src/mdlib/gmx_wallcycle.c

   1 /*  -*- mode: c; tab-width: 4; indent-tabs-mode: nil; c-basic-offset: 4; c-file-style: "stroustrup"; -*-
   2  *
   3  *
   4  *                This source code is part of
   5  *
   6  *                 G   R   O   M   A   C   S
   7  *
   8  *          GROningen MAchine for Chemical Simulations
   9  *
  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) 2001-2008, The GROMACS development team,
  13  * check out http://www.gromacs.org for more information.
  14
  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.
  19  *
  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.
  26  *
  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.
  29  *
  30  * For more info, check our website at http://www.gromacs.org
  31  *
  32  * And Hey:
  33  * Gallium Rubidium Oxygen Manganese Argon Carbon Silicon
  34  */
  35
  36
  37 #ifdef HAVE_CONFIG_H
  38 #include <config.h>
  39 #endif
  40
  41 #include <string.h>
  42 #include "gmx_wallcycle.h"
  43 #include "gmx_cyclecounter.h"
  44 #include "smalloc.h"
  45 #include "gmx_fatal.h"
  46
  47 #ifdef GMX_LIB_MPI
  48 #include <mpi.h>
  49 #endif
  50 #ifdef GMX_THREADS
  51 #include "tmpi.h"
  52 #endif
  53
  54 typedef struct
  55 {
  56     int          n;
  57     gmx_cycles_t c;
  58     gmx_cycles_t start;
  59     gmx_cycles_t last;
  60 } wallcc_t;
  61
  62 typedef struct gmx_wallcycle
  63 {
  64     wallcc_t     *wcc;
  65     /* variables for testing/debugging */
  66     bool         wc_barrier;
  67     wallcc_t     *wcc_all;
  68     int          wc_depth;
  69     int          ewc_prev;
  70     gmx_cycles_t cycle_prev;
  71     gmx_large_int_t   reset_counters;
  72 #ifdef GMX_MPI
  73     MPI_Comm     mpi_comm_mygroup;
  74 #endif
  75 } gmx_wallcycle_t_t;
  76
  77 /* Each name should not exceed 19 characters */
  78 static const char *wcn[ewcNR] =
  79 { "Run", "Step", "PP during PME", "Domain decomp.", "DD comm. load", "DD comm. bounds", "Vsite constr.", "Send X to PME", "Comm. coord.", "Neighbor search", "Born radii", "Force", "Wait + Comm. F", "PME mesh", "PME redist. X/F", "PME spread/gather", "PME 3D-FFT", "PME solve", "Wait + Comm. X/F", "Wait + Recv. PME F", "Vsite spread", "Write traj.", "Update", "Constraints", "Comm. energies", "Test" };
  80
  81 bool wallcycle_have_counter(void)
  82 {
  83   return gmx_cycles_have_counter();
  84 }
  85
  86 gmx_wallcycle_t wallcycle_init(FILE *fplog,int resetstep,t_commrec *cr)
  87 {
  88     gmx_wallcycle_t wc;
  89
  90
  91     if (!wallcycle_have_counter())
  92     {
  93         return NULL;
  94     }
  95
  96     snew(wc,1);
  97
  98     wc->wc_barrier = FALSE;
  99     wc->wcc_all    = NULL;
 100     wc->wc_depth   = 0;
 101     wc->ewc_prev   = -1;
 102     wc->reset_counters = resetstep;
 103
 104 #ifdef GMX_MPI
 105     if (PAR(cr) && getenv("GMX_CYCLE_BARRIER") != NULL)
 106     {
 107         if (fplog)
 108         {
 109             fprintf(fplog,"\nWill call MPI_Barrier before each cycle start/stop call\n\n");
 110         }
 111         wc->wc_barrier = TRUE;
 112         wc->mpi_comm_mygroup = cr->mpi_comm_mygroup;
 113     }
 114 #endif
 115
 116     snew(wc->wcc,ewcNR);
 117     if (getenv("GMX_CYCLE_ALL") != NULL)
 118     {
