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) 2012,2013, 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 "md_logging.h"
  51 #include "pme_loadbal.h"
  52
  53 /* Parameters and setting for one PP-PME setup */
  54 typedef struct {
  55     real      rcut_coulomb;    /* Coulomb cut-off                              */
  56     real      rlist;           /* pair-list cut-off                            */
  57     real      rlistlong;       /* LR pair-list cut-off                         */
  58     int       nstcalclr;       /* frequency of evaluating long-range forces for group scheme */
  59     real      spacing;         /* (largest) PME grid spacing                   */
  60     ivec      grid;            /* the PME grid dimensions                      */
  61     real      grid_efficiency; /* ineffiency factor for non-uniform grids <= 1 */
  62     real      ewaldcoeff;      /* the Ewald coefficient                        */
  63     gmx_pme_t pmedata;         /* the data structure used in the PME code      */
  64
  65     int       count;           /* number of times this setup has been timed    */
  66     double    cycles;          /* the fastest time for this setup in cycles    */
  67 } pme_setup_t;
  68
  69 /* In the initial scan, step by grids that are at least a factor 0.8 coarser */
  70 #define PME_LB_GRID_SCALE_FAC  0.8
  71 /* In the initial scan, try to skip grids with uneven x/y/z spacing,
  72  * checking if the "efficiency" is more than 5% worse than the previous grid.
  73  */
  74 #define PME_LB_GRID_EFFICIENCY_REL_FAC  1.05
  75 /* Rerun up till 12% slower setups than the fastest up till now */
  76 #define PME_LB_SLOW_FAC  1.12
  77 /* If setups get more than 2% faster, do another round to avoid
  78  * choosing a slower setup due to acceleration or fluctuations.
  79  */
  80 #define PME_LB_ACCEL_TOL 1.02
  81
  82 enum {
  83     epmelblimNO, epmelblimBOX, epmelblimDD, epmelblimNR
  84 };
  85
  86 const char *pmelblim_str[epmelblimNR] =
  87 { "no", "box size", "domain decompostion" };
  88
  89 struct pme_load_balancing {
  90     int          nstage;             /* the current maximum number of stages */
  91
  92     real         cut_spacing;        /* the minimum cutoff / PME grid spacing ratio */
  93     real         rcut_vdw;           /* Vdw cutoff (does not change) */
  94     real         rcut_coulomb_start; /* Initial electrostatics cutoff */
  95     int          nstcalclr_start;    /* Initial electrostatics cutoff */
  96     real         rbuf_coulomb;       /* the pairlist buffer size */
  97     real         rbuf_vdw;           /* the pairlist buffer size */
  98     matrix       box_start;          /* the initial simulation box */
  99     int          n;                  /* the count of setup as well as the allocation size */
 100     pme_setup_t *setup;              /* the PME+cutoff setups */
 101     int          cur;                /* the current setup */
 102     int          fastest;            /* fastest setup up till now */
 103     int          start;              /* start of setup range to consider in stage>0 */
 104     int          end;                /* end   of setup range to consider in stage>0 */
 105     int          elimited;           /* was the balancing limited, uses enum above */
 106     int          cutoff_scheme;      /* Verlet or group cut-offs */
 107
 108     int          stage;              /* the current stage */
 109 };
 110
 111 void pme_loadbal_init(pme_load_balancing_t *pme_lb_p,
 112                       const t_inputrec *ir, matrix box,
 113                       const interaction_const_t *ic,
 114                       gmx_pme_t pmedata)
 115 {
 116     pme_load_balancing_t pme_lb;
