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.
7 * Copyright (c) 2018,2019,2020, by the GROMACS development team, led by
8 * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
9 * and including many others, as listed in the AUTHORS file in the
10 * top-level source directory and at http://www.gromacs.org.
12 * GROMACS is free software; you can redistribute it and/or
13 * modify it under the terms of the GNU Lesser General Public License
14 * as published by the Free Software Foundation; either version 2.1
15 * of the License, or (at your option) any later version.
17 * GROMACS is distributed in the hope that it will be useful,
18 * but WITHOUT ANY WARRANTY; without even the implied warranty of
19 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
20 * Lesser General Public License for more details.
22 * You should have received a copy of the GNU Lesser General Public
23 * License along with GROMACS; if not, see
24 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
25 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
27 * If you want to redistribute modifications to GROMACS, please
28 * consider that scientific software is very special. Version
29 * control is crucial - bugs must be traceable. We will be happy to
30 * consider code for inclusion in the official distribution, but
31 * derived work must not be called official GROMACS. Details are found
32 * in the README & COPYING files - if they are missing, get the
33 * official version at http://www.gromacs.org.
35 * To help us fund GROMACS development, we humbly ask that you cite
36 * the research papers on the package. Check out http://www.gromacs.org.
49 #include "gromacs/commandline/filenm.h"
50 #include "gromacs/mdrunutility/multisim.h"
51 #include "gromacs/mdtypes/commrec.h"
52 #include "gromacs/utility/basenetwork.h"
53 #include "gromacs/utility/cstringutil.h"
54 #include "gromacs/utility/fatalerror.h"
55 #include "gromacs/utility/futil.h"
56 #include "gromacs/utility/gmxmpi.h"
57 #include "gromacs/utility/mpiinplacebuffers.h"
58 #include "gromacs/utility/real.h"
59 #include "gromacs/utility/smalloc.h"
61 /* The source code in this file should be thread-safe.
62 Please keep it that way. */
64 CommrecHandle init_commrec(MPI_Comm communicator, const gmx_multisim_t* ms)
72 int rankInCommunicator, sizeOfCommunicator;
75 GMX_RELEASE_ASSERT(gmx_mpi_initialized(), "Must have initialized MPI before building commrec");
77 MPI_Comm_rank(communicator, &rankInCommunicator);
78 MPI_Comm_size(communicator, &sizeOfCommunicator);
80 GMX_UNUSED_VALUE(communicator);
81 rankInCommunicator = 0;
82 sizeOfCommunicator = 1;
88 cr->nnodes = sizeOfCommunicator / ms->nsim;
89 MPI_Comm_split(communicator, ms->sim, rankInCommunicator, &cr->mpi_comm_mysim);
90 cr->mpi_comm_mygroup = cr->mpi_comm_mysim;
91 MPI_Comm_rank(cr->mpi_comm_mysim, &cr->sim_nodeid);
92 MPI_Comm_rank(cr->mpi_comm_mygroup, &cr->nodeid);
97 cr->nnodes = sizeOfCommunicator;
98 cr->nodeid = rankInCommunicator;
99 cr->sim_nodeid = cr->nodeid;
100 cr->mpi_comm_mysim = communicator;
101 cr->mpi_comm_mygroup = communicator;
104 // TODO cr->duty should not be initialized here
105 cr->duty = (DUTY_PP | DUTY_PME);
107 #if GMX_MPI && !MPI_IN_PLACE_EXISTS
108 /* initialize the MPI_IN_PLACE replacement buffers */
110 cr->mpb->ibuf = nullptr;
111 cr->mpb->libuf = nullptr;
112 cr->mpb->fbuf = nullptr;
113 cr->mpb->dbuf = nullptr;
114 cr->mpb->ibuf_alloc = 0;
115 cr->mpb->libuf_alloc = 0;
116 cr->mpb->fbuf_alloc = 0;
117 cr->mpb->dbuf_alloc = 0;
123 void done_commrec(t_commrec* cr)
127 if (nullptr != cr->dd)
130 // done_domdec(cr->dd);
132 done_mpi_in_place_buf(cr->mpb);
135 // TODO We need to be able to free communicators, but the
136 // structure of the commrec and domdec initialization code makes
137 // it hard to avoid both leaks and double frees.
