3 * This source code is part of
7 * GROningen MAchine for Chemical Simulations
10 * Written by David van der Spoel, Erik Lindahl, Berk Hess, and others.
11 * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
12 * Copyright (c) 2001-2004, The GROMACS development team,
13 * check out http://www.gromacs.org for more information.
15 * This program is free software; you can redistribute it and/or
16 * modify it under the terms of the GNU General Public License
17 * as published by the Free Software Foundation; either version 2
18 * of the License, or (at your option) any later version.
20 * If you want to redistribute modifications, please consider that
21 * scientific software is very special. Version control is crucial -
22 * bugs must be traceable. We will be happy to consider code for
23 * inclusion in the official distribution, but derived work must not
24 * be called official GROMACS. Details are found in the README & COPYING
25 * files - if they are missing, get the official version at www.gromacs.org.
27 * To help us fund GROMACS development, we humbly ask that you cite
28 * the papers on the package - you can find them in the top README file.
30 * For more info, check our website at http://www.gromacs.org
33 * GROningen Mixture of Alchemy and Childrens' Stories
40 #include "gmx_fatal.h"
57 #include "mpelogging.h"
59 /* The source code in this file should be thread-safe.
60 Please keep it that way. */
62 bool gmx_mpi_initialized(void)
74 int gmx_setup(int *argc,char **argv,int *nnodes)
77 gmx_call("gmx_setup");
81 int resultlen; /* actual length of node name */
85 char mpi_hostname[MPI_MAX_PROCESSOR_NAME];
87 /* Call the MPI routines */
90 (void) fah_MPI_Init(argc,&argv);
92 (void) MPI_Init(argc,&argv);
95 (void) MPI_Comm_size( MPI_COMM_WORLD, &mpi_num_nodes );
96 (void) MPI_Comm_rank( MPI_COMM_WORLD, &mpi_my_rank );
97 (void) MPI_Get_processor_name( mpi_hostname, &resultlen );
101 /* MPE logging routines. Get event IDs from MPE: */
103 ev_timestep1 = MPE_Log_get_event_number( );
104 ev_timestep2 = MPE_Log_get_event_number( );
105 ev_force_start = MPE_Log_get_event_number( );
106 ev_force_finish = MPE_Log_get_event_number( );
107 ev_do_fnbf_start = MPE_Log_get_event_number( );
108 ev_do_fnbf_finish = MPE_Log_get_event_number( );
109 ev_ns_start = MPE_Log_get_event_number( );
110 ev_ns_finish = MPE_Log_get_event_number( );
111 ev_calc_bonds_start = MPE_Log_get_event_number( );
112 ev_calc_bonds_finish = MPE_Log_get_event_number( );
113 ev_global_stat_start = MPE_Log_get_event_number( );
114 ev_global_stat_finish = MPE_Log_get_event_number( );
115 ev_virial_start = MPE_Log_get_event_number( );
116 ev_virial_finish = MPE_Log_get_event_number( );
118 /* Shift related events */
119 ev_shift_start = MPE_Log_get_event_number( );
120 ev_shift_finish = MPE_Log_get_event_number( );
121 ev_unshift_start = MPE_Log_get_event_number( );
122 ev_unshift_finish = MPE_Log_get_event_number( );
123 ev_mk_mshift_start = MPE_Log_get_event_number( );
124 ev_mk_mshift_finish = MPE_Log_get_event_number( );
126 /* PME related events */
127 ev_pme_start = MPE_Log_get_event_number( );
128 ev_pme_finish = MPE_Log_get_event_number( );
129 ev_spread_on_grid_start = MPE_Log_get_event_number( );
130 ev_spread_on_grid_finish = MPE_Log_get_event_number( );
131 ev_sum_qgrid_start = MPE_Log_get_event_number( );