 117     real                 spm, sp;
 118     int                  d;
 119
 120     snew(pme_lb, 1);
 121
 122     /* Any number of stages >= 2 is supported */
 123     pme_lb->nstage   = 2;
 124
 125     pme_lb->cutoff_scheme = ir->cutoff_scheme;
 126
 127     if (pme_lb->cutoff_scheme == ecutsVERLET)
 128     {
 129         pme_lb->rbuf_coulomb = ic->rlist - ic->rcoulomb;
 130         pme_lb->rbuf_vdw     = pme_lb->rbuf_coulomb;
 131     }
 132     else
 133     {
 134         if (ic->rcoulomb > ic->rlist)
 135         {
 136             pme_lb->rbuf_coulomb = ic->rlistlong - ic->rcoulomb;
 137         }
 138         else
 139         {
 140             pme_lb->rbuf_coulomb = ic->rlist - ic->rcoulomb;
 141         }
 142         if (ic->rvdw > ic->rlist)
 143         {
 144             pme_lb->rbuf_vdw = ic->rlistlong - ic->rvdw;
 145         }
 146         else
 147         {
 148             pme_lb->rbuf_vdw = ic->rlist - ic->rvdw;
 149         }
 150     }
 151
 152     copy_mat(box, pme_lb->box_start);
 153     if (ir->ePBC == epbcXY && ir->nwall == 2)
 154     {
 155         svmul(ir->wall_ewald_zfac, pme_lb->box_start[ZZ], pme_lb->box_start[ZZ]);
 156     }
 157
 158     pme_lb->n = 1;
 159     snew(pme_lb->setup, pme_lb->n);
 160
 161     pme_lb->rcut_vdw              = ic->rvdw;
 162     pme_lb->rcut_coulomb_start    = ir->rcoulomb;
 163     pme_lb->nstcalclr_start       = ir->nstcalclr;
 164
 165     pme_lb->cur                   = 0;
 166     pme_lb->setup[0].rcut_coulomb = ic->rcoulomb;
 167     pme_lb->setup[0].rlist        = ic->rlist;
 168     pme_lb->setup[0].rlistlong    = ic->rlistlong;
 169     pme_lb->setup[0].nstcalclr    = ir->nstcalclr;
 170     pme_lb->setup[0].grid[XX]     = ir->nkx;
 171     pme_lb->setup[0].grid[YY]     = ir->nky;
 172     pme_lb->setup[0].grid[ZZ]     = ir->nkz;
 173     pme_lb->setup[0].ewaldcoeff   = ic->ewaldcoeff;
 174
 175     pme_lb->setup[0].pmedata  = pmedata;
 176
 177     spm = 0;
 178     for (d = 0; d < DIM; d++)
 179     {
 180         sp = norm(pme_lb->box_start[d])/pme_lb->setup[0].grid[d];
 181         if (sp > spm)
 182         {
 183             spm = sp;
 184         }
 185     }
 186     pme_lb->setup[0].spacing = spm;
 187
 188     if (ir->fourier_spacing > 0)
 189     {
 190         pme_lb->cut_spacing = ir->rcoulomb/ir->fourier_spacing;
 191     }
 192     else
 193     {
 194         pme_lb->cut_spacing = ir->rcoulomb/pme_lb->setup[0].spacing;
 195     }
 196
 197     pme_lb->stage = 0;
 198
 199     pme_lb->fastest  = 0;
 200     pme_lb->start    = 0;
 201     pme_lb->end      = 0;
 202     pme_lb->elimited = epmelblimNO;
 203
 204     *pme_lb_p = pme_lb;
 205 }
 206
 207 static gmx_bool pme_loadbal_increase_cutoff(pme_load_balancing_t pme_lb,
 208                                             int                  pme_order)
 209 {
 210     pme_setup_t *set;
 211     real         fac, sp;
 212     real         tmpr_coulomb, tmpr_vdw;
 213     int          d;
 214
 215     /* Try to add a new setup with next larger cut-off to the list */
 216     pme_lb->n++;
 217     srenew(pme_lb->setup, pme_lb->n);
 218     set          = &pme_lb->setup[pme_lb->n-1];
 219     set->pmedata = NULL;
 220
 221     fac = 1;
 222     do
 223     {
 224         fac *= 1.01;
 225         clear_ivec(set->grid);
 226         sp = calc_grid(NULL, pme_lb->box_start,
 227                        fac*pme_lb->setup[pme_lb->cur].spacing,
 228                        &set->grid[XX],
 229                        &set->grid[YY],
 230                        &set->grid[ZZ]);
 231
 232         /* In parallel we can't have grids smaller than 2*pme_order,
 233          * and we would anyhow not gain much speed at these grid sizes.
