src/kernel/pme_loadbal.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
   5  * Copyright (c) 2001-2011, The GROMACS development team,
   6  * check out http://www.gromacs.org for more information.
   7  * Copyright (c) 2012, by the GROMACS development team, led by
   8  * David van der Spoel, Berk Hess, Erik Lindahl, and including many
   9  * others, as listed in the AUTHORS file in the top-level source
  10  * directory and at http://www.gromacs.org.
  11  *
  12  * GROMACS is free software; you can redistribute it and/or
  13  * modify it under the terms of the GNU Lesser General Public License
  14  * as published by the Free Software Foundation; either version 2.1
  15  * of the License, or (at your option) any later version.
  16  *
  17  * GROMACS is distributed in the hope that it will be useful,
  18  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  19  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  20  * Lesser General Public License for more details.
  21  *
  22  * You should have received a copy of the GNU Lesser General Public
  23  * License along with GROMACS; if not, see
  24  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  25  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  26  *
  27  * If you want to redistribute modifications to GROMACS, please
  28  * consider that scientific software is very special. Version
  29  * control is crucial - bugs must be traceable. We will be happy to
  30  * consider code for inclusion in the official distribution, but
  31  * derived work must not be called official GROMACS. Details are found
  32  * in the README & COPYING files - if they are missing, get the
  33  * official version at http://www.gromacs.org.
  34  *
  35  * To help us fund GROMACS development, we humbly ask that you cite
  36  * the research papers on the package. Check out http://www.gromacs.org.
  37  */
  38 #ifdef HAVE_CONFIG_H
  39 #include <config.h>
  40 #endif
  41
  42 #include "smalloc.h"
  43 #include "network.h"
  44 #include "calcgrid.h"
  45 #include "pme.h"
  46 #include "vec.h"
  47 #include "domdec.h"
  48 #include "nbnxn_cuda_data_mgmt.h"
  49 #include "force.h"
  50 #include "pme_loadbal.h"
  51
  52 /* Parameters and setting for one PP-PME setup */
  53 typedef struct {
  54     real rcut_coulomb;    /* Coulomb cut-off                              */
  55     real rlist;           /* pair-list cut-off                            */
  56     real rlistlong;       /* LR pair-list cut-off                         */
  57     int  nstcalclr;       /* frequency of evaluating long-range forces for group scheme */
  58     real spacing;         /* (largest) PME grid spacing                   */
  59     ivec grid;            /* the PME grid dimensions                      */
  60     real grid_efficiency; /* ineffiency factor for non-uniform grids <= 1 */
  61     real ewaldcoeff;      /* the Ewald coefficient                        */
  62     gmx_pme_t pmedata;    /* the data structure used in the PME code      */
  63
  64     int  count;           /* number of times this setup has been timed    */
  65     double cycles;        /* the fastest time for this setup in cycles    */
  66 } pme_setup_t;
  67
  68 /* In the initial scan, step by grids that are at least a factor 0.8 coarser */
  69 #define PME_LB_GRID_SCALE_FAC  0.8
  70 /* In the initial scan, try to skip grids with uneven x/y/z spacing,
  71  * checking if the "efficiency" is more than 5% worse than the previous grid.
  72  */
  73 #define PME_LB_GRID_EFFICIENCY_REL_FAC  1.05
  74 /* Rerun up till 12% slower setups than the fastest up till now */
  75 #define PME_LB_SLOW_FAC  1.12
  76 /* If setups get more than 2% faster, do another round to avoid
  77  * choosing a slower setup due to acceleration or fluctuations.
