src/programs/mdrun/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 "macros.h"
  49 #include "pme_loadbal.h"
  50
  51 /* Parameters and setting for one PP-PME setup */
  52 typedef struct {
  53     real rcut_coulomb;    /* Coulomb cut-off                              */
  54     real rlist;           /* pair-list cut-off                            */
  55     real rlistlong;       /* LR pair-list cut-off                         */
  56     int  nstcalclr;       /* frequency of evaluating long-range forces for group scheme */
  57     real spacing;         /* (largest) PME grid spacing                   */
  58     ivec grid;            /* the PME grid dimensions                      */
  59     real grid_efficiency; /* ineffiency factor for non-uniform grids <= 1 */
  60     real ewaldcoeff;      /* the Ewald coefficient                        */
  61     gmx_pme_t pmedata;    /* the data structure used in the PME code      */
  62
  63     int  count;           /* number of times this setup has been timed    */
  64     double cycles;        /* the fastest time for this setup in cycles    */
  65 } pme_setup_t;
  66
  67 /* In the initial scan, step by grids that are at least a factor 0.8 coarser */
  68 #define PME_LB_GRID_SCALE_FAC  0.8
  69 /* In the initial scan, try to skip grids with uneven x/y/z spacing,
  70  * checking if the "efficiency" is more than 5% worse than the previous grid.
  71  */
  72 #define PME_LB_GRID_EFFICIENCY_REL_FAC  1.05
  73 /* Rerun up till 12% slower setups than the fastest up till now */
  74 #define PME_LB_SLOW_FAC  1.12
  75 /* If setups get more than 2% faster, do another round to avoid
  76  * choosing a slower setup due to acceleration or fluctuations.
  77  */
  78 #define PME_LB_ACCEL_TOL 1.02
  79
  80 enum { epmelblimNO, epmelblimBOX, epmelblimDD, epmelblimNR };
  81
  82 const char *pmelblim_str[epmelblimNR] =
  83 { "no", "box size", "domain decompostion" };
  84
  85 struct pme_load_balancing {
  86     int  nstage;        /* the current maximum number of stages */
  87
  88     real cut_spacing;   /* the minimum cutoff / PME grid spacing ratio */
  89     real rcut_vdw;      /* Vdw cutoff (does not change) */
  90     real rcut_coulomb_start; /* Initial electrostatics cutoff */
  91     int  nstcalclr_start; /* Initial electrostatics cutoff */
  92     real rbuf_coulomb;  /* the pairlist buffer size */
  93     real rbuf_vdw;      /* the pairlist buffer size */
  94     matrix box_start;   /* the initial simulation box */
  95     int n;              /* the count of setup as well as the allocation size */
  96     pme_setup_t *setup; /* the PME+cutoff setups */
  97     int cur;            /* the current setup */
  98     int fastest;        /* fastest setup up till now */
  99     int start;          /* start of setup range to consider in stage>0 */
 100     int end;            /* end   of setup range to consider in stage>0 */
 101     int elimited;       /* was the balancing limited, uses enum above */
 102     int cutoff_scheme;  /* Verlet or group cut-offs */
 103
 104     int stage;          /* the current stage */
 105 };
 106
 107 void pme_loadbal_init(pme_load_balancing_t *pme_lb_p,
 108                       const t_inputrec *ir,matrix box,
 109                       const interaction_const_t *ic,
 110                       gmx_pme_t pmedata)
 111 {
 112     pme_load_balancing_t pme_lb;
 113     real spm,sp;
 114     int  d;
 115
 116     snew(pme_lb,1);
 117
