src/kernel/pme_loadbal.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  *                        VERSION 4.6.0
  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) 2001-2011, The GROMACS development team,
  14  * check out http://www.gromacs.org for more information.
  15
  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.
  20  *
  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.
  27  *
  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.
  30  *
  31  * For more info, check our website at http://www.gromacs.org
  32  *
  33  * And Hey:
  34  * Gallium Rubidium Oxygen Manganese Argon Carbon Silicon
  35  */
  36 #ifdef HAVE_CONFIG_H
  37 #include <config.h>
  38 #endif
  39
  40 #include "smalloc.h"
  41 #include "network.h"
  42 #include "calcgrid.h"
  43 #include "pme.h"
  44 #include "vec.h"
  45 #include "domdec.h"
  46 #include "nbnxn_cuda_data_mgmt.h"
  47 #include "force.h"
  48 #include "pme_loadbal.h"
  49
  50 /* Parameters and setting for one PP-PME setup */
  51 typedef struct {
  52     real rcut_coulomb;    /* Coulomb cut-off                              */
  53     real rlist;           /* pair-list cut-off                            */
  54     real rlistlong;       /* LR pair-list cut-off                         */
  55     int  nstcalclr;       /* frequency of evaluating long-range forces for group scheme */
  56     real spacing;         /* (largest) PME grid spacing                   */
  57     ivec grid;            /* the PME grid dimensions                      */
  58     real grid_efficiency; /* ineffiency factor for non-uniform grids <= 1 */
  59     real ewaldcoeff;      /* the Ewald coefficient                        */
  60     gmx_pme_t pmedata;    /* the data structure used in the PME code      */
  61
  62     int  count;           /* number of times this setup has been timed    */
  63     double cycles;        /* the fastest time for this setup in cycles    */
  64 } pme_setup_t;
  65
  66 /* In the initial scan, step by grids that are at least a factor 0.8 coarser */
  67 #define PME_LB_GRID_SCALE_FAC  0.8
  68 /* In the initial scan, try to skip grids with uneven x/y/z spacing,
  69  * checking if the "efficiency" is more than 5% worse than the previous grid.
  70  */
  71 #define PME_LB_GRID_EFFICIENCY_REL_FAC  1.05
  72 /* Rerun up till 12% slower setups than the fastest up till now */
  73 #define PME_LB_SLOW_FAC  1.12
  74 /* If setups get more than 2% faster, do another round to avoid
  75  * choosing a slower setup due to acceleration or fluctuations.
  76  */
  77 #define PME_LB_ACCEL_TOL 1.02
  78
  79 enum { epmelblimNO, epmelblimBOX, epmelblimDD, epmelblimNR };
  80
  81 const char *pmelblim_str[epmelblimNR] =
  82 { "no", "box size", "domain decompostion" };
  83
  84 struct pme_load_balancing {
  85     int  nstage;        /* the current maximum number of stages */
  86
  87     real cut_spacing;   /* the minimum cutoff / PME grid spacing ratio */
  88     real rcut_vdw;      /* Vdw cutoff (does not change) */
  89     real rcut_coulomb_start; /* Initial electrostatics cutoff */
  90     int  nstcalclr_start; /* Initial electrostatics cutoff */
  91     real rbuf_coulomb;  /* the pairlist buffer size */
  92     real rbuf_vdw;      /* the pairlist buffer size */
  93     matrix box_start;   /* the initial simulation box */
  94     int n;              /* the count of setup as well as the allocation size */
  95     pme_setup_t *setup; /* the PME+cutoff setups */
  96     int cur;            /* the current setup */
  97     int fastest;        /* fastest setup up till now */
  98     int start;          /* start of setup range to consider in stage>0 */
  99     int end;            /* end   of setup range to consider in stage>0 */
 100     int elimited;       /* was the balancing limited, uses enum above */
 101     int cutoff_scheme;  /* Verlet or group cut-offs */
 102