  78  */
  79 #define PME_LB_ACCEL_TOL 1.02
  80
  81 enum { epmelblimNO, epmelblimBOX, epmelblimDD, epmelblimNR };
  82
  83 const char *pmelblim_str[epmelblimNR] =
  84 { "no", "box size", "domain decompostion" };
  85
  86 struct pme_load_balancing {
  87     int  nstage;        /* the current maximum number of stages */
  88
  89     real cut_spacing;   /* the minimum cutoff / PME grid spacing ratio */
  90     real rcut_vdw;      /* Vdw cutoff (does not change) */
  91     real rcut_coulomb_start; /* Initial electrostatics cutoff */
  92     int  nstcalclr_start; /* Initial electrostatics cutoff */
  93     real rbuf_coulomb;  /* the pairlist buffer size */
  94     real rbuf_vdw;      /* the pairlist buffer size */
  95     matrix box_start;   /* the initial simulation box */
  96     int n;              /* the count of setup as well as the allocation size */
  97     pme_setup_t *setup; /* the PME+cutoff setups */
  98     int cur;            /* the current setup */
  99     int fastest;        /* fastest setup up till now */
 100     int start;          /* start of setup range to consider in stage>0 */
 101     int end;            /* end   of setup range to consider in stage>0 */
 102     int elimited;       /* was the balancing limited, uses enum above */
 103     int cutoff_scheme;  /* Verlet or group cut-offs */
 104
 105     int stage;          /* the current stage */
 106 };
 107
 108 void pme_loadbal_init(pme_load_balancing_t *pme_lb_p,
 109                       const t_inputrec *ir,matrix box,
 110                       const interaction_const_t *ic,
 111                       gmx_pme_t pmedata)
 112 {
 113     pme_load_balancing_t pme_lb;
 114     real spm,sp;
 115     int  d;
 116
 117     snew(pme_lb,1);
 118
 119     /* Any number of stages >= 2 is supported */
 120     pme_lb->nstage   = 2;
 121
 122     pme_lb->cutoff_scheme = ir->cutoff_scheme;
 123
 124     if(pme_lb->cutoff_scheme == ecutsVERLET)
 125     {
 126         pme_lb->rbuf_coulomb = ic->rlist - ic->rcoulomb;
 127         pme_lb->rbuf_vdw     = pme_lb->rbuf_coulomb;
 128     }
 129     else
 130     {
 131         if(ic->rcoulomb > ic->rlist)
 132         {
 133             pme_lb->rbuf_coulomb = ic->rlistlong - ic->rcoulomb;
 134         }
 135         else
 136         {
 137             pme_lb->rbuf_coulomb = ic->rlist - ic->rcoulomb;
 138         }
 139         if(ic->rvdw > ic->rlist)
 140         {
 141             pme_lb->rbuf_vdw = ic->rlistlong - ic->rvdw;
 142         }
 143         else
 144         {
 145             pme_lb->rbuf_vdw = ic->rlist - ic->rvdw;
 146         }
 147     }
 148
 149     copy_mat(box,pme_lb->box_start);
 150     if (ir->ePBC==epbcXY && ir->nwall==2)
 151     {
 152         svmul(ir->wall_ewald_zfac,pme_lb->box_start[ZZ],pme_lb->box_start[ZZ]);
 153     }
 154
 155     pme_lb->n = 1;
 156     snew(pme_lb->setup,pme_lb->n);
 157
 158     pme_lb->rcut_vdw              = ic->rvdw;
 159     pme_lb->rcut_coulomb_start    = ir->rcoulomb;
 160     pme_lb->nstcalclr_start       = ir->nstcalclr;
 161
 162     pme_lb->cur = 0;
 163     pme_lb->setup[0].rcut_coulomb = ic->rcoulomb;
 164     pme_lb->setup[0].rlist        = ic->rlist;
 165     pme_lb->setup[0].rlistlong    = ic->rlistlong;
 166     pme_lb->setup[0].nstcalclr    = ir->nstcalclr;
 167     pme_lb->setup[0].grid[XX]     = ir->nkx;
 168     pme_lb->setup[0].grid[YY]     = ir->nky;
 169     pme_lb->setup[0].grid[ZZ]     = ir->nkz;
 170     pme_lb->setup[0].ewaldcoeff   = ic->ewaldcoeff;
 171
 172     pme_lb->setup[0].pmedata  = pmedata;
 173
 174     spm = 0;
 175     for(d=0; d<DIM; d++)
 176     {
 177         sp = norm(pme_lb->box_start[d])/pme_lb->setup[0].grid[d];
 178         if (sp > spm)
 179         {
 180             spm = sp;
 181         }
 182     }
 183     pme_lb->setup[0].spacing = spm;
