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 "md_logging.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 {
  82     epmelblimNO, epmelblimBOX, epmelblimDD, epmelblimNR
  83 };
  84
  85 const char *pmelblim_str[epmelblimNR] =
  86 { "no", "box size", "domain decompostion" };
  87
  88 struct pme_load_balancing {
  89     int          nstage;             /* the current maximum number of stages */
  90
  91     real         cut_spacing;        /* the minimum cutoff / PME grid spacing ratio */
  92     real         rcut_vdw;           /* Vdw cutoff (does not change) */
  93     real         rcut_coulomb_start; /* Initial electrostatics cutoff */
  94     int          nstcalclr_start;    /* Initial electrostatics cutoff */
  95     real         rbuf_coulomb;       /* the pairlist buffer size */
  96     real         rbuf_vdw;           /* the pairlist buffer size */
  97     matrix       box_start;          /* the initial simulation box */
  98     int          n;                  /* the count of setup as well as the allocation size */
  99     pme_setup_t *setup;              /* the PME+cutoff setups */
 100     int          cur;                /* the current setup */
 101     int          fastest;            /* fastest setup up till now */
 102     int          start;              /* start of setup range to consider in stage>0 */
 103     int          end;                /* end   of setup range to consider in stage>0 */
 104     int          elimited;           /* was the balancing limited, uses enum above */
 105     int          cutoff_scheme;      /* Verlet or group cut-offs */
 106
 107     int          stage;              /* the current stage */
 108 };
 109
 110 void pme_loadbal_init(pme_load_balancing_t *pme_lb_p,
 111                       const t_inputrec *ir, matrix box,
 112                       const interaction_const_t *ic,
 113                       gmx_pme_t pmedata)
 114 {
 115     pme_load_balancing_t pme_lb;
 116     real                 spm, sp;
 117     int                  d;
 118
 119     snew(pme_lb, 1);
 120
 121     /* Any number of stages >= 2 is supported */
 122     pme_lb->nstage   = 2;
 123
 124     pme_lb->cutoff_scheme = ir->cutoff_scheme;
 125
 126     if (pme_lb->cutoff_scheme == ecutsVERLET)
 127     {
 128         pme_lb->rbuf_coulomb = ic->rlist - ic->rcoulomb;
 129         pme_lb->rbuf_vdw     = pme_lb->rbuf_coulomb;
 130     }
 131     else
 132     {
 133         if (ic->rcoulomb > ic->rlist)
 134         {
 135             pme_lb->rbuf_coulomb = ic->rlistlong - ic->rcoulomb;
 136         }
 137         else
 138         {
 139             pme_lb->rbuf_coulomb = ic->rlist - ic->rcoulomb;
 140         }
 141         if (ic->rvdw > ic->rlist)
 142         {
 143             pme_lb->rbuf_vdw = ic->rlistlong - ic->rvdw;
 144         }
 145         else
 146         {
 147             pme_lb->rbuf_vdw = ic->rlist - ic->rvdw;
 148         }
 149     }
 150
 151     copy_mat(box, pme_lb->box_start);
 152     if (ir->ePBC == epbcXY && ir->nwall == 2)
 153     {
 154         svmul(ir->wall_ewald_zfac, pme_lb->box_start[ZZ], pme_lb->box_start[ZZ]);
 155     }
 156
 157     pme_lb->n = 1;
 158     snew(pme_lb->setup, pme_lb->n);
 159
 160     pme_lb->rcut_vdw              = ic->rvdw;
 161     pme_lb->rcut_coulomb_start    = ir->rcoulomb;
 162     pme_lb->nstcalclr_start       = ir->nstcalclr;
 163
 164     pme_lb->cur                   = 0;
 165     pme_lb->setup[0].rcut_coulomb = ic->rcoulomb;
 166     pme_lb->setup[0].rlist        = ic->rlist;
 167     pme_lb->setup[0].rlistlong    = ic->rlistlong;
 168     pme_lb->setup[0].nstcalclr    = ir->nstcalclr;
 169     pme_lb->setup[0].grid[XX]     = ir->nkx;
 170     pme_lb->setup[0].grid[YY]     = ir->nky;
