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 {
  81     epmelblimNO, epmelblimBOX, epmelblimDD, epmelblimNR
  82 };
  83
  84 const char *pmelblim_str[epmelblimNR] =
  85 { "no", "box size", "domain decompostion" };
  86
  87 struct pme_load_balancing {
  88     int          nstage;             /* the current maximum number of stages */
  89
  90     real         cut_spacing;        /* the minimum cutoff / PME grid spacing ratio */
  91     real         rcut_vdw;           /* Vdw cutoff (does not change) */
  92     real         rcut_coulomb_start; /* Initial electrostatics cutoff */
  93     int          nstcalclr_start;    /* Initial electrostatics cutoff */
  94     real         rbuf_coulomb;       /* the pairlist buffer size */
  95     real         rbuf_vdw;           /* the pairlist buffer size */
  96     matrix       box_start;          /* the initial simulation box */
  97     int          n;                  /* the count of setup as well as the allocation size */
  98     pme_setup_t *setup;              /* the PME+cutoff setups */
  99     int          cur;                /* the current setup */
 100     int          fastest;            /* fastest setup up till now */
 101     int          start;              /* start of setup range to consider in stage>0 */
 102     int          end;                /* end   of setup range to consider in stage>0 */
 103     int          elimited;           /* was the balancing limited, uses enum above */
 104     int          cutoff_scheme;      /* Verlet or group cut-offs */
 105
 106     int          stage;              /* the current stage */
 107 };
 108
 109 void pme_loadbal_init(pme_load_balancing_t *pme_lb_p,
 110                       const t_inputrec *ir, matrix box,
 111                       const interaction_const_t *ic,
 112                       gmx_pme_t pmedata)
 113 {
 114     pme_load_balancing_t pme_lb;
 115     real                 spm, sp;
 116     int                  d;
 117
 118     snew(pme_lb, 1);
 119
 120     /* Any number of stages >= 2 is supported */
 121     pme_lb->nstage   = 2;
 122
 123     pme_lb->cutoff_scheme = ir->cutoff_scheme;
 124
 125     if (pme_lb->cutoff_scheme == ecutsVERLET)
 126     {
 127         pme_lb->rbuf_coulomb = ic->rlist - ic->rcoulomb;
 128         pme_lb->rbuf_vdw     = pme_lb->rbuf_coulomb;
 129     }
 130     else
 131     {
 132         if (ic->rcoulomb > ic->rlist)
 133         {
 134             pme_lb->rbuf_coulomb = ic->rlistlong - ic->rcoulomb;
 135         }
 136         else
 137         {
 138             pme_lb->rbuf_coulomb = ic->rlist - ic->rcoulomb;
 139         }
 140         if (ic->rvdw > ic->rlist)
 141         {
 142             pme_lb->rbuf_vdw = ic->rlistlong - ic->rvdw;
 143         }
 144         else
 145         {
 146             pme_lb->rbuf_vdw = ic->rlist - ic->rvdw;
 147         }
 148     }
 149
 150     copy_mat(box, pme_lb->box_start);
 151     if (ir->ePBC == epbcXY && ir->nwall == 2)
 152     {
 153         svmul(ir->wall_ewald_zfac, pme_lb->box_start[ZZ], pme_lb->box_start[ZZ]);
 154     }
 155
 156     pme_lb->n = 1;
 157     snew(pme_lb->setup, pme_lb->n);
 158
 159     pme_lb->rcut_vdw              = ic->rvdw;
 160     pme_lb->rcut_coulomb_start    = ir->rcoulomb;
 161     pme_lb->nstcalclr_start       = ir->nstcalclr;
 162
 163     pme_lb->cur                   = 0;
 164     pme_lb->setup[0].rcut_coulomb = ic->rcoulomb;
 165     pme_lb->setup[0].rlist        = ic->rlist;
 166     pme_lb->setup[0].rlistlong    = ic->rlistlong;
 167     pme_lb->setup[0].nstcalclr    = ir->nstcalclr;
 168     pme_lb->setup[0].grid[XX]     = ir->nkx;
 169     pme_lb->setup[0].grid[YY]     = ir->nky;