132 ev_sum_qgrid_finish = MPE_Log_get_event_number( );
133 ev_gmxfft3d_start = MPE_Log_get_event_number( );
134 ev_gmxfft3d_finish = MPE_Log_get_event_number( );
135 ev_solve_pme_start = MPE_Log_get_event_number( );
136 ev_solve_pme_finish = MPE_Log_get_event_number( );
137 ev_gather_f_bsplines_start = MPE_Log_get_event_number( );
138 ev_gather_f_bsplines_finish= MPE_Log_get_event_number( );
139 ev_reduce_start = MPE_Log_get_event_number( );
140 ev_reduce_finish = MPE_Log_get_event_number( );
141 ev_rscatter_start = MPE_Log_get_event_number( );
142 ev_rscatter_finish = MPE_Log_get_event_number( );
143 ev_alltoall_start = MPE_Log_get_event_number( );
144 ev_alltoall_finish = MPE_Log_get_event_number( );
145 ev_pmeredist_start = MPE_Log_get_event_number( );
146 ev_pmeredist_finish = MPE_Log_get_event_number( );
147 ev_init_pme_start = MPE_Log_get_event_number( );
148 ev_init_pme_finish = MPE_Log_get_event_number( );
149 ev_send_coordinates_start = MPE_Log_get_event_number( );
150 ev_send_coordinates_finish = MPE_Log_get_event_number( );
151 ev_update_fr_start = MPE_Log_get_event_number( );
152 ev_update_fr_finish = MPE_Log_get_event_number( );
153 ev_clear_rvecs_start = MPE_Log_get_event_number( );
154 ev_clear_rvecs_finish = MPE_Log_get_event_number( );
155 ev_update_start = MPE_Log_get_event_number( );
156 ev_update_finish = MPE_Log_get_event_number( );
157 ev_output_start = MPE_Log_get_event_number( );
158 ev_output_finish = MPE_Log_get_event_number( );
159 ev_sum_lrforces_start = MPE_Log_get_event_number( );
160 ev_sum_lrforces_finish = MPE_Log_get_event_number( );
161 ev_sort_start = MPE_Log_get_event_number( );
162 ev_sort_finish = MPE_Log_get_event_number( );
163 ev_sum_qgrid_start = MPE_Log_get_event_number( );
164 ev_sum_qgrid_finish = MPE_Log_get_event_number( );
166 /* Essential dynamics related events */
167 ev_edsam_start = MPE_Log_get_event_number( );
168 ev_edsam_finish = MPE_Log_get_event_number( );
169 ev_get_coords_start = MPE_Log_get_event_number( );
170 ev_get_coords_finish = MPE_Log_get_event_number( );
171 ev_ed_apply_cons_start = MPE_Log_get_event_number( );
172 ev_ed_apply_cons_finish = MPE_Log_get_event_number( );
173 ev_fit_to_reference_start = MPE_Log_get_event_number( );
174 ev_fit_to_reference_finish = MPE_Log_get_event_number( );
176 /* describe events: */
177 if ( mpi_my_rank == 0 )
180 MPE_Describe_state(ev_timestep1, ev_timestep2, "timestep START", "magenta" );
181 MPE_Describe_state(ev_force_start, ev_force_finish, "force", "cornflower blue" );
182 MPE_Describe_state(ev_do_fnbf_start, ev_do_fnbf_finish, "do_fnbf", "navy" );
183 MPE_Describe_state(ev_ns_start, ev_ns_finish, "neighbor search", "tomato" );
184 MPE_Describe_state(ev_calc_bonds_start, ev_calc_bonds_finish, "bonded forces", "slate blue" );
185 MPE_Describe_state(ev_global_stat_start, ev_global_stat_finish, "global stat", "firebrick3");
186 MPE_Describe_state(ev_update_fr_start, ev_update_fr_finish, "update forcerec", "goldenrod");
187 MPE_Describe_state(ev_clear_rvecs_start, ev_clear_rvecs_finish, "clear rvecs", "bisque");
188 MPE_Describe_state(ev_update_start, ev_update_finish, "update", "cornsilk");
189 MPE_Describe_state(ev_output_start, ev_output_finish, "output", "black");
190 MPE_Describe_state(ev_virial_start, ev_virial_finish, "calc_virial", "thistle4");
192 /* PME related events */