 171     pme_lb->setup[0].grid[ZZ]     = ir->nkz;
 172     pme_lb->setup[0].ewaldcoeff   = ic->ewaldcoeff;
 173
 174     pme_lb->setup[0].pmedata  = pmedata;
 175
 176     spm = 0;
 177     for (d = 0; d < DIM; d++)
 178     {
 179         sp = norm(pme_lb->box_start[d])/pme_lb->setup[0].grid[d];
 180         if (sp > spm)
 181         {
 182             spm = sp;
 183         }
 184     }
 185     pme_lb->setup[0].spacing = spm;
 186
 187     if (ir->fourier_spacing > 0)
 188     {
 189         pme_lb->cut_spacing = ir->rcoulomb/ir->fourier_spacing;
 190     }
 191     else
 192     {
 193         pme_lb->cut_spacing = ir->rcoulomb/pme_lb->setup[0].spacing;
 194     }
 195
 196     pme_lb->stage = 0;
 197
 198     pme_lb->fastest  = 0;
 199     pme_lb->start    = 0;
 200     pme_lb->end      = 0;
 201     pme_lb->elimited = epmelblimNO;
 202
 203     *pme_lb_p = pme_lb;
 204 }
 205
 206 static gmx_bool pme_loadbal_increase_cutoff(pme_load_balancing_t pme_lb,
 207                                             int                  pme_order)
 208 {
 209     pme_setup_t *set;
 210     real         fac, sp;
 211     real         tmpr_coulomb, tmpr_vdw;
 212     int          d;
 213
 214     /* Try to add a new setup with next larger cut-off to the list */
 215     pme_lb->n++;
 216     srenew(pme_lb->setup, pme_lb->n);
 217     set          = &pme_lb->setup[pme_lb->n-1];
 218     set->pmedata = NULL;
 219
 220     fac = 1;
 221     do
 222     {
 223         fac *= 1.01;
 224         clear_ivec(set->grid);
 225         sp = calc_grid(NULL, pme_lb->box_start,
 226                        fac*pme_lb->setup[pme_lb->cur].spacing,
 227                        &set->grid[XX],
 228                        &set->grid[YY],
 229                        &set->grid[ZZ]);
 230
 231         /* In parallel we can't have grids smaller than 2*pme_order,
 232          * and we would anyhow not gain much speed at these grid sizes.
 233          */
 234         for (d = 0; d < DIM; d++)
 235         {
 236             if (set->grid[d] <= 2*pme_order)
 237             {
 238                 pme_lb->n--;
 239
 240                 return FALSE;
 241             }
 242         }
 243     }
 244     while (sp <= 1.001*pme_lb->setup[pme_lb->cur].spacing);
 245
 246     set->rcut_coulomb = pme_lb->cut_spacing*sp;
 247
 248     if (pme_lb->cutoff_scheme == ecutsVERLET)
 249     {
 250         set->rlist        = set->rcut_coulomb + pme_lb->rbuf_coulomb;
 251         /* We dont use LR lists with Verlet, but this avoids if-statements in further checks */
 252         set->rlistlong    = set->rlist;
 253     }
 254     else
 255     {
 256         tmpr_coulomb          = set->rcut_coulomb + pme_lb->rbuf_coulomb;
 257         tmpr_vdw              = pme_lb->rcut_vdw + pme_lb->rbuf_vdw;
 258         set->rlist            = min(tmpr_coulomb, tmpr_vdw);
 259         set->rlistlong        = max(tmpr_coulomb, tmpr_vdw);
 260
 261         /* Set the long-range update frequency */
 262         if (set->rlist == set->rlistlong)
 263         {
 264             /* No long-range interactions if the short-/long-range cutoffs are identical */
 265             set->nstcalclr = 0;
 266         }
 267         else if (pme_lb->nstcalclr_start == 0 || pme_lb->nstcalclr_start == 1)
 268         {
 269             /* We were not doing long-range before, but now we are since rlist!=rlistlong */
 270             set->nstcalclr = 1;
 271         }
 272         else
 273         {
 274             /* We were already doing long-range interactions from the start */
 275             if (pme_lb->rcut_vdw > pme_lb->rcut_coulomb_start)
 276             {
 277                 /* We were originally doing long-range VdW-only interactions.