 170     pme_lb->setup[0].grid[ZZ]     = ir->nkz;
 171     pme_lb->setup[0].ewaldcoeff   = ic->ewaldcoeff;
 172
 173     pme_lb->setup[0].pmedata  = pmedata;
 174
 175     spm = 0;
 176     for (d = 0; d < DIM; d++)
 177     {
 178         sp = norm(pme_lb->box_start[d])/pme_lb->setup[0].grid[d];
 179         if (sp > spm)
 180         {
 181             spm = sp;
 182         }
 183     }
 184     pme_lb->setup[0].spacing = spm;
 185
 186     if (ir->fourier_spacing > 0)
 187     {
 188         pme_lb->cut_spacing = ir->rcoulomb/ir->fourier_spacing;
 189     }
 190     else
 191     {
 192         pme_lb->cut_spacing = ir->rcoulomb/pme_lb->setup[0].spacing;
 193     }
 194
 195     pme_lb->stage = 0;
 196
 197     pme_lb->fastest  = 0;
 198     pme_lb->start    = 0;
 199     pme_lb->end      = 0;
 200     pme_lb->elimited = epmelblimNO;
 201
 202     *pme_lb_p = pme_lb;
 203 }
 204
 205 static gmx_bool pme_loadbal_increase_cutoff(pme_load_balancing_t pme_lb,
 206                                             int                  pme_order)
 207 {
 208     pme_setup_t *set;
 209     real         fac, sp;
 210     real         tmpr_coulomb, tmpr_vdw;
 211     int          d;
 212
 213     /* Try to add a new setup with next larger cut-off to the list */
 214     pme_lb->n++;
 215     srenew(pme_lb->setup, pme_lb->n);
 216     set          = &pme_lb->setup[pme_lb->n-1];
 217     set->pmedata = NULL;
 218
 219     fac = 1;
 220     do
 221     {
 222         fac *= 1.01;
 223         clear_ivec(set->grid);
 224         sp = calc_grid(NULL, pme_lb->box_start,
 225                        fac*pme_lb->setup[pme_lb->cur].spacing,
 226                        &set->grid[XX],
 227                        &set->grid[YY],
 228                        &set->grid[ZZ]);
 229
 230         /* In parallel we can't have grids smaller than 2*pme_order,
 231          * and we would anyhow not gain much speed at these grid sizes.
 232          */
 233         for (d = 0; d < DIM; d++)
 234         {
 235             if (set->grid[d] <= 2*pme_order)
 236             {
 237                 pme_lb->n--;
 238
 239                 return FALSE;
 240             }
 241         }
 242     }
 243     while (sp <= 1.001*pme_lb->setup[pme_lb->cur].spacing);
 244
 245     set->rcut_coulomb = pme_lb->cut_spacing*sp;
 246
 247     if (pme_lb->cutoff_scheme == ecutsVERLET)
 248     {
 249         set->rlist        = set->rcut_coulomb + pme_lb->rbuf_coulomb;
 250         /* We dont use LR lists with Verlet, but this avoids if-statements in further checks */
 251         set->rlistlong    = set->rlist;
 252     }
 253     else
 254     {
 255         tmpr_coulomb          = set->rcut_coulomb + pme_lb->rbuf_coulomb;
 256         tmpr_vdw              = pme_lb->rcut_vdw + pme_lb->rbuf_vdw;
 257         set->rlist            = min(tmpr_coulomb, tmpr_vdw);
 258         set->rlistlong        = max(tmpr_coulomb, tmpr_vdw);
 259
 260         /* Set the long-range update frequency */
 261         if (set->rlist == set->rlistlong)
 262         {
 263             /* No long-range interactions if the short-/long-range cutoffs are identical */
 264             set->nstcalclr = 0;
 265         }
 266         else if (pme_lb->nstcalclr_start == 0 || pme_lb->nstcalclr_start == 1)
 267         {
 268             /* We were not doing long-range before, but now we are since rlist!=rlistlong */
 269             set->nstcalclr = 1;
 270         }
 271         else
 272         {
 273             /* We were already doing long-range interactions from the start */
 274             if (pme_lb->rcut_vdw > pme_lb->rcut_coulomb_start)
 275             {
 276                 /* We were originally doing long-range VdW-only interactions.