 103     int stage;          /* the current stage */
 104 };
 105
 106 void pme_loadbal_init(pme_load_balancing_t *pme_lb_p,
 107                       const t_inputrec *ir,matrix box,
 108                       const interaction_const_t *ic,
 109                       gmx_pme_t pmedata)
 110 {
 111     pme_load_balancing_t pme_lb;
 112     real spm,sp;
 113     int  d;
 114
 115     snew(pme_lb,1);
 116
 117     /* Any number of stages >= 2 is supported */
 118     pme_lb->nstage   = 2;
 119
 120     pme_lb->cutoff_scheme = ir->cutoff_scheme;
 121
 122     if(pme_lb->cutoff_scheme == ecutsVERLET)
 123     {
 124         pme_lb->rbuf_coulomb = ic->rlist - ic->rcoulomb;
 125         pme_lb->rbuf_vdw     = pme_lb->rbuf_coulomb;
 126     }
 127     else
 128     {
 129         if(ic->rcoulomb > ic->rlist)
 130         {
 131             pme_lb->rbuf_coulomb = ic->rlistlong - ic->rcoulomb;
 132         }
 133         else
 134         {
 135             pme_lb->rbuf_coulomb = ic->rlist - ic->rcoulomb;
 136         }
 137         if(ic->rvdw > ic->rlist)
 138         {
 139             pme_lb->rbuf_vdw = ic->rlistlong - ic->rvdw;
 140         }
 141         else
 142         {
 143             pme_lb->rbuf_vdw = ic->rlist - ic->rvdw;
 144         }
 145     }
 146
 147     copy_mat(box,pme_lb->box_start);
 148     if (ir->ePBC==epbcXY && ir->nwall==2)
 149     {
 150         svmul(ir->wall_ewald_zfac,pme_lb->box_start[ZZ],pme_lb->box_start[ZZ]);
 151     }
 152
 153     pme_lb->n = 1;
 154     snew(pme_lb->setup,pme_lb->n);
 155
 156     pme_lb->rcut_vdw              = ic->rvdw;
 157     pme_lb->rcut_coulomb_start    = ir->rcoulomb;
 158     pme_lb->nstcalclr_start       = ir->nstcalclr;
 159
 160     pme_lb->cur = 0;
 161     pme_lb->setup[0].rcut_coulomb = ic->rcoulomb;
 162     pme_lb->setup[0].rlist        = ic->rlist;
 163     pme_lb->setup[0].rlistlong    = ic->rlistlong;
 164     pme_lb->setup[0].nstcalclr    = ir->nstcalclr;
 165     pme_lb->setup[0].grid[XX]     = ir->nkx;
 166     pme_lb->setup[0].grid[YY]     = ir->nky;
 167     pme_lb->setup[0].grid[ZZ]     = ir->nkz;
 168     pme_lb->setup[0].ewaldcoeff   = ic->ewaldcoeff;
 169
 170     pme_lb->setup[0].pmedata  = pmedata;
 171
 172     spm = 0;
 173     for(d=0; d<DIM; d++)
 174     {
 175         sp = norm(pme_lb->box_start[d])/pme_lb->setup[0].grid[d];
 176         if (sp > spm)
 177         {
 178             spm = sp;
 179         }
 180     }
 181     pme_lb->setup[0].spacing = spm;
 182
 183     if (ir->fourier_spacing > 0)
 184     {
 185         pme_lb->cut_spacing = ir->rcoulomb/ir->fourier_spacing;
 186     }
 187     else
 188     {
 189         pme_lb->cut_spacing = ir->rcoulomb/pme_lb->setup[0].spacing;
 190     }
 191
 192     pme_lb->stage = 0;
 193
 194     pme_lb->fastest  = 0;
 195     pme_lb->start    = 0;
 196     pme_lb->end      = 0;
 197     pme_lb->elimited = epmelblimNO;
 198
 199     *pme_lb_p = pme_lb;
 200 }
 201
 202 static gmx_bool pme_loadbal_increase_cutoff(pme_load_balancing_t pme_lb,
 203                                             int pme_order)
 204 {
 205     pme_setup_t *set;
 206     real fac,sp;
 207     real tmpr_coulomb,tmpr_vdw;
 208     int d;
 209
 210     /* Try to add a new setup with next larger cut-off to the list */
 211     pme_lb->n++;
 212     srenew(pme_lb->setup,pme_lb->n);
 213     set = &pme_lb->setup[pme_lb->n-1];
 214     set->pmedata = NULL;
 215
 216     fac = 1;
 217     do
 218     {
 219         fac *= 1.01;
 220         clear_ivec(set->grid);
 221         sp = calc_grid(NULL,pme_lb->box_start,
 222                        fac*pme_lb->setup[pme_lb->cur].spacing,
 223                        &set->grid[XX],
 224                        &set->grid[YY],
 225                        &set->grid[ZZ]);
 226
 227         /* In parallel we can't have grids smaller than 2*pme_order,
 228          * and we would anyhow not gain much speed at these grid sizes.
