1 /* -*- mode: c; tab-width: 4; indent-tabs-mode: nil; c-basic-offset: 4; c-file-style: "stroustrup"; -*-
4 * This source code is part of
8 * GROningen MAchine for Chemical Simulations
11 * Written by David van der Spoel, Erik Lindahl, Berk Hess, and others.
12 * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
13 * Copyright (c) 2001-2004, The GROMACS development team,
14 * check out http://www.gromacs.org for more information.
16 * This program is free software; you can redistribute it and/or
17 * modify it under the terms of the GNU General Public License
18 * as published by the Free Software Foundation; either version 2
19 * of the License, or (at your option) any later version.
21 * If you want to redistribute modifications, please consider that
22 * scientific software is very special. Version control is crucial -
23 * bugs must be traceable. We will be happy to consider code for
24 * inclusion in the official distribution, but derived work must not
25 * be called official GROMACS. Details are found in the README & COPYING
26 * files - if they are missing, get the official version at www.gromacs.org.
28 * To help us fund GROMACS development, we humbly ask that you cite
29 * the papers on the package - you can find them in the top README file.
31 * For more info, check our website at http://www.gromacs.org
34 * GROningen Mixture of Alchemy and Childrens' Stories
39 #include "gromacs/utility/gmx_header_config.h"
47 #ifdef HAVE_SYS_TIME_H
52 #include "gmx_fatal.h"
63 #include "thread_mpi.h"
66 /* The source code in this file should be thread-safe.
67 Please keep it that way. */
74 #ifdef GMX_NATIVE_WINDOWS
79 /* Portable version of ctime_r implemented in src/gmxlib/string2.c, but we do not want it declared in public installed headers */
81 gmx_ctime_r(const time_t *clock,char *buf, int n);
87 static void par_fn(char *base,int ftp,const t_commrec *cr,
88 gmx_bool bAppendSimId,gmx_bool bAppendNodeId,
89 char buf[],int bufsize)
93 if((size_t)bufsize<(strlen(base)+10))
94 gmx_mem("Character buffer too small!");
96 /* Copy to buf, and strip extension */
98 buf[strlen(base) - strlen(ftp2ext(fn2ftp(base))) - 1] = '\0';
101 sprintf(buf+strlen(buf),"%d",cr->ms->sim);
105 sprintf(buf+strlen(buf),"%d",cr->nodeid);
109 /* Add extension again */
110 strcat(buf,(ftp == efTPX) ? "tpr" : (ftp == efEDR) ? "edr" : ftp2ext(ftp));
111 if (cr->nodeid == 0) {
112 printf("node %d par_fn '%s'\n",cr->nodeid,buf);
113 if (fn2ftp(buf) == efLOG) {
119 void check_multi_int(FILE *log,const gmx_multisim_t *ms,int val,
123 gmx_bool bCompatible;
126 fprintf(log,"Multi-checking %s ... ",name);
130 "check_multi_int called with a NULL communication pointer");
134 gmx_sumi_sim(ms->nsim,ibuf,ms);
137 for(p=1; p<ms->nsim; p++)
138 bCompatible = bCompatible && (ibuf[p-1] == ibuf[p]);
149 fprintf(log,"\n%s is not equal for all subsystems\n",name);
150 for(p=0; p<ms->nsim; p++)
151 fprintf(log," subsystem %d: %d\n",p,ibuf[p]);
153 gmx_fatal(FARGS,"The %d subsystems are not compatible\n",ms->nsim);
159 void check_multi_large_int(FILE *log,const gmx_multisim_t *ms,
160 gmx_large_int_t val, const char *name)
162 gmx_large_int_t *ibuf;
164 gmx_bool bCompatible;
167 fprintf(log,"Multi-checking %s ... ",name);
171 "check_multi_int called with a NULL communication pointer");
175 gmx_sumli_sim(ms->nsim,ibuf,ms);
178 for(p=1; p<ms->nsim; p++)
179 bCompatible = bCompatible && (ibuf[p-1] == ibuf[p]);
190 fprintf(log,"\n%s is not equal for all subsystems\n",name);
191 for(p=0; p<ms->nsim; p++)
194 /* first make the format string */
195 snprintf(strbuf, 255, " subsystem %%d: %s\n",
197 fprintf(log,strbuf,p,ibuf[p]);
200 gmx_fatal(FARGS,"The %d subsystems are not compatible\n",ms->nsim);
207 void gmx_log_open(const char *lognm,const t_commrec *cr,gmx_bool bMasterOnly,
208 unsigned long Flags, FILE** fplog)
211 char buf[256],host[256];
213 char timebuf[STRLEN];
217 gmx_bool bAppend = Flags & MD_APPENDFILES;
221 /* Communicate the filename for logfile */
222 if (cr->nnodes > 1 && !bMasterOnly
223 #ifdef GMX_THREAD_MPI
224 /* With thread MPI the non-master log files are opened later
225 * when the files names are already known on all nodes.