 119 /*#ifndef GMX_THREADS*/
 120         if (fplog)
 121         {
 122             fprintf(fplog,"\nWill time all the code during the run\n\n");
 123         }
 124         snew(wc->wcc_all,ewcNR*ewcNR);
 125 /*#else*/
 126         gmx_fatal(FARGS, "GMX_CYCLE_ALL is incompatible with threaded code");
 127 /*#endif*/
 128     }
 129
 130     return wc;
 131 }
 132
 133 void wallcycle_destroy(gmx_wallcycle_t wc)
 134 {
 135     if (wc == NULL)
 136     {
 137         return;
 138     }
 139
 140     if (wc->wcc != NULL)
 141     {
 142         sfree(wc->wcc);
 143     }
 144     if (wc->wcc_all != NULL)
 145     {
 146         sfree(wc->wcc_all);
 147     }
 148     sfree(wc);
 149 }
 150
 151 static void wallcycle_all_start(gmx_wallcycle_t wc,int ewc,gmx_cycles_t cycle)
 152 {
 153     wc->ewc_prev = ewc;
 154     wc->cycle_prev = cycle;
 155 }
 156
 157 static void wallcycle_all_stop(gmx_wallcycle_t wc,int ewc,gmx_cycles_t cycle)
 158 {
 159     wc->wcc_all[wc->ewc_prev*ewcNR+ewc].n += 1;
 160     wc->wcc_all[wc->ewc_prev*ewcNR+ewc].c += cycle - wc->cycle_prev;
 161 }
 162
 163 void wallcycle_start(gmx_wallcycle_t wc, int ewc)
 164 {
 165     gmx_cycles_t cycle;
 166
 167     if (wc == NULL)
 168     {
 169         return;
 170     }
 171
 172 #ifdef GMX_MPI
 173     if (wc->wc_barrier)
 174     {
 175         MPI_Barrier(wc->mpi_comm_mygroup);
 176     }
 177 #endif
 178
 179     cycle = gmx_cycles_read();
 180     wc->wcc[ewc].start = cycle;
 181     if (wc->wcc_all != NULL)
 182     {
 183         wc->wc_depth++;
 184         if (ewc == ewcRUN)
 185         {
 186             wallcycle_all_start(wc,ewc,cycle);
 187         }
 188         else if (wc->wc_depth == 3)
 189         {
 190             wallcycle_all_stop(wc,ewc,cycle);
 191         }
 192     }
 193 }
 194
 195 double wallcycle_stop(gmx_wallcycle_t wc, int ewc)
 196 {
 197     gmx_cycles_t cycle,last;
 198
 199     if (wc == NULL)
 200     {
 201         return 0;
 202     }
 203
 204 #ifdef GMX_MPI
 205     if (wc->wc_barrier)
 206     {
 207         MPI_Barrier(wc->mpi_comm_mygroup);
 208     }
 209 #endif
 210
 211     cycle = gmx_cycles_read();
 212     last = cycle - wc->wcc[ewc].start;
 213     wc->wcc[ewc].c += last;
 214     wc->wcc[ewc].n++;
 215     if (wc->wcc_all)
 216     {
 217         wc->wc_depth--;
 218         if (ewc == ewcRUN)
 219         {
 220             wallcycle_all_stop(wc,ewc,cycle);
 221         }
 222         else if (wc->wc_depth == 2)
 223         {
 224             wallcycle_all_start(wc,ewc,cycle);
 225         }
 226     }
 227
 228     return last;
 229 }
 230
 231 void wallcycle_reset_all(gmx_wallcycle_t wc)
 232 {
 233     int i;
 234
 235     if (wc == NULL)
 236     {
 237         return;
 238     }
 239
 240     for(i=0; i<ewcNR; i++)
 241     {
 242         wc->wcc[i].n = 0;
 243         wc->wcc[i].c = 0;
 244         wc->wcc[i].start = 0;
 245         wc->wcc[i].last = 0;
 246     }
 247 }
 248
 249 void wallcycle_sum(t_commrec *cr, gmx_wallcycle_t wc,double cycles[])
 250 {
 251     wallcc_t *wcc;
 252     double cycles_n[ewcNR],buf[ewcNR],*cyc_all,*buf_all;
 253     int    i;
 254
 255     if (wc == NULL)
 256     {
 257         return;
 258     }
 259
 260     wcc = wc->wcc;
 261