138 bool mySimIsMyGroup = (cr->mpi_comm_mysim == cr->mpi_comm_mygroup);
139 if (cr->mpi_comm_mysim != MPI_COMM_NULL && cr->mpi_comm_mysim != MPI_COMM_WORLD)
142 // MPI_Comm_free(&cr->mpi_comm_mysim);
144 if (!mySimIsMyGroup && cr->mpi_comm_mygroup != MPI_COMM_NULL && cr->mpi_comm_mygroup != MPI_COMM_WORLD)
147 // MPI_Comm_free(&cr->mpi_comm_mygroup);
153 void gmx_setup_nodecomm(FILE gmx_unused* fplog, t_commrec* cr)
157 /* Many MPI implementations do not optimize MPI_Allreduce
158 * (and probably also other global communication calls)
159 * for multi-core nodes connected by a network.
160 * We can optimize such communication by using one MPI call
161 * within each node and one between the nodes.
162 * For MVAPICH2 and Intel MPI this reduces the time for
163 * the global_stat communication by 25%
164 * for 2x2-core 3 GHz Woodcrest connected by mixed DDR/SDR Infiniband.
165 * B. Hess, November 2007
175 // TODO PhysicalNodeCommunicator could be extended/used to handle
176 // the need for per-node per-group communicators.
177 MPI_Comm_size(cr->mpi_comm_mygroup, &n);
178 MPI_Comm_rank(cr->mpi_comm_mygroup, &rank);
180 int nodehash = gmx_physicalnode_id_hash();
184 fprintf(debug, "In gmx_setup_nodecomm: splitting communicator of size %d\n", n);
188 /* The intra-node communicator, split on node number */
189 MPI_Comm_split(cr->mpi_comm_mygroup, nodehash, rank, &nc->comm_intra);
190 MPI_Comm_rank(nc->comm_intra, &nc->rank_intra);
193 fprintf(debug, "In gmx_setup_nodecomm: node ID %d rank within node %d\n", rank, nc->rank_intra);
195 /* The inter-node communicator, split on rank_intra.
196 * We actually only need the one for rank=0,
197 * but it is easier to create them all.
199 MPI_Comm_split(cr->mpi_comm_mygroup, nc->rank_intra, rank, &nc->comm_inter);
200 /* Check if this really created two step communication */
203 MPI_Comm_size(nc->comm_inter, &ng);
204 MPI_Comm_size(nc->comm_intra, &ni);
207 fprintf(debug, "In gmx_setup_nodecomm: groups %d, my group size %d\n", ng, ni);
210 if (getenv("GMX_NO_NODECOMM") == nullptr && ((ng > 1 && ng < n) || (ni > 1 && ni < n)))
215 fprintf(fplog, "Using two step summing over %d groups of on average %.1f ranks\n\n", ng,
218 if (nc->rank_intra > 0)
220 MPI_Comm_free(&nc->comm_inter);
225 /* One group or all processes in a separate group, use normal summing */
226 MPI_Comm_free(&nc->comm_inter);
227 MPI_Comm_free(&nc->comm_intra);
231 "In gmx_setup_nodecomm: not unsing separate inter- and intra-node "
237 /* tMPI runs only on a single node so just use the nodeid */
238 nc->rank_intra = cr->nodeid;
242 void gmx_barrier(MPI_Comm gmx_unused communicator)
245 gmx_call("gmx_barrier");
247 MPI_Barrier(communicator);
251 void gmx_bcast(int gmx_unused nbytes, void gmx_unused* b, MPI_Comm gmx_unused communicator)
254 gmx_call("gmx_bcast");
256 MPI_Bcast(b, nbytes, MPI_BYTE, 0, communicator);
260 void gmx_sumd(int gmx_unused nr, double gmx_unused r[], const t_commrec gmx_unused* cr)