 229          */
 230         for(d=0; d<DIM; d++)
 231         {
 232             if (set->grid[d] <= 2*pme_order)
 233             {
 234                 pme_lb->n--;
 235
 236                 return FALSE;
 237             }
 238         }
 239     }
 240     while (sp <= 1.001*pme_lb->setup[pme_lb->cur].spacing);
 241
 242     set->rcut_coulomb = pme_lb->cut_spacing*sp;
 243
 244     if(pme_lb->cutoff_scheme == ecutsVERLET)
 245     {
 246         set->rlist        = set->rcut_coulomb + pme_lb->rbuf_coulomb;
 247         /* We dont use LR lists with Verlet, but this avoids if-statements in further checks */
 248         set->rlistlong    = set->rlist;
 249     }
 250     else
 251     {
 252         tmpr_coulomb          = set->rcut_coulomb + pme_lb->rbuf_coulomb;
 253         tmpr_vdw              = pme_lb->rcut_vdw + pme_lb->rbuf_vdw;
 254         set->rlist            = min(tmpr_coulomb,tmpr_vdw);
 255         set->rlistlong        = max(tmpr_coulomb,tmpr_vdw);
 256
 257         /* Set the long-range update frequency */
 258         if(set->rlist == set->rlistlong)
 259         {
 260             /* No long-range interactions if the short-/long-range cutoffs are identical */
 261             set->nstcalclr = 0;
 262         }
 263         else if(pme_lb->nstcalclr_start==0 || pme_lb->nstcalclr_start==1)
 264         {
 265             /* We were not doing long-range before, but now we are since rlist!=rlistlong */
 266             set->nstcalclr = 1;
 267         }
 268         else
 269         {
 270             /* We were already doing long-range interactions from the start */
 271             if(pme_lb->rcut_vdw > pme_lb->rcut_coulomb_start)
 272             {
 273                 /* We were originally doing long-range VdW-only interactions.
 274                  * If rvdw is still longer than rcoulomb we keep the original nstcalclr,
 275                  * but if the coulomb cutoff has become longer we should update the long-range
 276                  * part every step.
 277                  */
 278                 set->nstcalclr = (tmpr_vdw > tmpr_coulomb) ? pme_lb->nstcalclr_start : 1;
 279             }
 280             else
 281             {
 282                 /* We were not doing any long-range interaction from the start,
 283                  * since it is not possible to do twin-range coulomb for the PME interaction.
 284                  */
 285                 set->nstcalclr = 1;
 286             }
 287         }
 288     }
 289
 290     set->spacing      = sp;
 291     /* The grid efficiency is the size wrt a grid with uniform x/y/z spacing */
 292     set->grid_efficiency = 1;
 293     for(d=0; d<DIM; d++)
 294     {
 295         set->grid_efficiency *= (set->grid[d]*sp)/norm(pme_lb->box_start[d]);
 296     }
 297     /* The Ewald coefficient is inversly proportional to the cut-off */
 298     set->ewaldcoeff =
 299         pme_lb->setup[0].ewaldcoeff*pme_lb->setup[0].rcut_coulomb/set->rcut_coulomb;
 300
 301     set->count   = 0;
 302     set->cycles  = 0;
 303
 304     if (debug)
 305     {
 306         fprintf(debug,"PME loadbal: grid %d %d %d, coulomb cutoff %f\n",
 307                 set->grid[XX],set->grid[YY],set->grid[ZZ],set->rcut_coulomb);
 308     }
 309     return TRUE;
 310 }
 311
 312 static void print_grid(FILE *fp_err,FILE *fp_log,
 313                        const char *pre,
 314                        const char *desc,
 315                        const pme_setup_t *set,
 316                        double cycles)
 317 {
 318     char buf[STRLEN],buft[STRLEN];
 319
 320     if (cycles >= 0)
 321     {
 322         sprintf(buft,": %.1f M-cycles",cycles*1e-6);
 323     }
 324     else
 325     {
 326         buft[0] = '\0';