 278                  * If rvdw is still longer than rcoulomb we keep the original nstcalclr,
 279                  * but if the coulomb cutoff has become longer we should update the long-range
 280                  * part every step.
 281                  */
 282                 set->nstcalclr = (tmpr_vdw > tmpr_coulomb) ? pme_lb->nstcalclr_start : 1;
 283             }
 284             else
 285             {
 286                 /* We were not doing any long-range interaction from the start,
 287                  * since it is not possible to do twin-range coulomb for the PME interaction.
 288                  */
 289                 set->nstcalclr = 1;
 290             }
 291         }
 292     }
 293
 294     set->spacing      = sp;
 295     /* The grid efficiency is the size wrt a grid with uniform x/y/z spacing */
 296     set->grid_efficiency = 1;
 297     for (d = 0; d < DIM; d++)
 298     {
 299         set->grid_efficiency *= (set->grid[d]*sp)/norm(pme_lb->box_start[d]);
 300     }
 301     /* The Ewald coefficient is inversly proportional to the cut-off */
 302     set->ewaldcoeff =
 303         pme_lb->setup[0].ewaldcoeff*pme_lb->setup[0].rcut_coulomb/set->rcut_coulomb;
 304
 305     set->count   = 0;
 306     set->cycles  = 0;
 307
 308     if (debug)
 309     {
 310         fprintf(debug, "PME loadbal: grid %d %d %d, coulomb cutoff %f\n",
 311                 set->grid[XX], set->grid[YY], set->grid[ZZ], set->rcut_coulomb);
 312     }
 313     return TRUE;
 314 }
 315
 316 static void print_grid(FILE *fp_err, FILE *fp_log,
 317                        const char *pre,
 318                        const char *desc,
 319                        const pme_setup_t *set,
 320                        double cycles)
 321 {
 322     char buf[STRLEN], buft[STRLEN];
 323
 324     if (cycles >= 0)
 325     {
 326         sprintf(buft, ": %.1f M-cycles", cycles*1e-6);
 327     }
 328     else
 329     {
 330         buft[0] = '\0';
 331     }
 332     sprintf(buf, "%-11s%10s pme grid %d %d %d, coulomb cutoff %.3f%s",
 333             pre,
 334             desc, set->grid[XX], set->grid[YY], set->grid[ZZ], set->rcut_coulomb,
 335             buft);
 336     if (fp_err != NULL)
 337     {
 338         fprintf(fp_err, "\r%s\n", buf);
 339     }
 340     if (fp_log != NULL)
 341     {
 342         fprintf(fp_log, "%s\n", buf);
 343     }
 344 }
 345
 346 static int pme_loadbal_end(pme_load_balancing_t pme_lb)
 347 {
 348     /* In the initial stage only n is set; end is not set yet */
 349     if (pme_lb->end > 0)
 350     {
 351         return pme_lb->end;
 352     }
 353     else
 354     {
 355         return pme_lb->n;
 356     }
 357 }
 358
 359 static void print_loadbal_limited(FILE *fp_err, FILE *fp_log,
 360                                   gmx_large_int_t step,
 361                                   pme_load_balancing_t pme_lb)
 362 {
 363     char buf[STRLEN], sbuf[22];
 364
 365     sprintf(buf, "step %4s: the %s limited the PME load balancing to a coulomb cut-off of %.3f",
 366             gmx_step_str(step, sbuf),
 367             pmelblim_str[pme_lb->elimited],
 368             pme_lb->setup[pme_loadbal_end(pme_lb)-1].rcut_coulomb);
 369     if (fp_err != NULL)
 370     {
 371         fprintf(fp_err, "\r%s\n", buf);
 372     }
 373     if (fp_log != NULL)
 374     {
 375         fprintf(fp_log, "%s\n", buf);
 376     }
 377 }
 378
 379 static void switch_to_stage1(pme_load_balancing_t pme_lb)
 380 {
 381     pme_lb->start = 0;
 382     while (pme_lb->start+1 < pme_lb->n &&
 383            (pme_lb->setup[pme_lb->start].count == 0 ||
 384             pme_lb->setup[pme_lb->start].cycles >
 385             pme_lb->setup[pme_lb->fastest].cycles*PME_LB_SLOW_FAC))
 386     {
 387         pme_lb->start++;
 388     }
 389     while (pme_lb->start > 0 && pme_lb->setup[pme_lb->start-1].cycles == 0)
 390     {
 391         pme_lb->start--;
 392     }
 393
 394     pme_lb->end = pme_lb->n;
 395     if (pme_lb->setup[pme_lb->end-1].count > 0 &&
 396         pme_lb->setup[pme_lb->end-1].cycles >
 397         pme_lb->setup[pme_lb->fastest].cycles*PME_LB_SLOW_FAC)
 398     {
 399         pme_lb->end--;
 400     }
 401
 402     pme_lb->stage = 1;
 403
 404     /* Next we want to choose setup pme_lb->start, but as we will increase
 405      * pme_ln->cur by one right after returning, we subtract 1 here.