 277                  * If rvdw is still longer than rcoulomb we keep the original nstcalclr,
 278                  * but if the coulomb cutoff has become longer we should update the long-range
 279                  * part every step.
 280                  */
 281                 set->nstcalclr = (tmpr_vdw > tmpr_coulomb) ? pme_lb->nstcalclr_start : 1;
 282             }
 283             else
 284             {
 285                 /* We were not doing any long-range interaction from the start,
 286                  * since it is not possible to do twin-range coulomb for the PME interaction.
 287                  */
 288                 set->nstcalclr = 1;
 289             }
 290         }
 291     }
 292
 293     set->spacing      = sp;
 294     /* The grid efficiency is the size wrt a grid with uniform x/y/z spacing */
 295     set->grid_efficiency = 1;
 296     for (d = 0; d < DIM; d++)
 297     {
 298         set->grid_efficiency *= (set->grid[d]*sp)/norm(pme_lb->box_start[d]);
 299     }
 300     /* The Ewald coefficient is inversly proportional to the cut-off */
 301     set->ewaldcoeff =
 302         pme_lb->setup[0].ewaldcoeff*pme_lb->setup[0].rcut_coulomb/set->rcut_coulomb;
 303
 304     set->count   = 0;
 305     set->cycles  = 0;
 306
 307     if (debug)
 308     {
 309         fprintf(debug, "PME loadbal: grid %d %d %d, coulomb cutoff %f\n",
 310                 set->grid[XX], set->grid[YY], set->grid[ZZ], set->rcut_coulomb);
 311     }
 312     return TRUE;
 313 }
 314
 315 static void print_grid(FILE *fp_err, FILE *fp_log,
 316                        const char *pre,
 317                        const char *desc,
 318                        const pme_setup_t *set,
 319                        double cycles)
 320 {
 321     char buf[STRLEN], buft[STRLEN];
 322
 323     if (cycles >= 0)
 324     {
 325         sprintf(buft, ": %.1f M-cycles", cycles*1e-6);
 326     }
 327     else
 328     {
 329         buft[0] = '\0';
 330     }
 331     sprintf(buf, "%-11s%10s pme grid %d %d %d, coulomb cutoff %.3f%s",
 332             pre,
 333             desc, set->grid[XX], set->grid[YY], set->grid[ZZ], set->rcut_coulomb,
 334             buft);
 335     if (fp_err != NULL)
 336     {
 337         fprintf(fp_err, "\r%s\n", buf);
 338     }
 339     if (fp_log != NULL)
 340     {
 341         fprintf(fp_log, "%s\n", buf);
 342     }
 343 }
 344
 345 static int pme_loadbal_end(pme_load_balancing_t pme_lb)
 346 {
 347     /* In the initial stage only n is set; end is not set yet */
 348     if (pme_lb->end > 0)
 349     {
 350         return pme_lb->end;
 351     }
 352     else
 353     {
 354         return pme_lb->n;
 355     }
 356 }
 357
 358 static void print_loadbal_limited(FILE *fp_err, FILE *fp_log,
 359                                   gmx_large_int_t step,
 360                                   pme_load_balancing_t pme_lb)
 361 {
 362     char buf[STRLEN], sbuf[22];
 363
 364     sprintf(buf, "step %4s: the %s limited the PME load balancing to a coulomb cut-off of %.3f",
 365             gmx_step_str(step, sbuf),
 366             pmelblim_str[pme_lb->elimited],
 367             pme_lb->setup[pme_loadbal_end(pme_lb)-1].rcut_coulomb);
 368     if (fp_err != NULL)