 118     /* Any number of stages >= 2 is supported */
 119     pme_lb->nstage   = 2;
 120
 121     pme_lb->cutoff_scheme = ir->cutoff_scheme;
 122
 123     if(pme_lb->cutoff_scheme == ecutsVERLET)
 124     {
 125         pme_lb->rbuf_coulomb = ic->rlist - ic->rcoulomb;
 126         pme_lb->rbuf_vdw     = pme_lb->rbuf_coulomb;
 127     }
 128     else
 129     {
 130         if(ic->rcoulomb > ic->rlist)
 131         {
 132             pme_lb->rbuf_coulomb = ic->rlistlong - ic->rcoulomb;
 133         }
 134         else
 135         {
 136             pme_lb->rbuf_coulomb = ic->rlist - ic->rcoulomb;
 137         }
 138         if(ic->rvdw > ic->rlist)
 139         {
 140             pme_lb->rbuf_vdw = ic->rlistlong - ic->rvdw;
 141         }
 142         else
 143         {
 144             pme_lb->rbuf_vdw = ic->rlist - ic->rvdw;
 145         }
 146     }
 147
 148     copy_mat(box,pme_lb->box_start);
 149     if (ir->ePBC==epbcXY && ir->nwall==2)
 150     {
 151         svmul(ir->wall_ewald_zfac,pme_lb->box_start[ZZ],pme_lb->box_start[ZZ]);
 152     }
 153
 154     pme_lb->n = 1;
 155     snew(pme_lb->setup,pme_lb->n);
 156
 157     pme_lb->rcut_vdw              = ic->rvdw;
 158     pme_lb->rcut_coulomb_start    = ir->rcoulomb;
 159     pme_lb->nstcalclr_start       = ir->nstcalclr;
 160
 161     pme_lb->cur = 0;
 162     pme_lb->setup[0].rcut_coulomb = ic->rcoulomb;
 163     pme_lb->setup[0].rlist        = ic->rlist;
 164     pme_lb->setup[0].rlistlong    = ic->rlistlong;
 165     pme_lb->setup[0].nstcalclr    = ir->nstcalclr;
 166     pme_lb->setup[0].grid[XX]     = ir->nkx;
 167     pme_lb->setup[0].grid[YY]     = ir->nky;
 168     pme_lb->setup[0].grid[ZZ]     = ir->nkz;
 169     pme_lb->setup[0].ewaldcoeff   = ic->ewaldcoeff;
 170
 171     pme_lb->setup[0].pmedata  = pmedata;
 172
 173     spm = 0;
 174     for(d=0; d<DIM; d++)
 175     {
 176         sp = norm(pme_lb->box_start[d])/pme_lb->setup[0].grid[d];
 177         if (sp > spm)
 178         {
 179             spm = sp;
 180         }
 181     }
 182     pme_lb->setup[0].spacing = spm;
 183
 184     if (ir->fourier_spacing > 0)
 185     {
 186         pme_lb->cut_spacing = ir->rcoulomb/ir->fourier_spacing;
 187     }
 188     else
 189     {
 190         pme_lb->cut_spacing = ir->rcoulomb/pme_lb->setup[0].spacing;
 191     }
 192
 193     pme_lb->stage = 0;
 194
 195     pme_lb->fastest  = 0;
 196     pme_lb->start    = 0;
 197     pme_lb->end      = 0;
 198     pme_lb->elimited = epmelblimNO;
 199
 200     *pme_lb_p = pme_lb;
 201 }
 202
 203 static gmx_bool pme_loadbal_increase_cutoff(pme_load_balancing_t pme_lb,
 204                                             int pme_order)
 205 {
 206     pme_setup_t *set;
 207     real fac,sp;
 208     real tmpr_coulomb,tmpr_vdw;
 209     int d;
 210
 211     /* Try to add a new setup with next larger cut-off to the list */
 212     pme_lb->n++;
 213     srenew(pme_lb->setup,pme_lb->n);
 214     set = &pme_lb->setup[pme_lb->n-1];
 215     set->pmedata = NULL;
 216
 217     fac = 1;
 218     do
 219     {
 220         fac *= 1.01;
 221         clear_ivec(set->grid);
 222         sp = calc_grid(NULL,pme_lb->box_start,
 223                        fac*pme_lb->setup[pme_lb->cur].spacing,
 224                        &set->grid[XX],
 225                        &set->grid[YY],
 226                        &set->grid[ZZ]);
 227
 228         /* In parallel we can't have grids smaller than 2*pme_order,
 229          * and we would anyhow not gain much speed at these grid sizes.