 262     if (wcc[ewcDDCOMMLOAD].n > 0)
 263     {
 264         wcc[ewcDOMDEC].c -= wcc[ewcDDCOMMLOAD].c;
 265     }
 266     if (wcc[ewcDDCOMMBOUND].n > 0)
 267     {
 268         wcc[ewcDOMDEC].c -= wcc[ewcDDCOMMBOUND].c;
 269     }
 270     if (cr->npmenodes == 0)
 271     {
 272         /* All nodes do PME (or no PME at all) */
 273         if (wcc[ewcPMEMESH].n > 0)
 274         {
 275             wcc[ewcFORCE].c -= wcc[ewcPMEMESH].c;
 276         }
 277     }
 278     else
 279     {
 280         /* The are PME-only nodes */
 281         if (wcc[ewcPMEMESH].n > 0)
 282         {
 283             /* This must be a PME only node, calculate the Wait + Comm. time */
 284             wcc[ewcPMEWAITCOMM].c = wcc[ewcRUN].c - wcc[ewcPMEMESH].c;
 285         }
 286     }
 287
 288     /* Store the cycles in a double buffer for summing */
 289     for(i=0; i<ewcNR; i++)
 290     {
 291         cycles_n[i] = (double)wcc[i].n;
 292         cycles[i]   = (double)wcc[i].c;
 293     }
 294
 295 #ifdef GMX_MPI
 296     if (cr->nnodes > 1)
 297     {
 298         MPI_Allreduce(cycles_n,buf,ewcNR,MPI_DOUBLE,MPI_MAX,
 299                       cr->mpi_comm_mysim);
 300         for(i=0; i<ewcNR; i++)
 301         {
 302             wcc[i].n = (int)(buf[i] + 0.5);
 303         }
 304         MPI_Allreduce(cycles,buf,ewcNR,MPI_DOUBLE,MPI_SUM,
 305                       cr->mpi_comm_mysim);
 306         for(i=0; i<ewcNR; i++)
 307         {
 308             cycles[i] = buf[i];
 309         }
 310
 311         if (wc->wcc_all != NULL)
 312         {
 313             snew(cyc_all,ewcNR*ewcNR);
 314             snew(buf_all,ewcNR*ewcNR);
 315             for(i=0; i<ewcNR*ewcNR; i++)
 316             {
 317                 cyc_all[i] = wc->wcc_all[i].c;
 318             }
 319             MPI_Allreduce(cyc_all,buf_all,ewcNR*ewcNR,MPI_DOUBLE,MPI_SUM,
 320                           cr->mpi_comm_mysim);
 321             for(i=0; i<ewcNR*ewcNR; i++)
 322             {
 323                 wc->wcc_all[i].c = buf_all[i];
 324             }
 325             sfree(buf_all);
 326             sfree(cyc_all);
 327         }
 328     }
 329 #endif
 330 }
 331
 332 static void print_cycles(FILE *fplog, double c2t, const char *name, int nnodes,
 333                          int n, gmx_cycles_t c, gmx_cycles_t tot)
 334 {
 335   char num[11];
 336
 337   if (c > 0) {
 338     if (n > 0)
 339       sprintf(num,"%10d",n);
 340     else
 341       sprintf(num,"          ");
 342     fprintf(fplog," %-19s %4d %10s %12.3f %10.1f   %5.1f\n",
 343             name,nnodes,num,c*1e-9,c*c2t,100*(double)c/(double)tot);
 344   }
 345 }
 346
 347 static bool subdivision(int ewc)
 348 {
 349     return (ewc >= ewcPME_REDISTXF && ewc <= ewcPME_SOLVE);
 350 }
 351
 352 void wallcycle_print(FILE *fplog, int nnodes, int npme, double realtime,
 353                      gmx_wallcycle_t wc, double cycles[])
 354 {
 355     double c2t,tot,sum;
 356     int    i,j,npp;
 357     char   buf[STRLEN];
 358     const char *myline = "-----------------------------------------------------------------------";
 359
 360     if (wc == NULL)
 361     {
 362         return;
 363     }
 364
 365     if (npme > 0)
 366     {
 367         npp = nnodes - npme;
 368     }
 369     else
