_SUM,
610 MPI_Bcast(r, nr, MPI_INT64_T, 0, cr->nc.comm_intra);
614 MPI_Allreduce(r, cr->mpb->libuf, nr, MPI_INT64_T, MPI_SUM,
615 cr->mpi_comm_mygroup);
616 for (i = 0; i < nr; i++)
618 r[i] = cr->mpb->libuf[i];
628 static void gmx_sumd_comm(int nr, double r[], MPI_Comm mpi_comm)
630 #if MPI_IN_PLACE_EXISTS
631 MPI_Allreduce(MPI_IN_PLACE, r, nr, MPI_DOUBLE, MPI_SUM, mpi_comm);
633 /* this function is only used in code that is not performance critical,
634 (during setup, when comm_rec is not the appropriate communication
635 structure), so this isn't as bad as it looks. */
640 MPI_Allreduce(r, buf, nr, MPI_DOUBLE, MPI_SUM, mpi_comm);
641 for (i = 0; i < nr; i++)
651 static void gmx_sumf_comm(int nr, float r[], MPI_Comm mpi_comm)
653 #if MPI_IN_PLACE_EXISTS
654 MPI_Allreduce(MPI_IN_PLACE, r, nr, MPI_FLOAT, MPI_SUM, mpi_comm);
656 /* this function is only used in code that is not performance critical,
657 (during setup, when comm_rec is not the appropriate communication
658 structure), so this isn't as bad as it looks. */
663 MPI_Allreduce(r, buf, nr, MPI_FLOAT, MPI_SUM, mpi_comm);
664 for (i = 0; i < nr; i++)
673 void gmx_sumd_sim(int gmx_unused nr, double gmx_unused r[], const gmx_multisim_t gmx_unused *ms)
676 gmx_call("gmx_sumd_sim");
678 gmx_sumd_comm(nr, r, ms->mpi_comm_masters);
682 void gmx_sumf_sim(int gmx_unused nr, float gmx_unused r[], const gmx_multisim_t gmx_unused *ms)
685 gmx_call("gmx_sumf_sim");
687 gmx_sumf_comm(nr, r, ms->mpi_comm_masters);
691 void gmx_sumi_sim(int gmx_unused nr, int gmx_unused r[], const gmx_multisim_t gmx_unused *ms)
694 gmx_call("gmx_sumi_sim");
696 #if MPI_IN_PLACE_EXISTS
697 MPI_Allreduce(MPI_IN_PLACE, r, nr, MPI_INT, MPI_SUM, ms->mpi_comm_masters);
699 /* this is thread-unsafe, but it will do for now: */
702 if (nr > ms->mpb->ibuf_alloc)
704 ms->mpb->ibuf_alloc = nr;
705 srenew(ms->mpb->ibuf, ms->mpb->ibuf_alloc);
707 MPI_Allreduce(r, ms->mpb->ibuf, nr, MPI_INT, MPI_SUM, ms->mpi_comm_masters);
708 for (i = 0; i < nr; i++)
710 r[i] = ms->mpb->ibuf[i];
716 void gmx_sumli_sim(int gmx_unused nr, gmx_int64_t gmx_unused r[], const gmx_multisim_t gmx_unused *ms)
719 gmx_call("gmx_sumli_sim");
721 #if MPI_IN_PLACE_EXISTS
722 MPI_Allreduce(MPI_IN_PLACE, r, nr, MPI_INT64_T, MPI_SUM,
723 ms->mpi_comm_masters);
725 /* this is thread-unsafe, but it will do for now: */
728 if (nr > ms->mpb->libuf_alloc)
730 ms->mpb->libuf_alloc = nr;
731 srenew(ms->mpb->libuf, ms->mpb->libuf_alloc);
733 MPI_Allreduce(r, ms->mpb->libuf, nr, MPI_INT64_T, MPI_SUM,
734 ms->mpi_comm_masters);
735 for (i = 0; i < nr; i++)
737 r[i] = ms->mpb->libuf[i];
743 const char *opt2fn_master(const char *opt, int nfile, const t_filenm fnm[],
746 return SIMMASTER(cr) ? opt2fn(opt, nfile, fnm) : nullptr;
749 void gmx_fatal_collective(int f_errno, const char *file, int line,
750 MPI_Comm comm, gmx_bool bMaster,
751 const char *fmt, ...)
757 /* Check if we are calling on all processes in MPI_COMM_WORLD */
758 MPI_Comm_compare(comm, MPI_COMM_WORLD, &result);
759 /* Any result except MPI_UNEQUAL allows us to call MPI_Finalize */
760 bFinalize = (result != MPI_UNEQUAL);
762 GMX_UNUSED_VALUE(comm);
767 gmx_fatal_mpi_va(f_errno, file, line, bMaster, bFinalize, fmt, ap);