 327     }
 328     sprintf(buf,"%-11s%10s pme grid %d %d %d, coulomb cutoff %.3f%s",
 329             pre,
 330             desc,set->grid[XX],set->grid[YY],set->grid[ZZ],set->rcut_coulomb,
 331             buft);
 332     if (fp_err != NULL)
 333     {
 334         fprintf(fp_err,"\r%s\n",buf);
 335     }
 336     if (fp_log != NULL)
 337     {
 338         fprintf(fp_log,"%s\n",buf);
 339     }
 340 }
 341
 342 static int pme_loadbal_end(pme_load_balancing_t pme_lb)
 343 {
 344     /* In the initial stage only n is set; end is not set yet */
 345     if (pme_lb->end > 0)
 346     {
 347         return pme_lb->end;
 348     }
 349     else
 350     {
 351         return pme_lb->n;
 352     }
 353 }
 354
 355 static void print_loadbal_limited(FILE *fp_err,FILE *fp_log,
 356                                   gmx_large_int_t step,
 357                                   pme_load_balancing_t pme_lb)
 358 {
 359     char buf[STRLEN],sbuf[22];
 360
 361     sprintf(buf,"step %4s: the %s limited the PME load balancing to a coulomb cut-off of %.3f",
 362             gmx_step_str(step,sbuf),
 363             pmelblim_str[pme_lb->elimited],
 364             pme_lb->setup[pme_loadbal_end(pme_lb)-1].rcut_coulomb);
 365     if (fp_err != NULL)
 366     {
 367         fprintf(fp_err,"\r%s\n",buf);
 368     }
 369     if (fp_log != NULL)
 370     {
 371         fprintf(fp_log,"%s\n",buf);
 372     }
 373 }
 374
 375 static void switch_to_stage1(pme_load_balancing_t pme_lb)
 376 {
 377     pme_lb->start = 0;
 378     while (pme_lb->start+1 < pme_lb->n &&
 379            (pme_lb->setup[pme_lb->start].count == 0 ||
 380             pme_lb->setup[pme_lb->start].cycles >
 381             pme_lb->setup[pme_lb->fastest].cycles*PME_LB_SLOW_FAC))
 382     {
 383         pme_lb->start++;
 384     }
 385     while (pme_lb->start > 0 && pme_lb->setup[pme_lb->start-1].cycles == 0)
 386     {
 387         pme_lb->start--;
 388     }
 389
 390     pme_lb->end = pme_lb->n;
 391     if (pme_lb->setup[pme_lb->end-1].count > 0 &&
 392         pme_lb->setup[pme_lb->end-1].cycles >
 393         pme_lb->setup[pme_lb->fastest].cycles*PME_LB_SLOW_FAC)
 394     {
 395         pme_lb->end--;
 396     }
 397
 398     pme_lb->stage = 1;
 399
 400     /* Next we want to choose setup pme_lb->start, but as we will increase
 401      * pme_ln->cur by one right after returning, we subtract 1 here.
 402      */
 403     pme_lb->cur = pme_lb->start - 1;
 404 }
 405
 406 gmx_bool pme_load_balance(pme_load_balancing_t pme_lb,
 407                           t_commrec *cr,
 408                           FILE *fp_err,
 409                           FILE *fp_log,
 410                           t_inputrec *ir,
 411                           t_state *state,
 412                           double cycles,
 413                           interaction_const_t *ic,
 414                           nonbonded_verlet_t *nbv,
 415                           gmx_pme_t *pmedata,
 416                           gmx_large_int_t step)
 417 {
 418     gmx_bool OK;
 419     pme_setup_t *set;
 420     double cycles_fast;
 421     char buf[STRLEN],sbuf[22];
 422     real rtab;
 423     gmx_bool bUsesSimpleTables = TRUE;
 424
 425     if (pme_lb->stage == pme_lb->nstage)
 426     {
 427         return FALSE;
 428     }
 429
 430     if (PAR(cr))
 431     {
 432         gmx_sumd(1,&cycles,cr);
 433         cycles /= cr->nnodes;
 434     }
 435
 436     set = &pme_lb->setup[pme_lb->cur];
 437     set->count++;
 438
 439     rtab = ir->rlistlong + ir->tabext;
 440
 441     if (set->count % 2 == 1)
 442     {
 443         /* Skip the first cycle, because the first step after a switch
 444          * is much slower due to allocation and/or caching effects.