 230          */
 231         for(d=0; d<DIM; d++)
 232         {
 233             if (set->grid[d] <= 2*pme_order)
 234             {
 235                 pme_lb->n--;
 236
 237                 return FALSE;
 238             }
 239         }
 240     }
 241     while (sp <= 1.001*pme_lb->setup[pme_lb->cur].spacing);
 242
 243     set->rcut_coulomb = pme_lb->cut_spacing*sp;
 244
 245     if(pme_lb->cutoff_scheme == ecutsVERLET)
 246     {
 247         set->rlist        = set->rcut_coulomb + pme_lb->rbuf_coulomb;
 248         /* We dont use LR lists with Verlet, but this avoids if-statements in further checks */
 249         set->rlistlong    = set->rlist;
 250     }
 251     else
 252     {
 253         tmpr_coulomb          = set->rcut_coulomb + pme_lb->rbuf_coulomb;
 254         tmpr_vdw              = pme_lb->rcut_vdw + pme_lb->rbuf_vdw;
 255         set->rlist            = min(tmpr_coulomb,tmpr_vdw);
 256         set->rlistlong        = max(tmpr_coulomb,tmpr_vdw);
 257
 258         /* Set the long-range update frequency */
 259         if(set->rlist == set->rlistlong)
 260         {
 261             /* No long-range interactions if the short-/long-range cutoffs are identical */
 262             set->nstcalclr = 0;
 263         }
 264         else if(pme_lb->nstcalclr_start==0 || pme_lb->nstcalclr_start==1)
 265         {
 266             /* We were not doing long-range before, but now we are since rlist!=rlistlong */
 267             set->nstcalclr = 1;
 268         }
 269         else
 270         {
 271             /* We were already doing long-range interactions from the start */
 272             if(pme_lb->rcut_vdw > pme_lb->rcut_coulomb_start)
 273             {
 274                 /* We were originally doing long-range VdW-only interactions.
 275                  * If rvdw is still longer than rcoulomb we keep the original nstcalclr,
 276                  * but if the coulomb cutoff has become longer we should update the long-range
 277                  * part every step.
 278                  */
 279                 set->nstcalclr = (tmpr_vdw > tmpr_coulomb) ? pme_lb->nstcalclr_start : 1;
 280             }
 281             else
 282             {
 283                 /* We were not doing any long-range interaction from the start,
 284                  * since it is not possible to do twin-range coulomb for the PME interaction.
 285                  */
 286                 set->nstcalclr = 1;
 287             }
 288         }
 289     }
 290
 291     set->spacing      = sp;
 292     /* The grid efficiency is the size wrt a grid with uniform x/y/z spacing */
 293     set->grid_efficiency = 1;
 294     for(d=0; d<DIM; d++)
 295     {
 296         set->grid_efficiency *= (set->grid[d]*sp)/norm(pme_lb->box_start[d]);
 297     }
 298     /* The Ewald coefficient is inversly proportional to the cut-off */
 299     set->ewaldcoeff =
 300         pme_lb->setup[0].ewaldcoeff*pme_lb->setup[0].rcut_coulomb/set->rcut_coulomb;
 301
 302     set->count   = 0;
 303     set->cycles  = 0;
 304
 305     if (debug)
 306     {
 307         fprintf(debug,"PME loadbal: grid %d %d %d, coulomb cutoff %f\n",
 308                 set->grid[XX],set->grid[YY],set->grid[ZZ],set->rcut_coulomb);
 309     }
 310     return TRUE;
 311 }
 312
 313 static void print_grid(FILE *fp_err,FILE *fp_log,
 314                        const char *pre,
 315                        const char *desc,
 316                        const pme_setup_t *set,
 317                        double cycles)
 318 {
 319     char buf[STRLEN],buft[STRLEN];
 320
 321     if (cycles >= 0)
 322     {
 323         sprintf(buft,": %.1f M-cycles",cycles*1e-6);
 324     }
 325     else
 326     {
 327         buft[0] = '\0';