 369     {
 370         fprintf(fp_err, "\r%s\n", buf);
 371     }
 372     if (fp_log != NULL)
 373     {
 374         fprintf(fp_log, "%s\n", buf);
 375     }
 376 }
 377
 378 static void switch_to_stage1(pme_load_balancing_t pme_lb)
 379 {
 380     pme_lb->start = 0;
 381     while (pme_lb->start+1 < pme_lb->n &&
 382            (pme_lb->setup[pme_lb->start].count == 0 ||
 383             pme_lb->setup[pme_lb->start].cycles >
 384             pme_lb->setup[pme_lb->fastest].cycles*PME_LB_SLOW_FAC))
 385     {
 386         pme_lb->start++;
 387     }
 388     while (pme_lb->start > 0 && pme_lb->setup[pme_lb->start-1].cycles == 0)
 389     {
 390         pme_lb->start--;
 391     }
 392
 393     pme_lb->end = pme_lb->n;
 394     if (pme_lb->setup[pme_lb->end-1].count > 0 &&
 395         pme_lb->setup[pme_lb->end-1].cycles >
 396         pme_lb->setup[pme_lb->fastest].cycles*PME_LB_SLOW_FAC)
 397     {
 398         pme_lb->end--;
 399     }
 400
 401     pme_lb->stage = 1;
 402
 403     /* Next we want to choose setup pme_lb->start, but as we will increase
 404      * pme_ln->cur by one right after returning, we subtract 1 here.
 405      */
 406     pme_lb->cur = pme_lb->start - 1;
 407 }
 408
 409 gmx_bool pme_load_balance(pme_load_balancing_t pme_lb,
 410                           t_commrec           *cr,
 411                           FILE                *fp_err,
 412                           FILE                *fp_log,
 413                           t_inputrec          *ir,
 414                           t_state             *state,
 415                           double               cycles,
 416                           interaction_const_t *ic,
 417                           nonbonded_verlet_t  *nbv,
 418                           gmx_pme_t           *pmedata,
 419                           gmx_large_int_t      step)
 420 {
 421     gmx_bool     OK;
 422     pme_setup_t *set;
 423     double       cycles_fast;
 424     char         buf[STRLEN], sbuf[22];
 425     real         rtab;
 426     gmx_bool     bUsesSimpleTables = TRUE;
 427
 428     if (pme_lb->stage == pme_lb->nstage)
 429     {
 430         return FALSE;
 431     }
 432
 433     if (PAR(cr))
 434     {
 435         gmx_sumd(1, &cycles, cr);
 436         cycles /= cr->nnodes;
 437     }
 438
 439     set = &pme_lb->setup[pme_lb->cur];
 440     set->count++;
 441
 442     rtab = ir->rlistlong + ir->tabext;
 443
 444     if (set->count % 2 == 1)
 445     {
 446         /* Skip the first cycle, because the first step after a switch
 447          * is much slower due to allocation and/or caching effects.
 448          */
 449         return TRUE;
 450     }
 451
 452     sprintf(buf, "step %4s: ", gmx_step_str(step, sbuf));
 453     print_grid(fp_err, fp_log, buf, "timed with", set, cycles);
 454
 455     if (set->count <= 2)
 456     {
 457         set->cycles = cycles;
 458     }
 459     else
 460     {
 461         if (cycles*PME_LB_ACCEL_TOL < set->cycles &&
 462             pme_lb->stage == pme_lb->nstage - 1)
 463         {
 464             /* The performance went up a lot (due to e.g. DD load balancing).
 465              * Add a stage, keep the minima, but rescan all setups.