 406      */
 407     pme_lb->cur = pme_lb->start - 1;
 408 }
 409
 410 gmx_bool pme_load_balance(pme_load_balancing_t pme_lb,
 411                           t_commrec           *cr,
 412                           FILE                *fp_err,
 413                           FILE                *fp_log,
 414                           t_inputrec          *ir,
 415                           t_state             *state,
 416                           double               cycles,
 417                           interaction_const_t *ic,
 418                           nonbonded_verlet_t  *nbv,
 419                           gmx_pme_t           *pmedata,
 420                           gmx_large_int_t      step)
 421 {
 422     gmx_bool     OK;
 423     pme_setup_t *set;
 424     double       cycles_fast;
 425     char         buf[STRLEN], sbuf[22];
 426     real         rtab;
 427     gmx_bool     bUsesSimpleTables = TRUE;
 428
 429     if (pme_lb->stage == pme_lb->nstage)
 430     {
 431         return FALSE;
 432     }
 433
 434     if (PAR(cr))
 435     {
 436         gmx_sumd(1, &cycles, cr);
 437         cycles /= cr->nnodes;
 438     }
 439
 440     set = &pme_lb->setup[pme_lb->cur];
 441     set->count++;
 442
 443     rtab = ir->rlistlong + ir->tabext;
 444
 445     if (set->count % 2 == 1)
 446     {
 447         /* Skip the first cycle, because the first step after a switch
 448          * is much slower due to allocation and/or caching effects.
 449          */
 450         return TRUE;
 451     }
 452
 453     sprintf(buf, "step %4s: ", gmx_step_str(step, sbuf));
 454     print_grid(fp_err, fp_log, buf, "timed with", set, cycles);
 455
 456     if (set->count <= 2)
 457     {
 458         set->cycles = cycles;
 459     }
 460     else
 461     {
 462         if (cycles*PME_LB_ACCEL_TOL < set->cycles &&
 463             pme_lb->stage == pme_lb->nstage - 1)
 464         {
 465             /* The performance went up a lot (due to e.g. DD load balancing).
 466              * Add a stage, keep the minima, but rescan all setups.
 467              */
 468             pme_lb->nstage++;
 469
 470             if (debug)
 471             {
 472                 fprintf(debug, "The performance for grid %d %d %d went from %.3f to %.1f M-cycles, this is more than %f\n"
 473                         "Increased the number stages to %d"
 474                         " and ignoring the previous performance\n",
 475                         set->grid[XX], set->grid[YY], set->grid[ZZ],
 476                         cycles*1e-6, set->cycles*1e-6, PME_LB_ACCEL_TOL,
 477                         pme_lb->nstage);
 478             }
 479         }
 480         set->cycles = min(set->cycles, cycles);
 481     }
 482
 483     if (set->cycles < pme_lb->setup[pme_lb->fastest].cycles)
 484     {
 485         pme_lb->fastest = pme_lb->cur;
 486
 487         if (DOMAINDECOMP(cr))
 488         {
 489             /* We found a new fastest setting, ensure that with subsequent
 490              * shorter cut-off's the dynamic load balancing does not make
 491              * the use of the current cut-off impossible. This solution is
 492              * a trade-off, as the PME load balancing and DD domain size
 493              * load balancing can interact in complex ways.