263 gmx_call("gmx_sumd");
265 # if MPI_IN_PLACE_EXISTS
268 if (cr->nc.rank_intra == 0)
270 /* Use two step summing. */
271 MPI_Reduce(MPI_IN_PLACE, r, nr, MPI_DOUBLE, MPI_SUM, 0, cr->nc.comm_intra);
272 /* Sum the roots of the internal (intra) buffers. */
273 MPI_Allreduce(MPI_IN_PLACE, r, nr, MPI_DOUBLE, MPI_SUM, cr->nc.comm_inter);
277 /* This is here because of the silly MPI specification
278 that MPI_IN_PLACE should be put in sendbuf instead of recvbuf */
279 MPI_Reduce(r, nullptr, nr, MPI_DOUBLE, MPI_SUM, 0, cr->nc.comm_intra);
281 MPI_Bcast(r, nr, MPI_DOUBLE, 0, cr->nc.comm_intra);
285 MPI_Allreduce(MPI_IN_PLACE, r, nr, MPI_DOUBLE, MPI_SUM, cr->mpi_comm_mygroup);
290 if (nr > cr->mpb->dbuf_alloc)
292 cr->mpb->dbuf_alloc = nr;
293 srenew(cr->mpb->dbuf, cr->mpb->dbuf_alloc);
297 /* Use two step summing */
298 MPI_Allreduce(r, cr->mpb->dbuf, nr, MPI_DOUBLE, MPI_SUM, cr->nc.comm_intra);
299 if (cr->nc.rank_intra == 0)
301 /* Sum with the buffers reversed */
302 MPI_Allreduce(cr->mpb->dbuf, r, nr, MPI_DOUBLE, MPI_SUM, cr->nc.comm_inter);
304 MPI_Bcast(r, nr, MPI_DOUBLE, 0, cr->nc.comm_intra);
308 MPI_Allreduce(r, cr->mpb->dbuf, nr, MPI_DOUBLE, MPI_SUM, cr->mpi_comm_mygroup);
309 for (i = 0; i < nr; i++)
311 r[i] = cr->mpb->dbuf[i];
318 void gmx_sumf(int gmx_unused nr, float gmx_unused r[], const t_commrec gmx_unused* cr)
321 gmx_call("gmx_sumf");
323 # if MPI_IN_PLACE_EXISTS
326 /* Use two step summing. */
327 if (cr->nc.rank_intra == 0)
329 MPI_Reduce(MPI_IN_PLACE, r, nr, MPI_FLOAT, MPI_SUM, 0, cr->nc.comm_intra);
330 /* Sum the roots of the internal (intra) buffers */
331 MPI_Allreduce(MPI_IN_PLACE, r, nr, MPI_FLOAT, MPI_SUM, cr->nc.comm_inter);
335 /* This is here because of the silly MPI specification
336 that MPI_IN_PLACE should be put in sendbuf instead of recvbuf */
337 MPI_Reduce(r, nullptr, nr, MPI_FLOAT, MPI_SUM, 0, cr->nc.comm_intra);
339 MPI_Bcast(r, nr, MPI_FLOAT, 0, cr->nc.comm_intra);
343 MPI_Allreduce(MPI_IN_PLACE, r, nr, MPI_FLOAT, MPI_SUM, cr->mpi_comm_mygroup);
348 if (nr > cr->mpb->fbuf_alloc)
350 cr->mpb->fbuf_alloc = nr;
351 srenew(cr->mpb->fbuf, cr->mpb->fbuf_alloc);
355 /* Use two step summing */
356 MPI_Allreduce(r, cr->mpb->fbuf, nr, MPI_FLOAT, MPI_SUM, cr->nc.comm_intra);
357 if (cr->nc.rank_intra == 0)
359 /* Sum with the buffers reversed */
360 MPI_Allreduce(cr->mpb->fbuf, r, nr, MPI_FLOAT, MPI_SUM, cr->nc.comm_inter);
362 MPI_Bcast(r, nr, MPI_FLOAT, 0, cr->nc.comm_intra);
366 MPI_Allreduce(r, cr->mpb->fbuf, nr, MPI_FLOAT, MPI_SUM, cr->mpi_comm_mygroup);
367 for (i = 0; i < nr; i++)
369 r[i] = cr->mpb->fbuf[i];
376 void gmx_sumi(int gmx_unused nr, int gmx_unused r[], const t_commrec gmx_unused* cr)
379 gmx_call("gmx_sumi");
381 # if MPI_IN_PLACE_EXISTS
384 /* Use two step summing */
385 if (cr->nc.rank_intra == 0)
387 MPI_Reduce(MPI_IN_PLACE, r, nr, MPI_INT, MPI_SUM, 0, cr->nc.comm_intra);