193 MPE_Describe_state(ev_pme_start, ev_pme_finish, "doing PME", "grey" );
194 MPE_Describe_state(ev_spread_on_grid_start, ev_spread_on_grid_finish, "spread", "dark orange" );
195 MPE_Describe_state(ev_sum_qgrid_start, ev_sum_qgrid_finish, "sum qgrid", "slate blue");
196 MPE_Describe_state(ev_gmxfft3d_start, ev_gmxfft3d_finish, "fft3d", "snow2" );
197 MPE_Describe_state(ev_solve_pme_start, ev_solve_pme_finish, "solve PME", "indian red" );
198 MPE_Describe_state(ev_gather_f_bsplines_start, ev_gather_f_bsplines_finish, "bsplines", "light sea green" );
199 MPE_Describe_state(ev_reduce_start, ev_reduce_finish, "reduce", "cyan1" );
200 MPE_Describe_state(ev_rscatter_start, ev_rscatter_finish, "rscatter", "cyan3" );
201 MPE_Describe_state(ev_alltoall_start, ev_alltoall_finish, "alltoall", "LightCyan4" );
202 MPE_Describe_state(ev_pmeredist_start, ev_pmeredist_finish, "pmeredist", "thistle" );
203 MPE_Describe_state(ev_init_pme_start, ev_init_pme_finish, "init PME", "snow4");
204 MPE_Describe_state(ev_send_coordinates_start, ev_send_coordinates_finish, "send_coordinates","blue");
205 MPE_Describe_state(ev_sum_lrforces_start, ev_sum_lrforces_finish, "sum_LRforces", "lime green");
206 MPE_Describe_state(ev_sort_start, ev_sort_finish, "sort pme atoms", "brown");
207 MPE_Describe_state(ev_sum_qgrid_start, ev_sum_qgrid_finish, "sum charge grid", "medium orchid");
209 /* Shift related events */
210 MPE_Describe_state(ev_shift_start, ev_shift_finish, "shift", "orange");
211 MPE_Describe_state(ev_unshift_start, ev_unshift_finish, "unshift", "dark orange");
212 MPE_Describe_state(ev_mk_mshift_start, ev_mk_mshift_finish, "mk_mshift", "maroon");
214 /* Essential dynamics related events */
215 MPE_Describe_state(ev_edsam_start, ev_edsam_finish, "EDSAM", "deep sky blue");
216 MPE_Describe_state(ev_get_coords_start, ev_get_coords_finish, "ED get coords", "steel blue");
217 MPE_Describe_state(ev_ed_apply_cons_start, ev_ed_apply_cons_finish, "ED apply constr", "forest green");
218 MPE_Describe_state(ev_fit_to_reference_start, ev_fit_to_reference_finish, "ED fit to ref", "lavender");
225 fprintf(stderr,"NNODES=%d, MYRANK=%d, HOSTNAME=%s\n",
226 mpi_num_nodes,mpi_my_rank,mpi_hostname);
229 *nnodes=mpi_num_nodes;
235 int gmx_node_num(void)
241 (void) MPI_Comm_size(MPI_COMM_WORLD, &i);
246 int gmx_node_rank(void)
252 (void) MPI_Comm_rank(MPI_COMM_WORLD, &i);
257 void gmx_setup_nodecomm(FILE *fplog,t_commrec *cr)
260 int n,rank,resultlen,hostnum,i,j,ng,ni;
262 char mpi_hostname[MPI_MAX_PROCESSOR_NAME],num[MPI_MAX_PROCESSOR_NAME];
265 /* Many MPI implementations do not optimize MPI_Allreduce
266 * (and probably also other global communication calls)
267 * for multi-core nodes connected by a network.
268 * We can optimize such communication by using one MPI call
269 * within each node and one between the nodes.
270 * For MVAPICH2 and Intel MPI this reduces the time for
271 * the global_stat communication by 25%
272 * for 2x2-core 3 GHz Woodcrest connected by mixed DDR/SDR Infiniband.
273 * B. Hess, November 2007
280 if (getenv("GMX_NO_NODECOMM") == NULL) {
282 MPI_Comm_size(cr->mpi_comm_mygroup,&n);
283 MPI_Comm_rank(cr->mpi_comm_mygroup,&rank);
284 MPI_Get_processor_name(mpi_hostname,&resultlen);
285 /* This procedure can only differentiate nodes with host names
286 * that end on unique numbers.