 328     }
 329     sprintf(buf,"%-11s%10s pme grid %d %d %d, coulomb cutoff %.3f%s",
 330             pre,
 331             desc,set->grid[XX],set->grid[YY],set->grid[ZZ],set->rcut_coulomb,
 332             buft);
 333     if (fp_err != NULL)
 334     {
 335         fprintf(fp_err,"\r%s\n",buf);
 336     }
 337     if (fp_log != NULL)
 338     {
 339         fprintf(fp_log,"%s\n",buf);
 340     }
 341 }
 342
 343 static int pme_loadbal_end(pme_load_balancing_t pme_lb)
 344 {
 345     /* In the initial stage only n is set; end is not set yet */
 346     if (pme_lb->end > 0)
 347     {
 348         return pme_lb->end;
 349     }
 350     else
 351     {
 352         return pme_lb->n;
 353     }
 354 }
 355
 356 static void print_loadbal_limited(FILE *fp_err,FILE *fp_log,
 357                                   gmx_large_int_t step,
 358                                   pme_load_balancing_t pme_lb)
 359 {
 360     char buf[STRLEN],sbuf[22];
 361
 362     sprintf(buf,"step %4s: the %s limited the PME load balancing to a coulomb cut-off of %.3f",
 363             gmx_step_str(step,sbuf),
 364             pmelblim_str[pme_lb->elimited],
 365             pme_lb->setup[pme_loadbal_end(pme_lb)-1].rcut_coulomb);
 366     if (fp_err != NULL)
 367     {
 368         fprintf(fp_err,"\r%s\n",buf);
 369     }
 370     if (fp_log != NULL)
 371     {
 372         fprintf(fp_log,"%s\n",buf);
 373     }
 374 }
 375
 376 static void switch_to_stage1(pme_load_balancing_t pme_lb)
 377 {
 378     pme_lb->start = 0;
 379     while (pme_lb->start+1 < pme_lb->n &&
 380            (pme_lb->setup[pme_lb->start].count == 0 ||
 381             pme_lb->setup[pme_lb->start].cycles >
 382             pme_lb->setup[pme_lb->fastest].cycles*PME_LB_SLOW_FAC))
 383     {
 384         pme_lb->start++;
 385     }
 386     while (pme_lb->start > 0 && pme_lb->setup[pme_lb->start-1].cycles == 0)
 387     {
 388         pme_lb->start--;
 389     }
 390
 391     pme_lb->end = pme_lb->n;
 392     if (pme_lb->setup[pme_lb->end-1].count > 0 &&
 393         pme_lb->setup[pme_lb->end-1].cycles >
 394         pme_lb->setup[pme_lb->fastest].cycles*PME_LB_SLOW_FAC)
 395     {
 396         pme_lb->end--;
 397     }
 398
 399     pme_lb->stage = 1;
 400
 401     /* Next we want to choose setup pme_lb->start, but as we will increase
 402      * pme_ln->cur by one right after returning, we subtract 1 here.
 403      */
 404     pme_lb->cur = pme_lb->start - 1;
 405 }
 406
 407 gmx_bool pme_load_balance(pme_load_balancing_t pme_lb,
 408                           t_commrec *cr,
 409                           FILE *fp_err,
 410                           FILE *fp_log,
 411                           t_inputrec *ir,
 412                           t_state *state,
 413                           double cycles,
 414                           interaction_const_t *ic,
 415                           nonbonded_verlet_t *nbv,
 416                           gmx_pme_t *pmedata,
 417                           gmx_large_int_t step)
 418 {
 419     gmx_bool OK;
 420     pme_setup_t *set;
 421     double cycles_fast;
 422     char buf[STRLEN],sbuf[22];
 423     real rtab;
 424     gmx_bool bUsesSimpleTables = TRUE;
 425
 426     if (pme_lb->stage == pme_lb->nstage)
 427     {
 428         return FALSE;
 429     }
 430
 431     if (PAR(cr))
 432     {
 433         gmx_sumd(1,&cycles,cr);
 434         cycles /= cr->nnodes;
 435     }
 436
 437     set = &pme_lb->setup[pme_lb->cur];
 438     set->count++;
 439
 440     rtab = ir->rlistlong + ir->tabext;
 441
 442     if (set->count % 2 == 1)
 443     {
 444         /* Skip the first cycle, because the first step after a switch
 445          * is much slower due to allocation and/or caching effects.