 184
 185     if (ir->fourier_spacing > 0)
 186     {
 187         pme_lb->cut_spacing = ir->rcoulomb/ir->fourier_spacing;
 188     }
 189     else
 190     {
 191         pme_lb->cut_spacing = ir->rcoulomb/pme_lb->setup[0].spacing;
 192     }
 193
 194     pme_lb->stage = 0;
 195
 196     pme_lb->fastest  = 0;
 197     pme_lb->start    = 0;
 198     pme_lb->end      = 0;
 199     pme_lb->elimited = epmelblimNO;
 200
 201     *pme_lb_p = pme_lb;
 202 }
 203
 204 static gmx_bool pme_loadbal_increase_cutoff(pme_load_balancing_t pme_lb,
 205                                             int pme_order)
 206 {
 207     pme_setup_t *set;
 208     real fac,sp;
 209     real tmpr_coulomb,tmpr_vdw;
 210     int d;
 211
 212     /* Try to add a new setup with next larger cut-off to the list */
 213     pme_lb->n++;
 214     srenew(pme_lb->setup,pme_lb->n);
 215     set = &pme_lb->setup[pme_lb->n-1];
 216     set->pmedata = NULL;
 217
 218     fac = 1;
 219     do
 220     {
 221         fac *= 1.01;
 222         clear_ivec(set->grid);
 223         sp = calc_grid(NULL,pme_lb->box_start,
 224                        fac*pme_lb->setup[pme_lb->cur].spacing,
 225                        &set->grid[XX],
 226                        &set->grid[YY],
 227                        &set->grid[ZZ]);
 228
 229         /* In parallel we can't have grids smaller than 2*pme_order,
 230          * and we would anyhow not gain much speed at these grid sizes.
 231          */
 232         for(d=0; d<DIM; d++)
 233         {
 234             if (set->grid[d] <= 2*pme_order)
 235             {
 236                 pme_lb->n--;
 237
 238                 return FALSE;
 239             }
 240         }
 241     }
 242     while (sp <= 1.001*pme_lb->setup[pme_lb->cur].spacing);
 243
 244     set->rcut_coulomb = pme_lb->cut_spacing*sp;
 245
 246     if(pme_lb->cutoff_scheme == ecutsVERLET)
 247     {
 248         set->rlist        = set->rcut_coulomb + pme_lb->rbuf_coulomb;
 249         /* We dont use LR lists with Verlet, but this avoids if-statements in further checks */
 250         set->rlistlong    = set->rlist;
 251     }
 252     else
 253     {
 254         tmpr_coulomb          = set->rcut_coulomb + pme_lb->rbuf_coulomb;
 255         tmpr_vdw              = pme_lb->rcut_vdw + pme_lb->rbuf_vdw;
 256         set->rlist            = min(tmpr_coulomb,tmpr_vdw);
 257         set->rlistlong        = max(tmpr_coulomb,tmpr_vdw);
 258
 259         /* Set the long-range update frequency */
 260         if(set->rlist == set->rlistlong)
 261         {
 262             /* No long-range interactions if the short-/long-range cutoffs are identical */
 263             set->nstcalclr = 0;
 264         }
 265         else if(pme_lb->nstcalclr_start==0 || pme_lb->nstcalclr_start==1)
 266         {
 267             /* We were not doing long-range before, but now we are since rlist!=rlistlong */
 268             set->nstcalclr = 1;
 269         }
 270         else
 271         {
 272             /* We were already doing long-range interactions from the start */
 273             if(pme_lb->rcut_vdw > pme_lb->rcut_coulomb_start)
 274             {
 275                 /* We were originally doing long-range VdW-only interactions.
 276                  * If rvdw is still longer than rcoulomb we keep the original nstcalclr,
 277                  * but if the coulomb cutoff has become longer we should update the long-range
 278                  * part every step.
 279                  */
 280                 set->nstcalclr = (tmpr_vdw > tmpr_coulomb) ? pme_lb->nstcalclr_start : 1;
 281             }
 282             else
 283             {
 284                 /* We were not doing any long-range interaction from the start,
 285                  * since it is not possible to do twin-range coulomb for the PME interaction.