 234          */
 235         for (d = 0; d < DIM; d++)
 236         {
 237             if (set->grid[d] <= 2*pme_order)
 238             {
 239                 pme_lb->n--;
 240
 241                 return FALSE;
 242             }
 243         }
 244     }
 245     while (sp <= 1.001*pme_lb->setup[pme_lb->cur].spacing);
 246
 247     set->rcut_coulomb = pme_lb->cut_spacing*sp;
 248
 249     if (pme_lb->cutoff_scheme == ecutsVERLET)
 250     {
 251         set->rlist        = set->rcut_coulomb + pme_lb->rbuf_coulomb;
 252         /* We dont use LR lists with Verlet, but this avoids if-statements in further checks */
 253         set->rlistlong    = set->rlist;
 254     }
 255     else
 256     {
 257         tmpr_coulomb          = set->rcut_coulomb + pme_lb->rbuf_coulomb;
 258         tmpr_vdw              = pme_lb->rcut_vdw + pme_lb->rbuf_vdw;
 259         set->rlist            = min(tmpr_coulomb, tmpr_vdw);
 260         set->rlistlong        = max(tmpr_coulomb, tmpr_vdw);
 261
 262         /* Set the long-range update frequency */
 263         if (set->rlist == set->rlistlong)
 264         {
 265             /* No long-range interactions if the short-/long-range cutoffs are identical */
 266             set->nstcalclr = 0;
 267         }
 268         else if (pme_lb->nstcalclr_start == 0 || pme_lb->nstcalclr_start == 1)
 269         {
 270             /* We were not doing long-range before, but now we are since rlist!=rlistlong */
 271             set->nstcalclr = 1;
 272         }
 273         else
 274         {
 275             /* We were already doing long-range interactions from the start */
 276             if (pme_lb->rcut_vdw > pme_lb->rcut_coulomb_start)
 277             {
 278                 /* We were originally doing long-range VdW-only interactions.
 279                  * If rvdw is still longer than rcoulomb we keep the original nstcalclr,
 280                  * but if the coulomb cutoff has become longer we should update the long-range
 281                  * part every step.
 282                  */
 283                 set->nstcalclr = (tmpr_vdw > tmpr_coulomb) ? pme_lb->nstcalclr_start : 1;
 284             }
 285             else
 286             {
 287                 /* We were not doing any long-range interaction from the start,
 288                  * since it is not possible to do twin-range coulomb for the PME interaction.