 446          */
 447         return TRUE;
 448     }
 449
 450     sprintf(buf, "step %4s: ", gmx_step_str(step,sbuf));
 451     print_grid(fp_err,fp_log,buf,"timed with",set,cycles);
 452
 453     if (set->count <= 2)
 454     {
 455         set->cycles = cycles;
 456     }
 457     else
 458     {
 459         if (cycles*PME_LB_ACCEL_TOL < set->cycles &&
 460             pme_lb->stage == pme_lb->nstage - 1)
 461         {
 462             /* The performance went up a lot (due to e.g. DD load balancing).
 463              * Add a stage, keep the minima, but rescan all setups.
 464              */
 465             pme_lb->nstage++;
 466
 467             if (debug)
 468             {
 469                 fprintf(debug,"The performance for grid %d %d %d went from %.3f to %.1f M-cycles, this is more than %f\n"
 470                         "Increased the number stages to %d"
 471                         " and ignoring the previous performance\n",
 472                         set->grid[XX],set->grid[YY],set->grid[ZZ],
 473                         cycles*1e-6,set->cycles*1e-6,PME_LB_ACCEL_TOL,
 474                         pme_lb->nstage);
 475             }
 476         }
 477         set->cycles = min(set->cycles,cycles);
 478     }
 479
 480     if (set->cycles < pme_lb->setup[pme_lb->fastest].cycles)
 481     {
 482         pme_lb->fastest = pme_lb->cur;
 483     }
 484     cycles_fast = pme_lb->setup[pme_lb->fastest].cycles;
 485
 486     /* Check in stage 0 if we should stop scanning grids.
 487      * Stop when the time is more than SLOW_FAC longer than the fastest.
 488      */
 489     if (pme_lb->stage == 0 && pme_lb->cur > 0 &&
 490         cycles > pme_lb->setup[pme_lb->fastest].cycles*PME_LB_SLOW_FAC)
 491     {
 492         pme_lb->n = pme_lb->cur + 1;
 493         /* Done with scanning, go to stage 1 */
 494         switch_to_stage1(pme_lb);
 495     }
 496
 497     if (pme_lb->stage == 0)
 498     {
 499         int gridsize_start;
 500
 501         gridsize_start = set->grid[XX]*set->grid[YY]*set->grid[ZZ];
 502
 503         do
 504         {
 505             if (pme_lb->cur+1 < pme_lb->n)
 506             {
 507                 /* We had already generated the next setup */
 508                 OK = TRUE;
 509             }
 510             else
 511             {
 512                 /* Find the next setup */
 513                 OK = pme_loadbal_increase_cutoff(pme_lb,ir->pme_order);
 514             }
 515
 516             if (OK && ir->ePBC != epbcNONE)
 517             {
 518                 OK = (sqr(pme_lb->setup[pme_lb->cur+1].rlistlong)
 519                       <= max_cutoff2(ir->ePBC,state->box));
 520                 if (!OK)
 521                 {
 522                     pme_lb->elimited = epmelblimBOX;
 523                 }
 524             }
 525
 526             if (OK)
 527             {
 528                 pme_lb->cur++;
 529
 530                 if (DOMAINDECOMP(cr))
 531                 {
 532                     OK = change_dd_cutoff(cr,state,ir,
 533                                           pme_lb->setup[pme_lb->cur].rlistlong);
 534                     if (!OK)
 535                     {
 536                         /* Failed: do not use this setup */
 537                         pme_lb->cur--;
 538                         pme_lb->elimited = epmelblimDD;
 539                     }
 540                 }
 541             }
 542             if (!OK)
 543             {
 544                 /* We hit the upper limit for the cut-off,
 545                  * the setup should not go further than cur.