290 /* Only parse the host name up to the first dot */
291 while(i < resultlen && mpi_hostname[i] != '.') {
292 if (isdigit(mpi_hostname[i])) {
293 num[j++] = mpi_hostname[i];
301 /* Use only the last 9 decimals, so we don't overflow an int */
302 hostnum = strtol(num + max(0,j-9), NULL, 10);
307 "In gmx_setup_nodecomm: splitting communicator of size %d\n",
309 fprintf(debug,"In gmx_setup_nodecomm: hostname '%s', hostnum %d\n",
310 mpi_hostname,hostnum);
313 /* The intra-node communicator, split on node number */
314 MPI_Comm_split(cr->mpi_comm_mygroup,hostnum,rank,&nc->comm_intra);
315 MPI_Comm_rank(nc->comm_intra,&nc->rank_intra);
317 fprintf(debug,"In gmx_setup_nodecomm: node rank %d rank_intra %d\n",
318 rank,nc->rank_intra);
320 /* The inter-node communicator, split on rank_intra.
321 * We actually only need the one for rank=0,
322 * but it is easier to create them all.
324 MPI_Comm_split(cr->mpi_comm_mygroup,nc->rank_intra,rank,&nc->comm_inter);
325 /* Check if this really created two step communication */
326 MPI_Comm_size(nc->comm_inter,&ng);
327 MPI_Comm_size(nc->comm_intra,&ni);
329 fprintf(debug,"In gmx_setup_nodecomm: groups %d, my group size %d\n",
332 if ((ng > 1 && ng < n) || (ni > 1 && ni < n)) {
335 fprintf(fplog,"Using two step summing over %d groups of on average %.1f processes\n\n",ng,(real)n/(real)ng);
336 if (nc->rank_intra > 0)
337 MPI_Comm_free(&nc->comm_inter);
339 /* One group or all processes in a separate group, use normal summing */
340 MPI_Comm_free(&nc->comm_inter);
341 MPI_Comm_free(&nc->comm_intra);
348 void gmx_barrier(const t_commrec *cr)
351 gmx_call("gmx_barrier");
353 MPI_Barrier(cr->mpi_comm_mygroup);
357 void gmx_abort(int noderank,int nnodes,int errorno)
360 gmx_call("gmx_abort");
363 fprintf(stderr,"Halting program %s\n",ShortProgram());
369 fprintf(stderr,"Halting parallel program %s on CPU %d out of %d\n",
370 ShortProgram(),noderank,nnodes);
374 fprintf(stderr,"Halting program %s\n",ShortProgram());
378 MPI_Abort(MPI_COMM_WORLD,errorno);
384 void gmx_bcast(int nbytes,void *b,const t_commrec *cr)
387 gmx_call("gmx_bast");
389 MPI_Bcast(b,nbytes,MPI_BYTE,MASTERRANK(cr),cr->mpi_comm_mygroup);
393 void gmx_bcast_sim(int nbytes,void *b,const t_commrec *cr)
396 gmx_call("gmx_bast");
398 MPI_Bcast(b,nbytes,MPI_BYTE,MASTERRANK(cr),cr->mpi_comm_mysim);
402 void gmx_sumd(int nr,double r[],const t_commrec *cr)
405 gmx_call("gmx_sumd");
407 #if defined(MPI_IN_PLACE_EXISTS) || defined(GMX_THREADS)
409 if (cr->nc.rank_intra == 0)
411 /* Use two step summing. */
412 MPI_Reduce(MPI_IN_PLACE,r,nr,MPI_DOUBLE,MPI_SUM,0,
414 /* Sum the roots of the internal (intra) buffers. */
415 MPI_Allreduce(MPI_IN_PLACE,r,nr,MPI_DOUBLE,MPI_SUM,
420 /* This is here because of the silly MPI specification
421 that MPI_IN_PLACE should be put in sendbuf instead of recvbuf */