 286                  */
 287                 set->nstcalclr = 1;
 288             }
 289         }
 290     }
 291
 292     set->spacing      = sp;
 293     /* The grid efficiency is the size wrt a grid with uniform x/y/z spacing */
 294     set->grid_efficiency = 1;
 295     for(d=0; d<DIM; d++)
 296     {
 297         set->grid_efficiency *= (set->grid[d]*sp)/norm(pme_lb->box_start[d]);
 298     }
 299     /* The Ewald coefficient is inversly proportional to the cut-off */
 300     set->ewaldcoeff =
 301         pme_lb->setup[0].ewaldcoeff*pme_lb->setup[0].rcut_coulomb/set->rcut_coulomb;
 302
 303     set->count   = 0;
 304     set->cycles  = 0;
 305
 306     if (debug)
 307     {
 308         fprintf(debug,"PME loadbal: grid %d %d %d, coulomb cutoff %f\n",
 309                 set->grid[XX],set->grid[YY],set->grid[ZZ],set->rcut_coulomb);
 310     }
 311     return TRUE;
 312 }
 313
 314 static void print_grid(FILE *fp_err,FILE *fp_log,
 315                        const char *pre,
 316                        const char *desc,
 317                        const pme_setup_t *set,
 318                        double cycles)
 319 {
 320     char buf[STRLEN],buft[STRLEN];
 321
 322     if (cycles >= 0)
 323     {
 324         sprintf(buft,": %.1f M-cycles",cycles*1e-6);
 325     }
 326     else
 327     {
 328         buft[0] = '\0';
 329     }
 330     sprintf(buf,"%-11s%10s pme grid %d %d %d, coulomb cutoff %.3f%s",
 331             pre,
 332             desc,set->grid[XX],set->grid[YY],set->grid[ZZ],set->rcut_coulomb,
 333             buft);
 334     if (fp_err != NULL)
 335     {
 336         fprintf(fp_err,"\r%s\n",buf);
 337     }
 338     if (fp_log != NULL)
 339     {
 340         fprintf(fp_log,"%s\n",buf);
 341     }
 342 }
 343
 344 static int pme_loadbal_end(pme_load_balancing_t pme_lb)
 345 {
 346     /* In the initial stage only n is set; end is not set yet */
 347     if (pme_lb->end > 0)
 348     {
 349         return pme_lb->end;
 350     }
 351     else
 352     {
 353         return pme_lb->n;
 354     }
 355 }
 356
 357 static void print_loadbal_limited(FILE *fp_err,FILE *fp_log,
 358                                   gmx_large_int_t step,
 359                                   pme_load_balancing_t pme_lb)
 360 {
 361     char buf[STRLEN],sbuf[22];
 362
 363     sprintf(buf,"step %4s: the %s limited the PME load balancing to a coulomb cut-off of %.3f",
 364             gmx_step_str(step,sbuf),
 365             pmelblim_str[pme_lb->elimited],
 366             pme_lb->setup[pme_loadbal_end(pme_lb)-1].rcut_coulomb);
 367     if (fp_err != NULL)
 368     {
 369         fprintf(fp_err,"\r%s\n",buf);
 370     }
 371     if (fp_log != NULL)
 372     {
 373         fprintf(fp_log,"%s\n",buf);
 374     }
 375 }
 376
 377 static void switch_to_stage1(pme_load_balancing_t pme_lb)
 378 {
 379     pme_lb->start = 0;
 380     while (pme_lb->start+1 < pme_lb->n &&
 381            (pme_lb->setup[pme_lb->start].count == 0 ||
 382             pme_lb->setup[pme_lb->start].cycles >
 383             pme_lb->setup[pme_lb->fastest].cycles*PME_LB_SLOW_FAC))
 384     {
 385         pme_lb->start++;
 386     }
 387     while (pme_lb->start > 0 && pme_lb->setup[pme_lb->start-1].cycles == 0)
 388     {
 389         pme_lb->start--;
 390     }
 391
 392     pme_lb->end = pme_lb->n;
 393     if (pme_lb->setup[pme_lb->end-1].count > 0 &&
 394         pme_lb->setup[pme_lb->end-1].cycles >
 395         pme_lb->setup[pme_lb->fastest].cycles*PME_LB_SLOW_FAC)
 396     {
 397         pme_lb->end--;
 398     }
 399
 400     pme_lb->stage = 1;
 401
 402     /* Next we want to choose setup pme_lb->start, but as we will increase
 403      * pme_ln->cur by one right after returning, we subtract 1 here.