 546                  */
 547                 pme_lb->n = pme_lb->cur + 1;
 548                 print_loadbal_limited(fp_err,fp_log,step,pme_lb);
 549                 /* Switch to the next stage */
 550                 switch_to_stage1(pme_lb);
 551             }
 552         }
 553         while (OK &&
 554                !(pme_lb->setup[pme_lb->cur].grid[XX]*
 555                  pme_lb->setup[pme_lb->cur].grid[YY]*
 556                  pme_lb->setup[pme_lb->cur].grid[ZZ] <
 557                  gridsize_start*PME_LB_GRID_SCALE_FAC
 558                  &&
 559                  pme_lb->setup[pme_lb->cur].grid_efficiency <
 560                  pme_lb->setup[pme_lb->cur-1].grid_efficiency*PME_LB_GRID_EFFICIENCY_REL_FAC));
 561     }
 562
 563     if (pme_lb->stage > 0 && pme_lb->end == 1)
 564     {
 565         pme_lb->cur = 0;
 566         pme_lb->stage = pme_lb->nstage;
 567     }
 568     else if (pme_lb->stage > 0 && pme_lb->end > 1)
 569     {
 570         /* If stage = nstage-1:
 571          *   scan over all setups, rerunning only those setups
 572          *   which are not much slower than the fastest
 573          * else:
 574          *   use the next setup
 575          */
 576         do
 577         {
 578             pme_lb->cur++;
 579             if (pme_lb->cur == pme_lb->end)
 580             {
 581                 pme_lb->stage++;
 582                 pme_lb->cur = pme_lb->start;
 583             }
 584         }
 585         while (pme_lb->stage == pme_lb->nstage - 1 &&
 586                pme_lb->setup[pme_lb->cur].count > 0 &&
 587                pme_lb->setup[pme_lb->cur].cycles > cycles_fast*PME_LB_SLOW_FAC);
 588
 589         if (pme_lb->stage == pme_lb->nstage)
 590         {
 591             /* We are done optimizing, use the fastest setup we found */
 592             pme_lb->cur = pme_lb->fastest;
 593         }
 594     }
 595
 596     if (DOMAINDECOMP(cr) && pme_lb->stage > 0)
 597     {
 598         OK = change_dd_cutoff(cr,state,ir,pme_lb->setup[pme_lb->cur].rlistlong);
 599         if (!OK)
 600         {
 601             /* Failsafe solution */
 602             if (pme_lb->cur > 1 && pme_lb->stage == pme_lb->nstage)
 603             {
 604                 pme_lb->stage--;
 605             }
 606             pme_lb->fastest  = 0;
 607             pme_lb->start    = 0;
 608             pme_lb->end      = pme_lb->cur;
 609             pme_lb->cur      = pme_lb->start;
 610             pme_lb->elimited = epmelblimDD;
 611             print_loadbal_limited(fp_err,fp_log,step,pme_lb);
 612         }
 613     }
 614
 615     /* Change the Coulomb cut-off and the PME grid */
 616
 617     set = &pme_lb->setup[pme_lb->cur];
 618
 619     ic->rcoulomb   = set->rcut_coulomb;
 620     ic->rlist      = set->rlist;
 621     ic->rlistlong  = set->rlistlong;
 622     ir->nstcalclr  = set->nstcalclr;
 623     ic->ewaldcoeff = set->ewaldcoeff;
 624
 625     bUsesSimpleTables = uses_simple_tables(ir->cutoff_scheme, nbv, 0);
 626     if (pme_lb->cutoff_scheme == ecutsVERLET &&
 627         nbv->grp[0].kernel_type == nbnxnk8x8x8_CUDA)
 628     {
 629         nbnxn_cuda_pme_loadbal_update_param(nbv->cu_nbv,ic);
 630     }
 631     else
 632     {
 633         init_interaction_const_tables(NULL,ic,bUsesSimpleTables,
 634                                       rtab);
 635     }
 636
 637     if (pme_lb->cutoff_scheme == ecutsVERLET && nbv->ngrp > 1)
 638     {
 639         init_interaction_const_tables(NULL,ic,bUsesSimpleTables,
 640                                       rtab);
 641     }
 642
 643     if (cr->duty & DUTY_PME)
 644     {
 645         if (pme_lb->setup[pme_lb->cur].pmedata == NULL)
 646         {
 647             /* Generate a new PME data structure,
 648              * copying part of the old pointers.