 289                  */
 290                 set->nstcalclr = 1;
 291             }
 292         }
 293     }
 294
 295     set->spacing      = sp;
 296     /* The grid efficiency is the size wrt a grid with uniform x/y/z spacing */
 297     set->grid_efficiency = 1;
 298     for (d = 0; d < DIM; d++)
 299     {
 300         set->grid_efficiency *= (set->grid[d]*sp)/norm(pme_lb->box_start[d]);
 301     }
 302     /* The Ewald coefficient is inversly proportional to the cut-off */
 303     set->ewaldcoeff =
 304         pme_lb->setup[0].ewaldcoeff*pme_lb->setup[0].rcut_coulomb/set->rcut_coulomb;
 305
 306     set->count   = 0;
 307     set->cycles  = 0;
 308
 309     if (debug)
 310     {
 311         fprintf(debug, "PME loadbal: grid %d %d %d, coulomb cutoff %f\n",
 312                 set->grid[XX], set->grid[YY], set->grid[ZZ], set->rcut_coulomb);
 313     }
 314     return TRUE;
 315 }
 316
 317 static void print_grid(FILE *fp_err, FILE *fp_log,
 318                        const char *pre,
 319                        const char *desc,
 320                        const pme_setup_t *set,
 321                        double cycles)
 322 {
 323     char buf[STRLEN], buft[STRLEN];
 324
 325     if (cycles >= 0)
 326     {
 327         sprintf(buft, ": %.1f M-cycles", cycles*1e-6);
 328     }
 329     else
 330     {
 331         buft[0] = '\0';
 332     }
 333     sprintf(buf, "%-11s%10s pme grid %d %d %d, coulomb cutoff %.3f%s",
 334             pre,
 335             desc, set->grid[XX], set->grid[YY], set->grid[ZZ], set->rcut_coulomb,
 336             buft);
 337     if (fp_err != NULL)
 338     {
 339         fprintf(fp_err, "\r%s\n", buf);
 340     }
 341     if (fp_log != NULL)
 342     {
 343         fprintf(fp_log, "%s\n", buf);
 344     }
 345 }
 346
 347 static int pme_loadbal_end(pme_load_balancing_t pme_lb)
 348 {
 349     /* In the initial stage only n is set; end is not set yet */
 350     if (pme_lb->end > 0)
 351     {
 352         return pme_lb->end;
 353     }
 354     else
 355     {
 356         return pme_lb->n;
 357     }
 358 }
 359
 360 static void print_loadbal_limited(FILE *fp_err, FILE *fp_log,
 361                                   gmx_large_int_t step,
 362                                   pme_load_balancing_t pme_lb)
 363 {
 364     char buf[STRLEN], sbuf[22];
 365
 366     sprintf(buf, "step %4s: the %s limited the PME load balancing to a coulomb cut-off of %.3f",
 367             gmx_step_str(step, sbuf),
 368             pmelblim_str[pme_lb->elimited],
 369             pme_lb->setup[pme_loadbal_end(pme_lb)-1].rcut_coulomb);
 370     if (fp_err != NULL)
 371     {
 372         fprintf(fp_err, "\r%s\n", buf);
 373     }
 374     if (fp_log != NULL)
 375     {
 376         fprintf(fp_log, "%s\n", buf);
 377     }
 378 }
 379
 380 static void switch_to_stage1(pme_load_balancing_t pme_lb)
 381 {
 382     pme_lb->start = 0;
 383     while (pme_lb->start+1 < pme_lb->n &&
 384            (pme_lb->setup[pme_lb->start].count == 0 ||
 385             pme_lb->setup[pme_lb->start].cycles >
 386             pme_lb->setup[pme_lb->fastest].cycles*PME_LB_SLOW_FAC))
 387     {
 388         pme_lb->start++;
 389     }
 390     while (pme_lb->start > 0 && pme_lb->setup[pme_lb->start-1].cycles == 0)
 391     {
 392         pme_lb->start--;
 393     }
 394
 395     pme_lb->end = pme_lb->n;
 396     if (pme_lb->setup[pme_lb->end-1].count > 0 &&
 397         pme_lb->setup[pme_lb->end-1].cycles >
 398         pme_lb->setup[pme_lb->fastest].cycles*PME_LB_SLOW_FAC)
 399     {
 400         pme_lb->end--;
 401     }
 402
 403     pme_lb->stage = 1;
 404
 405     /* Next we want to choose setup pme_lb->start, but as we will increase
 406      * pme_ln->cur by one right after returning, we subtract 1 here.