233 len = strlen(lognm) + 1;
235 gmx_bcast(sizeof(len),&len,cr);
242 tmpnm=gmx_strdup(lognm);
244 gmx_bcast(len*sizeof(*tmpnm),tmpnm,cr);
248 tmpnm=gmx_strdup(lognm);
253 if (!bMasterOnly && !MASTER(cr))
255 /* Since log always ends with '.log' let's use this info */
256 par_fn(tmpnm,efLOG,cr,FALSE,!bMasterOnly,buf,255);
257 fp = gmx_fio_fopen(buf, bAppend ? "a+" : "w+" );
261 fp = gmx_fio_fopen(tmpnm, bAppend ? "a+" : "w+" );
266 gmx_fatal_set_log_file(fp);
268 /* Get some machine parameters */
270 if (gethostname(host,255) != 0)
272 sprintf(host,"unknown");
275 sprintf(host,"unknown");
281 # ifdef GMX_NATIVE_WINDOWS
295 "-----------------------------------------------------------\n"
296 "Restarting from checkpoint, appending to previous log file.\n"
301 gmx_ctime_r(&t,timebuf,STRLEN);
304 "Log file opened on %s"
305 "Host: %s pid: %d nodeid: %d nnodes: %d\n",
306 timebuf,host,pid,cr->nodeid,cr->nnodes);
308 #if (defined BUILD_MACHINE && defined BUILD_TIME && defined BUILD_USER)
310 "The Gromacs distribution was built %s by\n"
311 "%s (%s)\n\n\n",BUILD_TIME,BUILD_USER,BUILD_MACHINE);
320 void gmx_log_close(FILE *fp)
323 gmx_fatal_set_log_file(NULL);
328 static void comm_args(const t_commrec *cr,int *argc,char ***argv)
333 gmx_bcast(sizeof(*argc),argc,cr);
337 fprintf(stderr,"NODEID=%d argc=%d\n",cr->nodeid,*argc);
338 for(i=0; (i<*argc); i++) {
340 len = strlen((*argv)[i])+1;
341 gmx_bcast(sizeof(len),&len,cr);
343 snew((*argv)[i],len);
344 /*gmx_bcast(len*sizeof((*argv)[i][0]),(*argv)[i],cr);*/
345 gmx_bcast(len*sizeof(char),(*argv)[i],cr);
350 void init_multisystem(t_commrec *cr,int nsim, char **multidirs,
351 int nfile, const t_filenm fnm[],gmx_bool bParFn)
354 int nnodes,nnodpersim,sim,i,ftp;
357 MPI_Group mpi_group_world;
364 gmx_fatal(FARGS,"This binary is compiled without MPI support, can not do multiple simulations.");
369 if (nnodes % nsim != 0)
371 gmx_fatal(FARGS,"The number of nodes (%d) is not a multiple of the number of simulations (%d)",nnodes,nsim);
374 nnodpersim = nnodes/nsim;
375 sim = cr->nodeid/nnodpersim;
379 fprintf(debug,"We have %d simulations, %d nodes per simulation, local simulation is %d\n",nsim,nnodpersim,sim);
387 /* Create a communicator for the master nodes */
389 for(i=0; i<ms->nsim; i++)
391 rank[i] = i*nnodpersim;
393 MPI_Comm_group(MPI_COMM_WORLD,&mpi_group_world);
394 MPI_Group_incl(mpi_group_world,nsim,rank,&ms->mpi_group_masters);
396 MPI_Comm_create(MPI_COMM_WORLD,ms->mpi_group_masters,
397 &ms->mpi_comm_masters);
399 #if !defined(GMX_THREAD_MPI) && !defined(MPI_IN_PLACE_EXISTS)
400 /* initialize the MPI_IN_PLACE replacement buffers */
406 ms->mpb->ibuf_alloc=0;
407 ms->mpb->libuf_alloc=0;