 466              */
 467             pme_lb->nstage++;
 468
 469             if (debug)
 470             {
 471                 fprintf(debug, "The performance for grid %d %d %d went from %.3f to %.1f M-cycles, this is more than %f\n"
 472                         "Increased the number stages to %d"
 473                         " and ignoring the previous performance\n",
 474                         set->grid[XX], set->grid[YY], set->grid[ZZ],
 475                         cycles*1e-6, set->cycles*1e-6, PME_LB_ACCEL_TOL,
 476                         pme_lb->nstage);
 477             }
 478         }
 479         set->cycles = min(set->cycles, cycles);
 480     }
 481
 482     if (set->cycles < pme_lb->setup[pme_lb->fastest].cycles)
 483     {
 484         pme_lb->fastest = pme_lb->cur;
 485
 486         if (DOMAINDECOMP(cr))
 487         {
 488             /* We found a new fastest setting, ensure that with subsequent
 489              * shorter cut-off's the dynamic load balancing does not make
 490              * the use of the current cut-off impossible. This solution is
 491              * a trade-off, as the PME load balancing and DD domain size
 492              * load balancing can interact in complex ways.
 493              * With the Verlet kernels, DD load imbalance will usually be
 494              * mainly due to bonded interaction imbalance, which will often
 495              * quickly push the domain boundaries beyond the limit for the
 496              * optimal, PME load balanced, cut-off. But it could be that
 497              * better overal performance can be obtained with a slightly
 498              * shorter cut-off and better DD load balancing.
 499              */
 500             change_dd_dlb_cutoff_limit(cr);
 501         }
 502     }
 503     cycles_fast = pme_lb->setup[pme_lb->fastest].cycles;
 504
 505     /* Check in stage 0 if we should stop scanning grids.
 506      * Stop when the time is more than SLOW_FAC longer than the fastest.
 507      */
 508     if (pme_lb->stage == 0 && pme_lb->cur > 0 &&
 509         cycles > pme_lb->setup[pme_lb->fastest].cycles*PME_LB_SLOW_FAC)
 510     {
 511         pme_lb->n = pme_lb->cur + 1;
 512         /* Done with scanning, go to stage 1 */
 513         switch_to_stage1(pme_lb);
 514     }
 515
 516     if (pme_lb->stage == 0)
 517     {
 518         int gridsize_start;
 519
 520         gridsize_start = set->grid[XX]*set->grid[YY]*set->grid[ZZ];
 521
 522         do
 523         {
 524             if (pme_lb->cur+1 < pme_lb->n)
 525             {
 526                 /* We had already generated the next setup */
 527                 OK = TRUE;
 528             }
 529             else
 530             {
 531                 /* Find the next setup */
 532                 OK = pme_loadbal_increase_cutoff(pme_lb, ir->pme_order);
 533             }
 534
 535             if (OK && ir->ePBC != epbcNONE)
 536             {
 537                 OK = (sqr(pme_lb->setup[pme_lb->cur+1].rlistlong)
 538                       <= max_cutoff2(ir->ePBC, state->box));
 539                 if (!OK)
 540                 {
 541                     pme_lb->elimited = epmelblimBOX;
 542                 }
 543             }
 544
 545             if (OK)
 546             {
 547                 pme_lb->cur++;
 548
 549                 if (DOMAINDECOMP(cr))
 550                 {
 551                     OK = change_dd_cutoff(cr, state, ir,
 552                                           pme_lb->setup[pme_lb->cur].rlistlong);
 553                     if (!OK)
 554                     {
 555                         /* Failed: do not use this setup */
 556                         pme_lb->cur--;
 557                         pme_lb->elimited = epmelblimDD;
 558                     }
 559                 }
 560             }
 561             if (!OK)
 562             {
 563                 /* We hit the upper limit for the cut-off,
 564                  * the setup should not go further than cur.