 407      */
 408     pme_lb->cur = pme_lb->start - 1;
 409 }
 410
 411 gmx_bool pme_load_balance(pme_load_balancing_t pme_lb,
 412                           t_commrec           *cr,
 413                           FILE                *fp_err,
 414                           FILE                *fp_log,
 415                           t_inputrec          *ir,
 416                           t_state             *state,
 417                           double               cycles,
 418                           interaction_const_t *ic,
 419                           nonbonded_verlet_t  *nbv,
 420                           gmx_pme_t           *pmedata,
 421                           gmx_large_int_t      step)
 422 {
 423     gmx_bool     OK;
 424     pme_setup_t *set;
 425     double       cycles_fast;
 426     char         buf[STRLEN], sbuf[22];
 427     real         rtab;
 428     gmx_bool     bUsesSimpleTables = TRUE;
 429
 430     if (pme_lb->stage == pme_lb->nstage)
 431     {
 432         return FALSE;
 433     }
 434
 435     if (PAR(cr))
 436     {
 437         gmx_sumd(1, &cycles, cr);
 438         cycles /= cr->nnodes;
 439     }
 440
 441     set = &pme_lb->setup[pme_lb->cur];
 442     set->count++;
 443
 444     rtab = ir->rlistlong + ir->tabext;
 445
 446     if (set->count % 2 == 1)
 447     {
 448         /* Skip the first cycle, because the first step after a switch
 449          * is much slower due to allocation and/or caching effects.
 450          */
 451         return TRUE;
 452     }
 453
 454     sprintf(buf, "step %4s: ", gmx_step_str(step, sbuf));
 455     print_grid(fp_err, fp_log, buf, "timed with", set, cycles);
 456
 457     if (set->count <= 2)
 458     {
 459         set->cycles = cycles;
 460     }
 461     else
 462     {
 463         if (cycles*PME_LB_ACCEL_TOL < set->cycles &&
 464             pme_lb->stage == pme_lb->nstage - 1)
 465         {
 466             /* The performance went up a lot (due to e.g. DD load balancing).
 467              * Add a stage, keep the minima, but rescan all setups.
 468              */
 469             pme_lb->nstage++;
 470
 471             if (debug)
 472             {
 473                 fprintf(debug, "The performance for grid %d %d %d went from %.3f to %.1f M-cycles, this is more than %f\n"
 474                         "Increased the number stages to %d"
 475                         " and ignoring the previous performance\n",
 476                         set->grid[XX], set->grid[YY], set->grid[ZZ],
 477                         cycles*1e-6, set->cycles*1e-6, PME_LB_ACCEL_TOL,
 478                         pme_lb->nstage);
 479             }
 480         }
 481         set->cycles = min(set->cycles, cycles);
 482     }
 483
 484     if (set->cycles < pme_lb->setup[pme_lb->fastest].cycles)
 485     {
 486         pme_lb->fastest = pme_lb->cur;
 487
 488         if (DOMAINDECOMP(cr))
 489         {
 490             /* We found a new fastest setting, ensure that with subsequent
 491              * shorter cut-off's the dynamic load balancing does not make
 492              * the use of the current cut-off impossible. This solution is
 493              * a trade-off, as the PME load balancing and DD domain size
 494              * load balancing can interact in complex ways.
 495              * With the Verlet kernels, DD load imbalance will usually be
 496              * mainly due to bonded interaction imbalance, which will often
 497              * quickly push the domain boundaries beyond the limit for the
 498              * optimal, PME load balanced, cut-off. But it could be that
 499              * better overal performance can be obtained with a slightly
 500              * shorter cut-off and better DD load balancing.
 501              */
 502             change_dd_dlb_cutoff_limit(cr);
 503         }
 504     }
 505     cycles_fast = pme_lb->setup[pme_lb->fastest].cycles;
 506
 507     /* Check in stage 0 if we should stop scanning grids.
 508      * Stop when the time is more than SLOW_FAC longer than the fastest.