422 MPI_Reduce(r,NULL,nr,MPI_DOUBLE,MPI_SUM,0,cr->nc.comm_intra);
424 MPI_Bcast(r,nr,MPI_DOUBLE,0,cr->nc.comm_intra);
428 MPI_Allreduce(MPI_IN_PLACE,r,nr,MPI_DOUBLE,MPI_SUM,
429 cr->mpi_comm_mygroup);
434 if (nr > cr->mpb->dbuf_alloc) {
435 cr->mpb->dbuf_alloc = nr;
436 srenew(cr->mpb->dbuf,cr->mpb->dbuf_alloc);
439 /* Use two step summing */
440 MPI_Allreduce(r,cr->mpb->dbuf,nr,MPI_DOUBLE,MPI_SUM,cr->nc.comm_intra);
441 if (cr->nc.rank_intra == 0) {
442 /* Sum with the buffers reversed */
443 MPI_Allreduce(cr->mpb->dbuf,r,nr,MPI_DOUBLE,MPI_SUM,
446 MPI_Bcast(r,nr,MPI_DOUBLE,0,cr->nc.comm_intra);
448 MPI_Allreduce(r,cr->mpb->dbuf,nr,MPI_DOUBLE,MPI_SUM,
449 cr->mpi_comm_mygroup);
451 r[i] = cr->mpb->dbuf[i];
457 void gmx_sumf(int nr,float r[],const t_commrec *cr)
460 gmx_call("gmx_sumf");
462 #if defined(MPI_IN_PLACE_EXISTS) || defined(GMX_THREADS)
464 /* Use two step summing. */
465 if (cr->nc.rank_intra == 0)
467 MPI_Reduce(MPI_IN_PLACE,r,nr,MPI_FLOAT,MPI_SUM,0,
469 /* Sum the roots of the internal (intra) buffers */
470 MPI_Allreduce(MPI_IN_PLACE,r,nr,MPI_FLOAT,MPI_SUM,
475 /* This is here because of the silly MPI specification
476 that MPI_IN_PLACE should be put in sendbuf instead of recvbuf */
477 MPI_Reduce(r,NULL,nr,MPI_FLOAT,MPI_SUM,0,cr->nc.comm_intra);
479 MPI_Bcast(r,nr,MPI_FLOAT,0,cr->nc.comm_intra);
483 MPI_Allreduce(MPI_IN_PLACE,r,nr,MPI_FLOAT,MPI_SUM,cr->mpi_comm_mygroup);
488 if (nr > cr->mpb->fbuf_alloc) {
489 cr->mpb->fbuf_alloc = nr;
490 srenew(cr->mpb->fbuf,cr->mpb->fbuf_alloc);
493 /* Use two step summing */
494 MPI_Allreduce(r,cr->mpb->fbuf,nr,MPI_FLOAT,MPI_SUM,cr->nc.comm_intra);
495 if (cr->nc.rank_intra == 0) {
496 /* Sum with the buffers reversed */
497 MPI_Allreduce(cr->mpb->fbuf,r,nr,MPI_FLOAT,MPI_SUM,
500 MPI_Bcast(r,nr,MPI_FLOAT,0,cr->nc.comm_intra);
502 MPI_Allreduce(r,cr->mpb->fbuf,nr,MPI_FLOAT,MPI_SUM,
503 cr->mpi_comm_mygroup);
505 r[i] = cr->mpb->fbuf[i];
511 void gmx_sumi(int nr,int r[],const t_commrec *cr)
514 gmx_call("gmx_sumi");
516 #if defined(MPI_IN_PLACE_EXISTS) || defined(GMX_THREADS)
518 /* Use two step summing */
519 if (cr->nc.rank_intra == 0)
521 MPI_Reduce(MPI_IN_PLACE,r,nr,MPI_INT,MPI_SUM,0,cr->nc.comm_intra);
522 /* Sum with the buffers reversed */
523 MPI_Allreduce(MPI_IN_PLACE,r,nr,MPI_INT,MPI_SUM,cr->nc.comm_inter);
527 /* This is here because of the silly MPI specification
528 that MPI_IN_PLACE should be put in sendbuf instead of recvbuf */
529 MPI_Reduce(r,NULL,nr,MPI_INT,MPI_SUM,0,cr->nc.comm_intra);
531 MPI_Bcast(r,nr,MPI_INT,0,cr->nc.comm_intra);
535 MPI_Allreduce(MPI_IN_PLACE,r,nr,MPI_INT,MPI_SUM,cr->mpi_comm_mygroup);
540 if (nr > cr->mpb->ibuf_alloc) {
541 cr->mpb->ibuf_alloc = nr;
542 srenew(cr->mpb->ibuf,cr->mpb->ibuf_alloc);
545 /* Use two step summing */