 565                  */
 566                 pme_lb->n = pme_lb->cur + 1;
 567                 print_loadbal_limited(fp_err, fp_log, step, pme_lb);
 568                 /* Switch to the next stage */
 569                 switch_to_stage1(pme_lb);
 570             }
 571         }
 572         while (OK &&
 573                !(pme_lb->setup[pme_lb->cur].grid[XX]*
 574                  pme_lb->setup[pme_lb->cur].grid[YY]*
 575                  pme_lb->setup[pme_lb->cur].grid[ZZ] <
 576                  gridsize_start*PME_LB_GRID_SCALE_FAC
 577                  &&
 578                  pme_lb->setup[pme_lb->cur].grid_efficiency <
 579                  pme_lb->setup[pme_lb->cur-1].grid_efficiency*PME_LB_GRID_EFFICIENCY_REL_FAC));
 580     }
 581
 582     if (pme_lb->stage > 0 && pme_lb->end == 1)
 583     {
 584         pme_lb->cur   = 0;
 585         pme_lb->stage = pme_lb->nstage;
 586     }
 587     else if (pme_lb->stage > 0 && pme_lb->end > 1)
 588     {
 589         /* If stage = nstage-1:
 590          *   scan over all setups, rerunning only those setups
 591          *   which are not much slower than the fastest
 592          * else:
 593          *   use the next setup
 594          */
 595         do
 596         {
 597             pme_lb->cur++;
 598             if (pme_lb->cur == pme_lb->end)
 599             {
 600                 pme_lb->stage++;
 601                 pme_lb->cur = pme_lb->start;
 602             }
 603         }
 604         while (pme_lb->stage == pme_lb->nstage - 1 &&
 605                pme_lb->setup[pme_lb->cur].count > 0 &&
 606                pme_lb->setup[pme_lb->cur].cycles > cycles_fast*PME_LB_SLOW_FAC);
 607
 608         if (pme_lb->stage == pme_lb->nstage)
 609         {
 610             /* We are done optimizing, use the fastest setup we found */
 611             pme_lb->cur = pme_lb->fastest;
 612         }
 613     }
 614
 615     if (DOMAINDECOMP(cr) && pme_lb->stage > 0)
 616     {
 617         OK = change_dd_cutoff(cr, state, ir, pme_lb->setup[pme_lb->cur].rlistlong);
 618         if (!OK)
 619         {
 620             /* Failsafe solution */
 621             if (pme_lb->cur > 1 && pme_lb->stage == pme_lb->nstage)
 622             {
 623                 pme_lb->stage--;
 624             }
 625             pme_lb->fastest  = 0;
 626             pme_lb->start    = 0;
 627             pme_lb->end      = pme_lb->cur;
 628             pme_lb->cur      = pme_lb->start;
 629             pme_lb->elimited = epmelblimDD;
 630             print_loadbal_limited(fp_err, fp_log, step, pme_lb);
 631         }
 632     }
 633
 634     /* Change the Coulomb cut-off and the PME grid */
 635
 636     set = &pme_lb->setup[pme_lb->cur];
 637
 638     ic->rcoulomb   = set->rcut_coulomb;
 639     ic->rlist      = set->rlist;
 640     ic->rlistlong  = set->rlistlong;
 641     ir->nstcalclr  = set->nstcalclr;
 642     ic->ewaldcoeff = set->ewaldcoeff;
 643
 644     bUsesSimpleTables = uses_simple_tables(ir->cutoff_scheme, nbv, 0);
 645     if (pme_lb->cutoff_scheme == ecutsVERLET &&
 646         nbv->grp[0].kernel_type == nbnxnk8x8x8_CUDA)
 647     {
 648         nbnxn_cuda_pme_loadbal_update_param(nbv->cu_nbv, ic);
 649     }
 650     else
 651     {
 652         init_interaction_const_tables(NULL, ic, bUsesSimpleTables,
 653                                       rtab);
 654     }
 655
 656     if (pme_lb->cutoff_scheme == ecutsVERLET && nbv->ngrp > 1)
 657     {
 658         init_interaction_const_tables(NULL, ic, bUsesSimpleTables,
 659                                       rtab);
 660     }
 661
 662     if (cr->duty & DUTY_PME)
 663     {
 664         if (pme_lb->setup[pme_lb->cur].pmedata == NULL)
 665         {
 666             /* Generate a new PME data structure,
 667              * copying part of the old pointers.