 509      */
 510     if (pme_lb->stage == 0 && pme_lb->cur > 0 &&
 511         cycles > pme_lb->setup[pme_lb->fastest].cycles*PME_LB_SLOW_FAC)
 512     {
 513         pme_lb->n = pme_lb->cur + 1;
 514         /* Done with scanning, go to stage 1 */
 515         switch_to_stage1(pme_lb);
 516     }
 517
 518     if (pme_lb->stage == 0)
 519     {
 520         int gridsize_start;
 521
 522         gridsize_start = set->grid[XX]*set->grid[YY]*set->grid[ZZ];
 523
 524         do
 525         {
 526             if (pme_lb->cur+1 < pme_lb->n)
 527             {
 528                 /* We had already generated the next setup */
 529                 OK = TRUE;
 530             }
 531             else
 532             {
 533                 /* Find the next setup */
 534                 OK = pme_loadbal_increase_cutoff(pme_lb, ir->pme_order);
 535             }
 536
 537             if (OK && ir->ePBC != epbcNONE)
 538             {
 539                 OK = (sqr(pme_lb->setup[pme_lb->cur+1].rlistlong)
 540                       <= max_cutoff2(ir->ePBC, state->box));
 541                 if (!OK)
 542                 {
 543                     pme_lb->elimited = epmelblimBOX;
 544                 }
 545             }
 546
 547             if (OK)
 548             {
 549                 pme_lb->cur++;
 550
 551                 if (DOMAINDECOMP(cr))
 552                 {
 553                     OK = change_dd_cutoff(cr, state, ir,
 554                                           pme_lb->setup[pme_lb->cur].rlistlong);
 555                     if (!OK)
 556                     {
 557                         /* Failed: do not use this setup */
 558                         pme_lb->cur--;
 559                         pme_lb->elimited = epmelblimDD;
 560                     }
 561                 }
 562             }
 563             if (!OK)
 564             {
 565                 /* We hit the upper limit for the cut-off,
 566                  * the setup should not go further than cur.
 567                  */
 568                 pme_lb->n = pme_lb->cur + 1;
 569                 print_loadbal_limited(fp_err, fp_log, step, pme_lb);
 570                 /* Switch to the next stage */
 571                 switch_to_stage1(pme_lb);
 572             }
 573         }
 574         while (OK &&
 575                !(pme_lb->setup[pme_lb->cur].grid[XX]*
 576                  pme_lb->setup[pme_lb->cur].grid[YY]*
 577                  pme_lb->setup[pme_lb->cur].grid[ZZ] <
 578                  gridsize_start*PME_LB_GRID_SCALE_FAC
 579                  &&
 580                  pme_lb->setup[pme_lb->cur].grid_efficiency <
 581                  pme_lb->setup[pme_lb->cur-1].grid_efficiency*PME_LB_GRID_EFFICIENCY_REL_FAC));
 582     }
 583
 584     if (pme_lb->stage > 0 && pme_lb->end == 1)
 585     {
 586         pme_lb->cur   = 0;
 587         pme_lb->stage = pme_lb->nstage;
 588     }
 589     else if (pme_lb->stage > 0 && pme_lb->end > 1)
 590     {
 591         /* If stage = nstage-1:
 592          *   scan over all setups, rerunning only those setups
 593          *   which are not much slower than the fastest
 594          * else:
 595          *   use the next setup
 596          */
 597         do
 598         {
 599             pme_lb->cur++;
 600             if (pme_lb->cur == pme_lb->end)
 601             {
 602                 pme_lb->stage++;
 603                 pme_lb->cur = pme_lb->start;
 604             }
 605         }
 606         while (pme_lb->stage == pme_lb->nstage - 1 &&