 370     {
 371         npp  = nnodes;
 372         npme = nnodes;
 373     }
 374     tot = cycles[ewcRUN];
 375     /* Conversion factor from cycles to seconds */
 376     if (tot > 0)
 377     {
 378       c2t = nnodes*realtime/tot;
 379     }
 380     else
 381     {
 382       c2t = 0;
 383     }
 384
 385     fprintf(fplog,"\n     R E A L   C Y C L E   A N D   T I M E   A C C O U N T I N G\n\n");
 386
 387     fprintf(fplog," Computing:         Nodes     Number     G-Cycles    Seconds     %c\n",'%');
 388     fprintf(fplog,"%s\n",myline);
 389     sum = 0;
 390     for(i=ewcPPDURINGPME+1; i<ewcNR; i++)
 391     {
 392         if (!subdivision(i))
 393         {
 394             print_cycles(fplog,c2t,wcn[i],
 395                          (i==ewcPMEMESH || i==ewcPMEWAITCOMM) ? npme : npp,
 396                          wc->wcc[i].n,cycles[i],tot);
 397             sum += cycles[i];
 398         }
 399     }
 400     if (wc->wcc_all != NULL)
 401     {
 402         for(i=0; i<ewcNR; i++)
 403         {
 404             for(j=0; j<ewcNR; j++)
 405             {
 406                 sprintf(buf,"%-9s",wcn[i]);
 407                 buf[9] = ' ';
 408                 sprintf(buf+10,"%-9s",wcn[j]);
 409                 buf[19] = '\0';
 410                 print_cycles(fplog,c2t,buf,
 411                              (i==ewcPMEMESH || i==ewcPMEWAITCOMM) ? npme : npp,
 412                              wc->wcc_all[i*ewcNR+j].n,
 413                              wc->wcc_all[i*ewcNR+j].c,
 414                              tot);
 415             }
 416         }
 417     }
 418     print_cycles(fplog,c2t,"Rest",npp,0,tot-sum,tot);
 419     fprintf(fplog,"%s\n",myline);
 420     print_cycles(fplog,c2t,"Total",nnodes,0,tot,tot);
 421     fprintf(fplog,"%s\n",myline);
 422
 423     if (wc->wcc[ewcPMEMESH].n > 0)
 424     {
 425         fprintf(fplog,"%s\n",myline);
 426         for(i=ewcPPDURINGPME+1; i<ewcNR; i++)
 427         {
 428             if (subdivision(i))
 429             {
 430                 print_cycles(fplog,c2t,wcn[i],
 431                              (i>=ewcPMEMESH || i<=ewcPME_SOLVE) ? npme : npp,
 432                              wc->wcc[i].n,cycles[i],tot);
 433             }
 434         }
 435         fprintf(fplog,"%s\n",myline);
 436     }
 437
 438     if (cycles[ewcMoveE] > tot*0.05)
 439     {
 440         sprintf(buf,
 441                 "NOTE: %d %% of the run time was spent communicating energies,\n"
 442                 "      you might want to use the -gcom option of mdrun\n",
 443                 (int)(100*cycles[ewcMoveE]/tot+0.5));
 444         if (fplog)
 445         {
 446             fprintf(fplog,"\n%s\n",buf);
 447         }
 448         /* Only the sim master calls this function, so always print to stderr */
 449         fprintf(stderr,"\n%s\n",buf);
 450     }
 451 }
 452
 453 extern gmx_large_int_t wcycle_get_reset_counters(gmx_wallcycle_t wc)
 454 {
 455     if (wc == NULL)
 456     {
 457         return -1;
 458     }
 459
 460     return wc->reset_counters;
 461 }
 462
 463 extern void wcycle_set_reset_counters(gmx_wallcycle_t wc, gmx_large_int_t reset_counters)
 464 {
 465     if (wc == NULL)
 466         return;
 467
 468     wc->reset_counters = reset_counters;
 469 }