 445          */
 446         return TRUE;
 447     }
 448
 449     sprintf(buf, "step %4s: ", gmx_step_str(step,sbuf));
 450     print_grid(fp_err,fp_log,buf,"timed with",set,cycles);
 451
 452     if (set->count <= 2)
 453     {
 454         set->cycles = cycles;
 455     }
 456     else
 457     {
 458         if (cycles*PME_LB_ACCEL_TOL < set->cycles &&
 459             pme_lb->stage == pme_lb->nstage - 1)
 460         {
 461             /* The performance went up a lot (due to e.g. DD load balancing).
 462              * Add a stage, keep the minima, but rescan all setups.
 463              */
 464             pme_lb->nstage++;
 465
 466             if (debug)
 467             {
 468                 fprintf(debug,"The performance for grid %d %d %d went from %.3f to %.1f M-cycles, this is more than %f\n"
 469                         "Increased the number stages to %d"
 470                         " and ignoring the previous performance\n",
 471                         set->grid[XX],set->grid[YY],set->grid[ZZ],
 472                         cycles*1e-6,set->cycles*1e-6,PME_LB_ACCEL_TOL,
 473                         pme_lb->nstage);
 474             }
 475         }
 476         set->cycles = min(set->cycles,cycles);
 477     }
 478
 479     if (set->cycles < pme_lb->setup[pme_lb->fastest].cycles)
 480     {
 481         pme_lb->fastest = pme_lb->cur;
 482     }
 483     cycles_fast = pme_lb->setup[pme_lb->fastest].cycles;
 484
 485     /* Check in stage 0 if we should stop scanning grids.
 486      * Stop when the time is more than SLOW_FAC longer than the fastest.
 487      */
 488     if (pme_lb->stage == 0 && pme_lb->cur > 0 &&
 489         cycles > pme_lb->setup[pme_lb->fastest].cycles*PME_LB_SLOW_FAC)
 490     {
 491         pme_lb->n = pme_lb->cur + 1;
 492         /* Done with scanning, go to stage 1 */
 493         switch_to_stage1(pme_lb);
 494     }
 495
 496     if (pme_lb->stage == 0)
 497     {
 498         int gridsize_start;
 499
 500         gridsize_start = set->grid[XX]*set->grid[YY]*set->grid[ZZ];
 501
 502         do
 503         {
 504             if (pme_lb->cur+1 < pme_lb->n)
 505             {
 506                 /* We had already generated the next setup */
 507                 OK = TRUE;
 508             }
 509             else
 510             {
 511                 /* Find the next setup */
 512                 OK = pme_loadbal_increase_cutoff(pme_lb,ir->pme_order);
 513             }
 514
 515             if (OK && ir->ePBC != epbcNONE)
 516             {
 517                 OK = (sqr(pme_lb->setup[pme_lb->cur+1].rlistlong)
 518                       <= max_cutoff2(ir->ePBC,state->box));
 519                 if (!OK)
 520                 {
 521                     pme_lb->elimited = epmelblimBOX;
 522                 }
 523             }
 524
 525             if (OK)
 526             {
 527                 pme_lb->cur++;
 528
 529                 if (DOMAINDECOMP(cr))
 530                 {
 531                     OK = change_dd_cutoff(cr,state,ir,
 532                                           pme_lb->setup[pme_lb->cur].rlistlong);
 533                     if (!OK)
 534                     {
 535                         /* Failed: do not use this setup */
 536                         pme_lb->cur--;
 537                         pme_lb->elimited = epmelblimDD;
 538                     }
 539                 }
 540             }
 541             if (!OK)
 542             {
 543                 /* We hit the upper limit for the cut-off,
 544                  * the setup should not go further than cur.