 649              */
 650             gmx_pme_reinit(&set->pmedata,
 651                            cr,pme_lb->setup[0].pmedata,ir,
 652                            set->grid);
 653         }
 654         *pmedata = set->pmedata;
 655     }
 656     else
 657     {
 658         /* Tell our PME-only node to switch grid */
 659         gmx_pme_send_switch(cr, set->grid, set->ewaldcoeff);
 660     }
 661
 662     if (debug)
 663     {
 664         print_grid(NULL,debug,"","switched to",set,-1);
 665     }
 666
 667     if (pme_lb->stage == pme_lb->nstage)
 668     {
 669         print_grid(fp_err,fp_log,"","optimal",set,-1);
 670     }
 671
 672     return TRUE;
 673 }
 674
 675 void restart_pme_loadbal(pme_load_balancing_t pme_lb, int n)
 676 {
 677     pme_lb->nstage += n;
 678 }
 679
 680 static int pme_grid_points(const pme_setup_t *setup)
 681 {
 682     return setup->grid[XX]*setup->grid[YY]*setup->grid[ZZ];
 683 }
 684
 685 static void print_pme_loadbal_setting(FILE *fplog,
 686                                      char *name,
 687                                      const pme_setup_t *setup)
 688 {
 689     fprintf(fplog,
 690             "   %-7s %6.3f nm %6.3f nm     %3d %3d %3d   %5.3f nm  %5.3f nm\n",
 691             name,
 692             setup->rcut_coulomb,setup->rlist,
 693             setup->grid[XX],setup->grid[YY],setup->grid[ZZ],
 694             setup->spacing,1/setup->ewaldcoeff);
 695 }
 696
 697 static void print_pme_loadbal_settings(pme_load_balancing_t pme_lb,
 698                                        FILE *fplog)
 699 {
 700     double pp_ratio,grid_ratio;
 701
 702     pp_ratio   = pow(pme_lb->setup[pme_lb->cur].rlist/pme_lb->setup[0].rlistlong,3.0);
 703     grid_ratio = pme_grid_points(&pme_lb->setup[pme_lb->cur])/
 704         (double)pme_grid_points(&pme_lb->setup[0]);
 705
 706     fprintf(fplog,"\n");
 707     fprintf(fplog,"       P P   -   P M E   L O A D   B A L A N C I N G\n");
 708     fprintf(fplog,"\n");
 709     /* Here we only warn when the optimal setting is the last one */
 710     if (pme_lb->elimited != epmelblimNO &&
 711         pme_lb->cur == pme_loadbal_end(pme_lb)-1)
 712     {
 713         fprintf(fplog," NOTE: The PP/PME load balancing was limited by the %s,\n",
 714                 pmelblim_str[pme_lb->elimited]);
 715         fprintf(fplog,"       you might not have reached a good load balance.\n");
 716         if (pme_lb->elimited == epmelblimDD)
 717         {
 718             fprintf(fplog,"       Try different mdrun -dd settings or lower the -dds value.\n");
 719         }
 720         fprintf(fplog,"\n");
 721     }
 722     fprintf(fplog," PP/PME load balancing changed the cut-off and PME settings:\n");
 723     fprintf(fplog,"           particle-particle                    PME\n");
 724     fprintf(fplog,"            rcoulomb  rlist            grid      spacing   1/beta\n");
 725     print_pme_loadbal_setting(fplog,"initial",&pme_lb->setup[0]);
 726     print_pme_loadbal_setting(fplog,"final"  ,&pme_lb->setup[pme_lb->cur]);
 727     fprintf(fplog," cost-ratio           %4.2f             %4.2f\n",
 728             pp_ratio,grid_ratio);
 729     fprintf(fplog," (note that these numbers concern only part of the total PP and PME load)\n");
 730     fprintf(fplog,"\n");
 731 }
 732
 733 void pme_loadbal_done(pme_load_balancing_t pme_lb, FILE *fplog)
 734 {
 735     if (fplog != NULL && (pme_lb->cur > 0 || pme_lb->elimited != epmelblimNO))
 736     {
 737         print_pme_loadbal_settings(pme_lb,fplog);
 738     }
 739
 740     /* TODO: Here we should free all pointers in pme_lb,
 741      * but as it contains pme data structures,
 742      * we need to first make pme.c free all data.
 743      */
 744 }