 494              * With the Verlet kernels, DD load imbalance will usually be
 495              * mainly due to bonded interaction imbalance, which will often
 496              * quickly push the domain boundaries beyond the limit for the
 497              * optimal, PME load balanced, cut-off. But it could be that
 498              * better overal performance can be obtained with a slightly
 499              * shorter cut-off and better DD load balancing.
 500              */
 501             change_dd_dlb_cutoff_limit(cr);
 502         }
 503     }
 504     cycles_fast = pme_lb->setup[pme_lb->fastest].cycles;
 505
 506     /* Check in stage 0 if we should stop scanning grids.
 507      * Stop when the time is more than SLOW_FAC longer than the fastest.
 508      */
 509     if (pme_lb->stage == 0 && pme_lb->cur > 0 &&
 510         cycles > pme_lb->setup[pme_lb->fastest].cycles*PME_LB_SLOW_FAC)
 511     {
 512         pme_lb->n = pme_lb->cur + 1;
 513         /* Done with scanning, go to stage 1 */
 514         switch_to_stage1(pme_lb);
 515     }
 516
 517     if (pme_lb->stage == 0)
 518     {
 519         int gridsize_start;
 520
 521         gridsize_start = set->grid[XX]*set->grid[YY]*set->grid[ZZ];
 522
 523         do
 524         {
 525             if (pme_lb->cur+1 < pme_lb->n)
 526             {
 527                 /* We had already generated the next setup */
 528                 OK = TRUE;
 529             }
 530             else
 531             {
 532                 /* Find the next setup */
 533                 OK = pme_loadbal_increase_cutoff(pme_lb, ir->pme_order);
 534             }
 535
 536             if (OK && ir->ePBC != epbcNONE)
 537             {
 538                 OK = (sqr(pme_lb->setup[pme_lb->cur+1].rlistlong)
 539                       <= max_cutoff2(ir->ePBC, state->box));
 540                 if (!OK)
 541                 {
 542                     pme_lb->elimited = epmelblimBOX;
 543                 }
 544             }
 545
 546             if (OK)
 547             {
 548                 pme_lb->cur++;
 549
 550                 if (DOMAINDECOMP(cr))
 551                 {
 552                     OK = change_dd_cutoff(cr, state, ir,
 553                                           pme_lb->setup[pme_lb->cur].rlistlong);
 554                     if (!OK)
 555                     {
 556                         /* Failed: do not use this setup */
 557                         pme_lb->cur--;
 558                         pme_lb->elimited = epmelblimDD;
 559                     }
 560                 }
 561             }
 562             if (!OK)
 563             {
 564                 /* We hit the upper limit for the cut-off,
 565                  * the setup should not go further than cur.
 566                  */
 567                 pme_lb->n = pme_lb->cur + 1;
 568                 print_loadbal_limited(fp_err, fp_log, step, pme_lb);
 569                 /* Switch to the next stage */
 570                 switch_to_stage1(pme_lb);
 571             }
 572         }
 573         while (OK &&
 574                !(pme_lb->setup[pme_lb->cur].grid[XX]*
 575                  pme_lb->setup[pme_lb->cur].grid[YY]*
 576                  pme_lb->setup[pme_lb->cur].grid[ZZ] <
 577                  gridsize_start*PME_LB_GRID_SCALE_FAC
 578                  &&
 579                  pme_lb->setup[pme_lb->cur].grid_efficiency <
 580                  pme_lb->setup[pme_lb->cur-1].grid_efficiency*PME_LB_GRID_EFFICIENCY_REL_FAC));
 581     }
 582
 583     if (pme_lb->stage > 0 && pme_lb->end == 1)
 584     {
 585         pme_lb->cur   = 0;
 586         pme_lb->stage = pme_lb->nstage;
 587     }
 588     else if (pme_lb->stage > 0 && pme_lb->end > 1)
 589     {
 590         /* If stage = nstage-1:
 591          *   scan over all setups, rerunning only those setups
 592          *   which are not much slower than the fastest