388 /* Sum with the buffers reversed */
389 MPI_Allreduce(MPI_IN_PLACE, r, nr, MPI_INT, MPI_SUM, cr->nc.comm_inter);
393 /* This is here because of the silly MPI specification
394 that MPI_IN_PLACE should be put in sendbuf instead of recvbuf */
395 MPI_Reduce(r, nullptr, nr, MPI_INT, MPI_SUM, 0, cr->nc.comm_intra);
397 MPI_Bcast(r, nr, MPI_INT, 0, cr->nc.comm_intra);
401 MPI_Allreduce(MPI_IN_PLACE, r, nr, MPI_INT, MPI_SUM, cr->mpi_comm_mygroup);
406 if (nr > cr->mpb->ibuf_alloc)
408 cr->mpb->ibuf_alloc = nr;
409 srenew(cr->mpb->ibuf, cr->mpb->ibuf_alloc);
413 /* Use two step summing */
414 MPI_Allreduce(r, cr->mpb->ibuf, nr, MPI_INT, MPI_SUM, cr->nc.comm_intra);
415 if (cr->nc.rank_intra == 0)
417 /* Sum with the buffers reversed */
418 MPI_Allreduce(cr->mpb->ibuf, r, nr, MPI_INT, MPI_SUM, cr->nc.comm_inter);
420 MPI_Bcast(r, nr, MPI_INT, 0, cr->nc.comm_intra);
424 MPI_Allreduce(r, cr->mpb->ibuf, nr, MPI_INT, MPI_SUM, cr->mpi_comm_mygroup);
425 for (i = 0; i < nr; i++)
427 r[i] = cr->mpb->ibuf[i];
434 void gmx_sumli(int gmx_unused nr, int64_t gmx_unused r[], const t_commrec gmx_unused* cr)
437 gmx_call("gmx_sumli");
439 # if MPI_IN_PLACE_EXISTS
442 /* Use two step summing */
443 if (cr->nc.rank_intra == 0)
445 MPI_Reduce(MPI_IN_PLACE, r, nr, MPI_INT64_T, MPI_SUM, 0, cr->nc.comm_intra);
446 /* Sum with the buffers reversed */
447 MPI_Allreduce(MPI_IN_PLACE, r, nr, MPI_INT64_T, MPI_SUM, cr->nc.comm_inter);
451 /* This is here because of the silly MPI specification
452 that MPI_IN_PLACE should be put in sendbuf instead of recvbuf */
453 MPI_Reduce(r, nullptr, nr, MPI_INT64_T, MPI_SUM, 0, cr->nc.comm_intra);
455 MPI_Bcast(r, nr, MPI_INT64_T, 0, cr->nc.comm_intra);
459 MPI_Allreduce(MPI_IN_PLACE, r, nr, MPI_INT64_T, MPI_SUM, cr->mpi_comm_mygroup);
464 if (nr > cr->mpb->libuf_alloc)
466 cr->mpb->libuf_alloc = nr;
467 srenew(cr->mpb->libuf, cr->mpb->libuf_alloc);
471 /* Use two step summing */
472 MPI_Allreduce(r, cr->mpb->libuf, nr, MPI_INT64_T, MPI_SUM, cr->nc.comm_intra);
473 if (cr->nc.rank_intra == 0)
475 /* Sum with the buffers reversed */
476 MPI_Allreduce(cr->mpb->libuf, r, nr, MPI_INT64_T, MPI_SUM, cr->nc.comm_inter);
478 MPI_Bcast(r, nr, MPI_INT64_T, 0, cr->nc.comm_intra);
482 MPI_Allreduce(r, cr->mpb->libuf, nr, MPI_INT64_T, MPI_SUM, cr->mpi_comm_mygroup);
483 for (i = 0; i < nr; i++)
485 r[i] = cr->mpb->libuf[i];
492 const char* opt2fn_master(const char* opt, int nfile, const t_filenm fnm[], t_commrec* cr)
494 return SIMMASTER(cr) ? opt2fn(opt, nfile, fnm) : nullptr;
497 void gmx_fatal_collective(int f_errno,
502 gmx_fmtstr const char* fmt,
509 /* Check if we are calling on all processes in MPI_COMM_WORLD */
510 MPI_Comm_compare(comm, MPI_COMM_WORLD, &result);
511 /* Any result except MPI_UNEQUAL allows us to call MPI_Finalize */
512 bFinalize = (result != MPI_UNEQUAL);
514 GMX_UNUSED_VALUE(comm);
519 gmx_fatal_mpi_va(f_errno, file, line, bMaster, bFinalize, fmt, ap);