 404      */
 405     pme_lb->cur = pme_lb->start - 1;
 406 }
 407
 408 gmx_bool pme_load_balance(pme_load_balancing_t pme_lb,
 409                           t_commrec *cr,
 410                           FILE *fp_err,
 411                           FILE *fp_log,
 412                           t_inputrec *ir,
 413                           t_state *state,
 414                           double cycles,
 415                           interaction_const_t *ic,
 416                           nonbonded_verlet_t *nbv,
 417                           gmx_pme_t *pmedata,
 418                           gmx_large_int_t step)
 419 {
 420     gmx_bool OK;
 421     pme_setup_t *set;
 422     double cycles_fast;
 423     char buf[STRLEN],sbuf[22];
 424     real rtab;
 425     gmx_bool bUsesSimpleTables = TRUE;
 426
 427     if (pme_lb->stage == pme_lb->nstage)
 428     {
 429         return FALSE;
 430     }
 431
 432     if (PAR(cr))
 433     {
 434         gmx_sumd(1,&cycles,cr);
 435         cycles /= cr->nnodes;
 436     }
 437
 438     set = &pme_lb->setup[pme_lb->cur];
 439     set->count++;
 440
 441     rtab = ir->rlistlong + ir->tabext;
 442
 443     if (set->count % 2 == 1)
 444     {
 445         /* Skip the first cycle, because the first step after a switch
 446          * is much slower due to allocation and/or caching effects.
 447          */
 448         return TRUE;
 449     }
 450
 451     sprintf(buf, "step %4s: ", gmx_step_str(step,sbuf));
 452     print_grid(fp_err,fp_log,buf,"timed with",set,cycles);
 453
 454     if (set->count <= 2)
 455     {
 456         set->cycles = cycles;
 457     }
 458     else
 459     {
 460         if (cycles*PME_LB_ACCEL_TOL < set->cycles &&
 461             pme_lb->stage == pme_lb->nstage - 1)
 462         {
 463             /* The performance went up a lot (due to e.g. DD load balancing).
 464              * Add a stage, keep the minima, but rescan all setups.
 465              */
 466             pme_lb->nstage++;
 467
 468             if (debug)
 469             {
 470                 fprintf(debug,"The performance for grid %d %d %d went from %.3f to %.1f M-cycles, this is more than %f\n"
 471                         "Increased the number stages to %d"
 472                         " and ignoring the previous performance\n",
 473                         set->grid[XX],set->grid[YY],set->grid[ZZ],
 474                         cycles*1e-6,set->cycles*1e-6,PME_LB_ACCEL_TOL,
 475                         pme_lb->nstage);
 476             }
 477         }
 478         set->cycles = min(set->cycles,cycles);
 479     }
 480
 481     if (set->cycles < pme_lb->setup[pme_lb->fastest].cycles)
 482     {
 483         pme_lb->fastest = pme_lb->cur;
 484     }
 485     cycles_fast = pme_lb->setup[pme_lb->fastest].cycles;
 486
 487     /* Check in stage 0 if we should stop scanning grids.
 488      * Stop when the time is more than SLOW_FAC longer than the fastest.
 489      */
 490     if (pme_lb->stage == 0 && pme_lb->cur > 0 &&
 491         cycles > pme_lb->setup[pme_lb->fastest].cycles*PME_LB_SLOW_FAC)
 492     {
 493         pme_lb->n = pme_lb->cur + 1;
 494         /* Done with scanning, go to stage 1 */
 495         switch_to_stage1(pme_lb);
 496     }
 497
 498     if (pme_lb->stage == 0)
 499     {
 500         int gridsize_start;
 501
 502         gridsize_start = set->grid[XX]*set->grid[YY]*set->grid[ZZ];
 503
 504         do
 505         {
 506             if (pme_lb->cur+1 < pme_lb->n)
 507             {
 508                 /* We had already generated the next setup */
 509                 OK = TRUE;
 510             }
 511             else
 512             {
 513                 /* Find the next setup */
 514                 OK = pme_loadbal_increase_cutoff(pme_lb,ir->pme_order);
 515             }
 516
 517             if (OK && ir->ePBC != epbcNONE)
 518             {
 519                 OK = (sqr(pme_lb->setup[pme_lb->cur+1].rlistlong)
 520                       <= max_cutoff2(ir->ePBC,state->box));
 521                 if (!OK)
 522                 {
 523                     pme_lb->elimited = epmelblimBOX;
 524                 }
 525             }
 526
 527             if (OK)
 528             {
 529                 pme_lb->cur++;
 530
 531                 if (DOMAINDECOMP(cr))
 532                 {
 533                     OK = change_dd_cutoff(cr,state,ir,
 534                                           pme_lb->setup[pme_lb->cur].rlistlong);
 535                     if (!OK)
 536                     {
 537                         /* Failed: do not use this setup */
 538                         pme_lb->cur--;
 539                         pme_lb->elimited = epmelblimDD;
 540                     }
 541                 }
 542             }
 543             if (!OK)
 544             {
 545                 /* We hit the upper limit for the cut-off,
 546                  * the setup should not go further than cur.