 593          * else:
 594          *   use the next setup
 595          */
 596         do
 597         {
 598             pme_lb->cur++;
 599             if (pme_lb->cur == pme_lb->end)
 600             {
 601                 pme_lb->stage++;
 602                 pme_lb->cur = pme_lb->start;
 603             }
 604         }
 605         while (pme_lb->stage == pme_lb->nstage - 1 &&
 606                pme_lb->setup[pme_lb->cur].count > 0 &&
 607                pme_lb->setup[pme_lb->cur].cycles > cycles_fast*PME_LB_SLOW_FAC);
 608
 609         if (pme_lb->stage == pme_lb->nstage)
 610         {
 611             /* We are done optimizing, use the fastest setup we found */
 612             pme_lb->cur = pme_lb->fastest;
 613         }
 614     }
 615
 616     if (DOMAINDECOMP(cr) && pme_lb->stage > 0)
 617     {
 618         OK = change_dd_cutoff(cr, state, ir, pme_lb->setup[pme_lb->cur].rlistlong);
 619         if (!OK)
 620         {
 621             /* Failsafe solution */
 622             if (pme_lb->cur > 1 && pme_lb->stage == pme_lb->nstage)
 623             {
 624                 pme_lb->stage--;
 625             }
 626             pme_lb->fastest  = 0;
 627             pme_lb->start    = 0;
 628             pme_lb->end      = pme_lb->cur;
 629             pme_lb->cur      = pme_lb->start;
 630             pme_lb->elimited = epmelblimDD;
 631             print_loadbal_limited(fp_err, fp_log, step, pme_lb);
 632         }
 633     }
 634
 635     /* Change the Coulomb cut-off and the PME grid */
 636
 637     set = &pme_lb->setup[pme_lb->cur];
 638
 639     ic->rcoulomb   = set->rcut_coulomb;
 640     ic->rlist      = set->rlist;
 641     ic->rlistlong  = set->rlistlong;
 642     ir->nstcalclr  = set->nstcalclr;
 643     ic->ewaldcoeff = set->ewaldcoeff;
 644
 645     bUsesSimpleTables = uses_simple_tables(ir->cutoff_scheme, nbv, 0);
 646     if (pme_lb->cutoff_scheme == ecutsVERLET &&
 647         nbv->grp[0].kernel_type == nbnxnk8x8x8_CUDA)
 648     {
 649         nbnxn_cuda_pme_loadbal_update_param(nbv->cu_nbv, ic);
 650     }
 651     else
 652     {
 653         init_interaction_const_tables(NULL, ic, bUsesSimpleTables,
 654                                       rtab);
 655     }
 656
 657     if (pme_lb->cutoff_scheme == ecutsVERLET && nbv->ngrp > 1)
 658     {
 659         init_interaction_const_tables(NULL, ic, bUsesSimpleTables,
 660                                       rtab);
 661     }
 662
 663     if (cr->duty & DUTY_PME)
 664     {
 665         if (pme_lb->setup[pme_lb->cur].pmedata == NULL)
 666         {
 667             /* Generate a new PME data structure,
 668              * copying part of the old pointers.
 669              */
 670             gmx_pme_reinit(&set->pmedata,
 671                            cr, pme_lb->setup[0].pmedata, ir,
 672                            set->grid);
 673         }
 674         *pmedata = set->pmedata;
 675     }
 676     else
 677     {
 678         /* Tell our PME-only node to switch grid */
 679         gmx_pme_send_switchgrid(cr, set->grid, set->ewaldcoeff);
 680     }
 681
 682     if (debug)
 683     {
 684         print_grid(NULL, debug, "", "switched to", set, -1);
 685     }
 686
 687     if (pme_lb->stage == pme_lb->nstage)
 688     {
 689         print_grid(fp_err, fp_log, "", "optimal", set, -1);
 690     }
 691
 692     return TRUE;
 693 }
 694
 695 void restart_pme_loadbal(pme_load_balancing_t pme_lb, int n)
 696 {
 697     pme_lb->nstage += n;
 698 }
 699
 700 static int pme_grid_points(const pme_setup_t *setup)
 701 {
 702     return setup->grid[XX]*setup->grid[YY]*setup->grid[ZZ];
 703 }
 704
 705 static real pme_loadbal_rlist(const pme_setup_t *setup)
 706 {
 707     /* With the group cut-off scheme we can have twin-range either
 708      * for Coulomb or for VdW, so we need a check here.
 709      * With the Verlet cut-off scheme rlist=rlistlong.