 607                pme_lb->setup[pme_lb->cur].count > 0 &&
 608                pme_lb->setup[pme_lb->cur].cycles > cycles_fast*PME_LB_SLOW_FAC);
 609
 610         if (pme_lb->stage == pme_lb->nstage)
 611         {
 612             /* We are done optimizing, use the fastest setup we found */
 613             pme_lb->cur = pme_lb->fastest;
 614         }
 615     }
 616
 617     if (DOMAINDECOMP(cr) && pme_lb->stage > 0)
 618     {
 619         OK = change_dd_cutoff(cr, state, ir, pme_lb->setup[pme_lb->cur].rlistlong);
 620         if (!OK)
 621         {
 622             /* Failsafe solution */
 623             if (pme_lb->cur > 1 && pme_lb->stage == pme_lb->nstage)
 624             {
 625                 pme_lb->stage--;
 626             }
 627             pme_lb->fastest  = 0;
 628             pme_lb->start    = 0;
 629             pme_lb->end      = pme_lb->cur;
 630             pme_lb->cur      = pme_lb->start;
 631             pme_lb->elimited = epmelblimDD;
 632             print_loadbal_limited(fp_err, fp_log, step, pme_lb);
 633         }
 634     }
 635
 636     /* Change the Coulomb cut-off and the PME grid */
 637
 638     set = &pme_lb->setup[pme_lb->cur];
 639
 640     ic->rcoulomb   = set->rcut_coulomb;
 641     ic->rlist      = set->rlist;
 642     ic->rlistlong  = set->rlistlong;
 643     ir->nstcalclr  = set->nstcalclr;
 644     ic->ewaldcoeff = set->ewaldcoeff;
 645
 646     bUsesSimpleTables = uses_simple_tables(ir->cutoff_scheme, nbv, 0);
 647     if (pme_lb->cutoff_scheme == ecutsVERLET &&
 648         nbv->grp[0].kernel_type == nbnxnk8x8x8_CUDA)
 649     {
 650         nbnxn_cuda_pme_loadbal_update_param(nbv->cu_nbv, ic);
 651     }
 652     else
 653     {
 654         init_interaction_const_tables(NULL, ic, bUsesSimpleTables,
 655                                       rtab);
 656     }
 657
 658     if (pme_lb->cutoff_scheme == ecutsVERLET && nbv->ngrp > 1)
 659     {
 660         init_interaction_const_tables(NULL, ic, bUsesSimpleTables,
 661                                       rtab);
 662     }
 663
 664     if (cr->duty & DUTY_PME)
 665     {
 666         if (pme_lb->setup[pme_lb->cur].pmedata == NULL)
 667         {
 668             /* Generate a new PME data structure,
 669              * copying part of the old pointers.
 670              */
 671             gmx_pme_reinit(&set->pmedata,
 672                            cr, pme_lb->setup[0].pmedata, ir,
 673                            set->grid);
 674         }
 675         *pmedata = set->pmedata;
 676     }
 677     else
 678     {
 679         /* Tell our PME-only node to switch grid */
 680         gmx_pme_send_switchgrid(cr, set->grid, set->ewaldcoeff);
 681     }
 682
 683     if (debug)
 684     {
 685         print_grid(NULL, debug, "", "switched to", set, -1);
 686     }
 687
 688     if (pme_lb->stage == pme_lb->nstage)
 689     {
 690         print_grid(fp_err, fp_log, "", "optimal", set, -1);
 691     }
 692
 693     return TRUE;
 694 }
 695
 696 void restart_pme_loadbal(pme_load_balancing_t pme_lb, int n)
 697 {
 698     pme_lb->nstage += n;
 699 }
 700
 701 static int pme_grid_points(const pme_setup_t *setup)
 702 {
 703     return setup->grid[XX]*setup->grid[YY]*setup->grid[ZZ];
 704 }
 705
 706 static real pme_loadbal_rlist(const pme_setup_t *setup)
 707 {
 708     /* With the group cut-off scheme we can have twin-range either
 709      * for Coulomb or for VdW, so we need a check here.
 710      * With the Verlet cut-off scheme rlist=rlistlong.