 547                  */
 548                 pme_lb->n = pme_lb->cur + 1;
 549                 print_loadbal_limited(fp_err,fp_log,step,pme_lb);
 550                 /* Switch to the next stage */
 551                 switch_to_stage1(pme_lb);
 552             }
 553         }
 554         while (OK &&
 555                !(pme_lb->setup[pme_lb->cur].grid[XX]*
 556                  pme_lb->setup[pme_lb->cur].grid[YY]*
 557                  pme_lb->setup[pme_lb->cur].grid[ZZ] <
 558                  gridsize_start*PME_LB_GRID_SCALE_FAC
 559                  &&
 560                  pme_lb->setup[pme_lb->cur].grid_efficiency <
 561                  pme_lb->setup[pme_lb->cur-1].grid_efficiency*PME_LB_GRID_EFFICIENCY_REL_FAC));
 562     }
 563
 564     if (pme_lb->stage > 0 && pme_lb->end == 1)
 565     {
 566         pme_lb->cur = 0;
 567         pme_lb->stage = pme_lb->nstage;
 568     }
 569     else if (pme_lb->stage > 0 && pme_lb->end > 1)
 570     {
 571         /* If stage = nstage-1:
 572          *   scan over all setups, rerunning only those setups
 573          *   which are not much slower than the fastest
 574          * else:
 575          *   use the next setup
 576          */
 577         do
 578         {
 579             pme_lb->cur++;
 580             if (pme_lb->cur == pme_lb->end)
 581             {
 582                 pme_lb->stage++;
 583                 pme_lb->cur = pme_lb->start;
 584             }
 585         }
 586         while (pme_lb->stage == pme_lb->nstage - 1 &&
 587                pme_lb->setup[pme_lb->cur].count > 0 &&
 588                pme_lb->setup[pme_lb->cur].cycles > cycles_fast*PME_LB_SLOW_FAC);
 589
 590         if (pme_lb->stage == pme_lb->nstage)
 591         {
 592             /* We are done optimizing, use the fastest setup we found */
 593             pme_lb->cur = pme_lb->fastest;
 594         }
 595     }
 596
 597     if (DOMAINDECOMP(cr) && pme_lb->stage > 0)
 598     {
 599         OK = change_dd_cutoff(cr,state,ir,pme_lb->setup[pme_lb->cur].rlistlong);
 600         if (!OK)
 601         {
 602             /* Failsafe solution */
 603             if (pme_lb->cur > 1 && pme_lb->stage == pme_lb->nstage)
 604             {
 605                 pme_lb->stage--;
 606             }
 607             pme_lb->fastest  = 0;
 608             pme_lb->start    = 0;
 609             pme_lb->end      = pme_lb->cur;
 610             pme_lb->cur      = pme_lb->start;
 611             pme_lb->elimited = epmelblimDD;
 612             print_loadbal_limited(fp_err,fp_log,step,pme_lb);
 613         }
 614     }
 615
 616     /* Change the Coulomb cut-off and the PME grid */
 617
 618     set = &pme_lb->setup[pme_lb->cur];
 619
 620     ic->rcoulomb   = set->rcut_coulomb;
 621     ic->rlist      = set->rlist;
 622     ic->rlistlong  = set->rlistlong;
 623     ir->nstcalclr  = set->nstcalclr;
 624     ic->ewaldcoeff = set->ewaldcoeff;
 625
 626     bUsesSimpleTables = uses_simple_tables(ir->cutoff_scheme, nbv, 0);
 627     if (pme_lb->cutoff_scheme == ecutsVERLET && nbv->grp[0].kernel_type == nbk8x8x8_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 }