408 ms->mpb->fbuf_alloc=0;
409 ms->mpb->dbuf_alloc=0;
414 /* Reduce the intra-simulation communication */
415 cr->sim_nodeid = cr->nodeid % nnodpersim;
416 cr->nnodes = nnodpersim;
418 MPI_Comm_split(MPI_COMM_WORLD,sim,cr->sim_nodeid,&cr->mpi_comm_mysim);
419 cr->mpi_comm_mygroup = cr->mpi_comm_mysim;
420 cr->nodeid = cr->sim_nodeid;
425 fprintf(debug,"This is simulation %d",cr->ms->sim);
428 fprintf(debug,", local number of nodes %d, local nodeid %d",
429 cr->nnodes,cr->sim_nodeid);
431 fprintf(debug,"\n\n");
439 fprintf(debug,"Changing to directory %s\n",multidirs[cr->ms->sim]);
441 gmx_chdir(multidirs[cr->ms->sim]);
445 /* Patch output and tpx, cpt and rerun input file names */
446 for(i=0; (i<nfile); i++)
448 /* Because of possible multiple extensions per type we must look
449 * at the actual file name
451 if (is_output(&fnm[i]) ||
452 fnm[i].ftp == efTPX || fnm[i].ftp == efCPT ||
453 strcmp(fnm[i].opt,"-rerun") == 0)
455 ftp = fn2ftp(fnm[i].fns[0]);
456 par_fn(fnm[i].fns[0],ftp,cr,TRUE,FALSE,buf,255);
457 sfree(fnm[i].fns[0]);
458 fnm[i].fns[0] = gmx_strdup(buf);
464 t_commrec *init_par(int *argc,char ***argv_ptr)
475 #if defined GMX_MPI && !defined GMX_THREAD_MPI
476 cr->sim_nodeid = gmx_setup(argc,argv,&cr->nnodes);
478 if (!PAR(cr) && (cr->sim_nodeid != 0))
480 gmx_comm("(!PAR(cr) && (cr->sim_nodeid != 0))");
483 cr->mpi_comm_mysim = MPI_COMM_WORLD;
484 cr->mpi_comm_mygroup = cr->mpi_comm_mysim;
486 /* These should never be accessed */
487 cr->mpi_comm_mysim = NULL;
488 cr->mpi_comm_mygroup = NULL;
493 cr->nodeid = cr->sim_nodeid;
495 cr->duty = (DUTY_PP | DUTY_PME);
497 /* Communicate arguments if parallel */
498 #ifndef GMX_THREAD_MPI
501 comm_args(cr,argc,argv_ptr);
503 #endif /* GMX_THREAD_MPI */
506 #if !defined(GMX_THREAD_MPI) && !defined(MPI_IN_PLACE_EXISTS)
507 /* initialize the MPI_IN_PLACE replacement buffers */
513 cr->mpb->ibuf_alloc=0;
514 cr->mpb->libuf_alloc=0;
515 cr->mpb->fbuf_alloc=0;
516 cr->mpb->dbuf_alloc=0;
523 t_commrec *init_par_threads(const t_commrec *cro)
525 #ifdef GMX_THREAD_MPI
529 /* make a thread-specific commrec */
531 /* now copy the whole thing, so settings like the number of PME nodes
535 /* and we start setting our own thread-specific values for things */
536 MPI_Initialized(&initialized);
539 gmx_comm("Initializing threads without comm");
541 /* once threads will be used together with MPI, we'll
542 fill the cr structure with distinct data here. This might even work: */
543 cr->sim_nodeid = gmx_setup(0,NULL, &cr->nnodes);
545 cr->mpi_comm_mysim = MPI_COMM_WORLD;
546 cr->mpi_comm_mygroup = cr->mpi_comm_mysim;
547 cr->nodeid = cr->sim_nodeid;
548 cr->duty = (DUTY_PP | DUTY_PME);