 545                  */
 546                 pme_lb->n = pme_lb->cur + 1;
 547                 print_loadbal_limited(fp_err,fp_log,step,pme_lb);
 548                 /* Switch to the next stage */
 549                 switch_to_stage1(pme_lb);
 550             }
 551         }
 552         while (OK &&
 553                !(pme_lb->setup[pme_lb->cur].grid[XX]*
 554                  pme_lb->setup[pme_lb->cur].grid[YY]*
 555                  pme_lb->setup[pme_lb->cur].grid[ZZ] <
 556                  gridsize_start*PME_LB_GRID_SCALE_FAC
 557                  &&
 558                  pme_lb->setup[pme_lb->cur].grid_efficiency <
 559                  pme_lb->setup[pme_lb->cur-1].grid_efficiency*PME_LB_GRID_EFFICIENCY_REL_FAC));
 560     }
 561
 562     if (pme_lb->stage > 0 && pme_lb->end == 1)
 563     {
 564         pme_lb->cur = 0;
 565         pme_lb->stage = pme_lb->nstage;
 566     }
 567     else if (pme_lb->stage > 0 && pme_lb->end > 1)
 568     {
 569         /* If stage = nstage-1:
 570          *   scan over all setups, rerunning only those setups
 571          *   which are not much slower than the fastest
 572          * else:
 573          *   use the next setup
 574          */
 575         do
 576         {
 577             pme_lb->cur++;
 578             if (pme_lb->cur == pme_lb->end)
 579             {
 580                 pme_lb->stage++;
 581                 pme_lb->cur = pme_lb->start;
 582             }
 583         }
 584         while (pme_lb->stage == pme_lb->nstage - 1 &&
 585                pme_lb->setup[pme_lb->cur].count > 0 &&
 586                pme_lb->setup[pme_lb->cur].cycles > cycles_fast*PME_LB_SLOW_FAC);
 587
 588         if (pme_lb->stage == pme_lb->nstage)
 589         {
 590             /* We are done optimizing, use the fastest setup we found */
 591             pme_lb->cur = pme_lb->fastest;
 592         }
 593     }
 594
 595     if (DOMAINDECOMP(cr) && pme_lb->stage > 0)
 596     {
 597         OK = change_dd_cutoff(cr,state,ir,pme_lb->setup[pme_lb->cur].rlistlong);
 598         if (!OK)
 599         {
 600             /* Failsafe solution */
 601             if (pme_lb->cur > 1 && pme_lb->stage == pme_lb->nstage)
 602             {
 603                 pme_lb->stage--;
 604             }
 605             pme_lb->fastest  = 0;
 606             pme_lb->start    = 0;
 607             pme_lb->end      = pme_lb->cur;
 608             pme_lb->cur      = pme_lb->start;
 609             pme_lb->elimited = epmelblimDD;
 610             print_loadbal_limited(fp_err,fp_log,step,pme_lb);
 611         }
 612     }
 613
 614     /* Change the Coulomb cut-off and the PME grid */
 615
 616     set = &pme_lb->setup[pme_lb->cur];
 617
 618     ic->rcoulomb   = set->rcut_coulomb;
 619     ic->rlist      = set->rlist;
 620     ic->rlistlong  = set->rlistlong;
 621     ir->nstcalclr  = set->nstcalclr;
 622     ic->ewaldcoeff = set->ewaldcoeff;
 623
 624     bUsesSimpleTables = uses_simple_tables(ir->cutoff_scheme, nbv, 0);
 625     if (pme_lb->cutoff_scheme == ecutsVERLET && nbv->grp[0].kernel_type == nbk8x8x8_CUDA)
 626     {
 627         nbnxn_cuda_pme_loadbal_update_param(nbv->cu_nbv,ic);
 628     }
 629     else
 630     {
 631         init_interaction_const_tables(NULL,ic,bUsesSimpleTables,
 632                                       rtab);
 633     }
 634
 635     if (pme_lb->cutoff_scheme == ecutsVERLET && nbv->ngrp > 1)
 636     {
 637         init_interaction_const_tables(NULL,ic,bUsesSimpleTables,
 638                                       rtab);
 639     }
 640
 641     if (cr->duty & DUTY_PME)
 642     {
 643         if (pme_lb->setup[pme_lb->cur].pmedata == NULL)
 644         {
 645             /* Generate a new PME data structure,
 646              * copying part of the old pointers.