 710      */
 711     if (setup->rcut_coulomb > setup->rlist)
 712     {
 713         return setup->rlistlong;
 714     }
 715     else
 716     {
 717         return setup->rlist;
 718     }
 719 }
 720
 721 static void print_pme_loadbal_setting(FILE              *fplog,
 722                                       char              *name,
 723                                       const pme_setup_t *setup)
 724 {
 725     fprintf(fplog,
 726             "   %-7s %6.3f nm %6.3f nm     %3d %3d %3d   %5.3f nm  %5.3f nm\n",
 727             name,
 728             setup->rcut_coulomb, pme_loadbal_rlist(setup),
 729             setup->grid[XX], setup->grid[YY], setup->grid[ZZ],
 730             setup->spacing, 1/setup->ewaldcoeff);
 731 }
 732
 733 static void print_pme_loadbal_settings(pme_load_balancing_t pme_lb,
 734                                        t_commrec           *cr,
 735                                        FILE                *fplog,
 736                                        gmx_bool             bNonBondedOnGPU)
 737 {
 738     double pp_ratio, grid_ratio;
 739
 740     pp_ratio   = pow(pme_loadbal_rlist(&pme_lb->setup[pme_lb->cur])/pme_loadbal_rlist(&pme_lb->setup[0]), 3.0);
 741     grid_ratio = pme_grid_points(&pme_lb->setup[pme_lb->cur])/
 742         (double)pme_grid_points(&pme_lb->setup[0]);
 743
 744     fprintf(fplog, "\n");
 745     fprintf(fplog, "       P P   -   P M E   L O A D   B A L A N C I N G\n");
 746     fprintf(fplog, "\n");
 747     /* Here we only warn when the optimal setting is the last one */
 748     if (pme_lb->elimited != epmelblimNO &&
 749         pme_lb->cur == pme_loadbal_end(pme_lb)-1)
 750     {
 751         fprintf(fplog, " NOTE: The PP/PME load balancing was limited by the %s,\n",
 752                 pmelblim_str[pme_lb->elimited]);
 753         fprintf(fplog, "       you might not have reached a good load balance.\n");
 754         if (pme_lb->elimited == epmelblimDD)
 755         {
 756             fprintf(fplog, "       Try different mdrun -dd settings or lower the -dds value.\n");
 757         }
 758         fprintf(fplog, "\n");
 759     }
 760     fprintf(fplog, " PP/PME load balancing changed the cut-off and PME settings:\n");
 761     fprintf(fplog, "           particle-particle                    PME\n");
 762     fprintf(fplog, "            rcoulomb  rlist            grid      spacing   1/beta\n");
 763     print_pme_loadbal_setting(fplog, "initial", &pme_lb->setup[0]);
 764     print_pme_loadbal_setting(fplog, "final", &pme_lb->setup[pme_lb->cur]);
 765     fprintf(fplog, " cost-ratio           %4.2f             %4.2f\n",
 766             pp_ratio, grid_ratio);
 767     fprintf(fplog, " (note that these numbers concern only part of the total PP and PME load)\n");
 768
 769     if (pp_ratio > 1.5 && !bNonBondedOnGPU)
 770     {
 771         md_print_warn(cr, fplog,
 772                       "NOTE: PME load balancing increased the non-bonded workload by more than 50%%.\n"
 773                       "      For better performance use (more) PME nodes (mdrun -npme),\n"
 774                       "      or in case you are beyond the scaling limit, use less nodes in total.\n");
 775     }
 776     else
 777     {
 778         fprintf(fplog, "\n");
 779     }
 780 }
 781
 782 void pme_loadbal_done(pme_load_balancing_t pme_lb,
 783                       t_commrec *cr, FILE *fplog,
 784                       gmx_bool bNonBondedOnGPU)
 785 {
 786     if (fplog != NULL && (pme_lb->cur > 0 || pme_lb->elimited != epmelblimNO))
 787     {
 788         print_pme_loadbal_settings(pme_lb, cr, fplog, bNonBondedOnGPU);
 789     }
 790
 791     /* TODO: Here we should free all pointers in pme_lb,
 792      * but as it contains pme data structures,
 793      * we need to first make pme.c free all data.
 794      */
 795 }