 668              */
 669             gmx_pme_reinit(&set->pmedata,
 670                            cr, pme_lb->setup[0].pmedata, ir,
 671                            set->grid);
 672         }
 673         *pmedata = set->pmedata;
 674     }
 675     else
 676     {
 677         /* Tell our PME-only node to switch grid */
 678         gmx_pme_send_switch(cr, set->grid, set->ewaldcoeff);
 679     }
 680
 681     if (debug)
 682     {
 683         print_grid(NULL, debug, "", "switched to", set, -1);
 684     }
 685
 686     if (pme_lb->stage == pme_lb->nstage)
 687     {
 688         print_grid(fp_err, fp_log, "", "optimal", set, -1);
 689     }
 690
 691     return TRUE;
 692 }
 693
 694 void restart_pme_loadbal(pme_load_balancing_t pme_lb, int n)
 695 {
 696     pme_lb->nstage += n;
 697 }
 698
 699 static int pme_grid_points(const pme_setup_t *setup)
 700 {
 701     return setup->grid[XX]*setup->grid[YY]*setup->grid[ZZ];
 702 }
 703
 704 static void print_pme_loadbal_setting(FILE              *fplog,
 705                                       char              *name,
 706                                       const pme_setup_t *setup)
 707 {
 708     fprintf(fplog,
 709             "   %-7s %6.3f nm %6.3f nm     %3d %3d %3d   %5.3f nm  %5.3f nm\n",
 710             name,
 711             setup->rcut_coulomb, setup->rlist,
 712             setup->grid[XX], setup->grid[YY], setup->grid[ZZ],
 713             setup->spacing, 1/setup->ewaldcoeff);
 714 }
 715
 716 static void print_pme_loadbal_settings(pme_load_balancing_t pme_lb,
 717                                        FILE                *fplog)
 718 {
 719     double pp_ratio, grid_ratio;
 720
 721     pp_ratio   = pow(pme_lb->setup[pme_lb->cur].rlist/pme_lb->setup[0].rlistlong, 3.0);
 722     grid_ratio = pme_grid_points(&pme_lb->setup[pme_lb->cur])/
 723         (double)pme_grid_points(&pme_lb->setup[0]);
 724
 725     fprintf(fplog, "\n");
 726     fprintf(fplog, "       P P   -   P M E   L O A D   B A L A N C I N G\n");
 727     fprintf(fplog, "\n");
 728     /* Here we only warn when the optimal setting is the last one */
 729     if (pme_lb->elimited != epmelblimNO &&
 730         pme_lb->cur == pme_loadbal_end(pme_lb)-1)
 731     {
 732         fprintf(fplog, " NOTE: The PP/PME load balancing was limited by the %s,\n",
 733                 pmelblim_str[pme_lb->elimited]);
 734         fprintf(fplog, "       you might not have reached a good load balance.\n");
 735         if (pme_lb->elimited == epmelblimDD)
 736         {
 737             fprintf(fplog, "       Try different mdrun -dd settings or lower the -dds value.\n");
 738         }
 739         fprintf(fplog, "\n");
 740     }
 741     fprintf(fplog, " PP/PME load balancing changed the cut-off and PME settings:\n");
 742     fprintf(fplog, "           particle-particle                    PME\n");
 743     fprintf(fplog, "            rcoulomb  rlist            grid      spacing   1/beta\n");
 744     print_pme_loadbal_setting(fplog, "initial", &pme_lb->setup[0]);
 745     print_pme_loadbal_setting(fplog, "final", &pme_lb->setup[pme_lb->cur]);
 746     fprintf(fplog, " cost-ratio           %4.2f             %4.2f\n",
 747             pp_ratio, grid_ratio);
 748     fprintf(fplog, " (note that these numbers concern only part of the total PP and PME load)\n");
 749     fprintf(fplog, "\n");
 750 }
 751
 752 void pme_loadbal_done(pme_load_balancing_t pme_lb, FILE *fplog)
 753 {
 754     if (fplog != NULL && (pme_lb->cur > 0 || pme_lb->elimited != epmelblimNO))
 755     {
 756         print_pme_loadbal_settings(pme_lb, fplog);
 757     }
 758
 759     /* TODO: Here we should free all pointers in pme_lb,
 760      * but as it contains pme data structures,
 761      * we need to first make pme.c free all data.
 762      */
 763 }