 711      */
 712     if (setup->rcut_coulomb > setup->rlist)
 713     {
 714         return setup->rlistlong;
 715     }
 716     else
 717     {
 718         return setup->rlist;
 719     }
 720 }
 721
 722 static void print_pme_loadbal_setting(FILE              *fplog,
 723                                       char              *name,
 724                                       const pme_setup_t *setup)
 725 {
 726     fprintf(fplog,
 727             "   %-7s %6.3f nm %6.3f nm     %3d %3d %3d   %5.3f nm  %5.3f nm\n",
 728             name,
 729             setup->rcut_coulomb, pme_loadbal_rlist(setup),
 730             setup->grid[XX], setup->grid[YY], setup->grid[ZZ],
 731             setup->spacing, 1/setup->ewaldcoeff);
 732 }
 733
 734 static void print_pme_loadbal_settings(pme_load_balancing_t pme_lb,
 735                                        t_commrec           *cr,
 736                                        FILE                *fplog,
 737                                        gmx_bool             bNonBondedOnGPU)
 738 {
 739     double pp_ratio, grid_ratio;
 740
 741     pp_ratio   = pow(pme_loadbal_rlist(&pme_lb->setup[pme_lb->cur])/pme_loadbal_rlist(&pme_lb->setup[0]), 3.0);
 742     grid_ratio = pme_grid_points(&pme_lb->setup[pme_lb->cur])/
 743         (double)pme_grid_points(&pme_lb->setup[0]);
 744
 745     fprintf(fplog, "\n");
 746     fprintf(fplog, "       P P   -   P M E   L O A D   B A L A N C I N G\n");
 747     fprintf(fplog, "\n");
 748     /* Here we only warn when the optimal setting is the last one */
 749     if (pme_lb->elimited != epmelblimNO &&
 750         pme_lb->cur == pme_loadbal_end(pme_lb)-1)
 751     {
 752         fprintf(fplog, " NOTE: The PP/PME load balancing was limited by the %s,\n",
 753                 pmelblim_str[pme_lb->elimited]);
 754         fprintf(fplog, "       you might not have reached a good load balance.\n");
 755         if (pme_lb->elimited == epmelblimDD)
 756         {
 757             fprintf(fplog, "       Try different mdrun -dd settings or lower the -dds value.\n");
 758         }
 759         fprintf(fplog, "\n");
 760     }
 761     fprintf(fplog, " PP/PME load balancing changed the cut-off and PME settings:\n");
 762     fprintf(fplog, "           particle-particle                    PME\n");
 763     fprintf(fplog, "            rcoulomb  rlist            grid      spacing   1/beta\n");
 764     print_pme_loadbal_setting(fplog, "initial", &pme_lb->setup[0]);
 765     print_pme_loadbal_setting(fplog, "final", &pme_lb->setup[pme_lb->cur]);
 766     fprintf(fplog, " cost-ratio           %4.2f             %4.2f\n",
 767             pp_ratio, grid_ratio);
 768     fprintf(fplog, " (note that these numbers concern only part of the total PP and PME load)\n");
 769
 770     if (pp_ratio > 1.5 && !bNonBondedOnGPU)
 771     {
 772         md_print_warn(cr, fplog,
 773                       "NOTE: PME load balancing increased the non-bonded workload by more than 50%%.\n"
 774                       "      For better performance use (more) PME nodes (mdrun -npme),\n"
 775                       "      or in case you are beyond the scaling limit, use less nodes in total.\n");
 776     }
 777     else
 778     {
 779         fprintf(fplog, "\n");
 780     }
 781 }
 782
 783 void pme_loadbal_done(pme_load_balancing_t pme_lb,
 784                       t_commrec *cr, FILE *fplog,
 785                       gmx_bool bNonBondedOnGPU)
 786 {
 787     if (fplog != NULL && (pme_lb->cur > 0 || pme_lb->elimited != epmelblimNO))
 788     {
 789         print_pme_loadbal_settings(pme_lb, cr, fplog, bNonBondedOnGPU);
 790     }
 791
 792     /* TODO: Here we should free all pointers in pme_lb,
 793      * but as it contains pme data structures,
 794      * we need to first make pme.c free all data.
 795      */
 796 }