546 MPI_Allreduce(r,cr->mpb->ibuf,nr,MPI_INT,MPI_SUM,cr->nc.comm_intra);
547 if (cr->nc.rank_intra == 0) {
548 /* Sum with the buffers reversed */
549 MPI_Allreduce(cr->mpb->ibuf,r,nr,MPI_INT,MPI_SUM,cr->nc.comm_inter);
551 MPI_Bcast(r,nr,MPI_INT,0,cr->nc.comm_intra);
553 MPI_Allreduce(r,cr->mpb->ibuf,nr,MPI_INT,MPI_SUM,cr->mpi_comm_mygroup);
555 r[i] = cr->mpb->ibuf[i];
562 void gmx_sumd_comm(int nr,double r[],MPI_Comm mpi_comm)
564 #if defined(MPI_IN_PLACE_EXISTS) || defined(GMX_THREADS)
565 MPI_Allreduce(MPI_IN_PLACE,r,nr,MPI_DOUBLE,MPI_SUM,mpi_comm);
567 /* this function is only used in code that is not performance critical,
568 (during setup, when comm_rec is not the appropriate communication
569 structure), so this isn't as bad as it looks. */
574 MPI_Allreduce(r,buf,nr,MPI_DOUBLE,MPI_SUM,mpi_comm);
583 void gmx_sumf_comm(int nr,float r[],MPI_Comm mpi_comm)
585 #if defined(MPI_IN_PLACE_EXISTS) || defined(GMX_THREADS)
586 MPI_Allreduce(MPI_IN_PLACE,r,nr,MPI_FLOAT,MPI_SUM,mpi_comm);
588 /* this function is only used in code that is not performance critical,
589 (during setup, when comm_rec is not the appropriate communication
590 structure), so this isn't as bad as it looks. */
595 MPI_Allreduce(r,buf,nr,MPI_FLOAT,MPI_SUM,mpi_comm);
603 void gmx_sumd_sim(int nr,double r[],const gmx_multisim_t *ms)
606 gmx_call("gmx_sumd");
608 gmx_sumd_comm(nr,r,ms->mpi_comm_masters);
612 void gmx_sumf_sim(int nr,float r[],const gmx_multisim_t *ms)
615 gmx_call("gmx_sumf");
617 gmx_sumf_comm(nr,r,ms->mpi_comm_masters);
621 void gmx_sumi_sim(int nr,int r[], const gmx_multisim_t *ms)
624 gmx_call("gmx_sumd");
626 #if defined(MPI_IN_PLACE_EXISTS) || defined(GMX_THREADS)
627 MPI_Allreduce(MPI_IN_PLACE,r,nr,MPI_INT,MPI_SUM,ms->mpi_comm_masters);
629 /* this is thread-unsafe, but it will do for now: */
632 if (nr > ms->mpb->ibuf_alloc) {
633 ms->mpb->ibuf_alloc = nr;
634 srenew(ms->mpb->ibuf,ms->mpb->ibuf_alloc);
636 MPI_Allreduce(r,ms->mpb->ibuf,nr,MPI_INT,MPI_SUM,ms->mpi_comm_masters);
638 r[i] = ms->mpb->ibuf[i];
643 void gmx_finalize(void)
646 gmx_call("gmx_finalize");
650 /* just as a check; we don't want to finalize twice */
652 MPI_Finalized(&finalized);
656 /* We sync the processes here to try to avoid problems
657 * with buggy MPI implementations that could cause
658 * unfinished processes to terminate.
660 MPI_Barrier(MPI_COMM_WORLD);
663 if (DOMAINDECOMP(cr)) {
664 if (cr->npmenodes > 0 || cr->dd->bCartesian)
665 MPI_Comm_free(&cr->mpi_comm_mygroup);
666 if (cr->dd->bCartesian)
667 MPI_Comm_free(&cr->mpi_comm_mysim);
671 /* Apparently certain mpich implementations cause problems
672 * with MPI_Finalize. In that case comment out MPI_Finalize.
675 fprintf(debug,"Will call MPI_Finalize now\n");
677 ret = MPI_Finalize();
679 fprintf(debug,"Return code from MPI_Finalize = %d\n",ret);