 647              */
 648             gmx_pme_reinit(&set->pmedata,
 649                            cr,pme_lb->setup[0].pmedata,ir,
 650                            set->grid);
 651         }
 652         *pmedata = set->pmedata;
 653     }
 654     else
 655     {
 656         /* Tell our PME-only node to switch grid */
 657         gmx_pme_send_switch(cr, set->grid, set->ewaldcoeff);
 658     }
 659
 660     if (debug)
 661     {
 662         print_grid(NULL,debug,"","switched to",set,-1);
 663     }
 664
 665     if (pme_lb->stage == pme_lb->nstage)
 666     {
 667         print_grid(fp_err,fp_log,"","optimal",set,-1);
 668     }
 669
 670     return TRUE;
 671 }
 672
 673 void restart_pme_loadbal(pme_load_balancing_t pme_lb, int n)
 674 {
 675     pme_lb->nstage += n;
 676 }
 677
 678 static int pme_grid_points(const pme_setup_t *setup)
 679 {
 680     return setup->grid[XX]*setup->grid[YY]*setup->grid[ZZ];
 681 }
 682
 683 static void print_pme_loadbal_setting(FILE *fplog,
 684                                      char *name,
 685                                      const pme_setup_t *setup)
 686 {
 687     fprintf(fplog,
 688             "   %-7s %6.3f nm %6.3f nm     %3d %3d %3d   %5.3f nm  %5.3f nm\n",
 689             name,
 690             setup->rcut_coulomb,setup->rlist,
 691             setup->grid[XX],setup->grid[YY],setup->grid[ZZ],
 692             setup->spacing,1/setup->ewaldcoeff);
 693 }
 694
 695 static void print_pme_loadbal_settings(pme_load_balancing_t pme_lb,
 696                                        FILE *fplog)
 697 {
 698     double pp_ratio,grid_ratio;
 699
 700     pp_ratio   = pow(pme_lb->setup[pme_lb->cur].rlist/pme_lb->setup[0].rlistlong,3.0);
 701     grid_ratio = pme_grid_points(&pme_lb->setup[pme_lb->cur])/
 702         (double)pme_grid_points(&pme_lb->setup[0]);
 703
 704     fprintf(fplog,"\n");
 705     fprintf(fplog,"       P P   -   P M E   L O A D   B A L A N C I N G\n");
 706     fprintf(fplog,"\n");
 707     /* Here we only warn when the optimal setting is the last one */
 708     if (pme_lb->elimited != epmelblimNO &&
 709         pme_lb->cur == pme_loadbal_end(pme_lb)-1)
 710     {
 711         fprintf(fplog," NOTE: The PP/PME load balancing was limited by the %s,\n",
 712                 pmelblim_str[pme_lb->elimited]);
 713         fprintf(fplog,"       you might not have reached a good load balance.\n");
 714         if (pme_lb->elimited == epmelblimDD)
 715         {
 716             fprintf(fplog,"       Try different mdrun -dd settings or lower the -dds value.\n");
 717         }
 718         fprintf(fplog,"\n");
 719     }
 720     fprintf(fplog," PP/PME load balancing changed the cut-off and PME settings:\n");
 721     fprintf(fplog,"           particle-particle                    PME\n");
 722     fprintf(fplog,"            rcoulomb  rlist            grid      spacing   1/beta\n");
 723     print_pme_loadbal_setting(fplog,"initial",&pme_lb->setup[0]);
 724     print_pme_loadbal_setting(fplog,"final"  ,&pme_lb->setup[pme_lb->cur]);
 725     fprintf(fplog," cost-ratio           %4.2f             %4.2f\n",
 726             pp_ratio,grid_ratio);
 727     fprintf(fplog," (note that these numbers concern only part of the total PP and PME load)\n");
 728     fprintf(fplog,"\n");
 729 }
 730
 731 void pme_loadbal_done(pme_load_balancing_t pme_lb, FILE *fplog)
 732 {
 733     if (fplog != NULL && (pme_lb->cur > 0 || pme_lb->elimited != epmelblimNO))
 734     {
 735         print_pme_loadbal_settings(pme_lb,fplog);
 736     }
 737
 738     /* TODO: Here we should free all pointers in pme_lb,
 739      * but as it contains pme data structures,
 740      * we need to first make pme.c free all data.
 741      */
 742 }