src/programs/mdrun/pme_loadbal.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 #ifdef HAVE_CONFIG_H
  36 #include <config.h>
  37 #endif
  38
  39 #include "gromacs/utility/smalloc.h"
  40 #include "types/commrec.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_q;    /* Electrostatic Ewald coefficient            */
  62     real      ewaldcoeff_lj;   /* LJ Ewald coefficient, only for the call to send_switchgrid */
  63     gmx_pme_t pmedata;         /* the data structure used in the PME code      */
  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, epmelblimPMEGRID, epmelblimNR
  83 };
  84
  85 const char *pmelblim_str[epmelblimNR] =
  86 { "no", "box size", "domain decompostion", "PME grid restriction" };
  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_q    = ic->ewaldcoeff_q;
 173     pme_lb->setup[0].ewaldcoeff_lj   = ic->ewaldcoeff_lj;
 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                                             const gmx_domdec_t   *dd)
 210 {
 211     pme_setup_t *set;
 212     int          npmenodes_x, npmenodes_y;
 213     real         fac, sp;
 214     real         tmpr_coulomb, tmpr_vdw;
 215     int          d;
 216     gmx_bool     grid_ok;
 217
 218     /* Try to add a new setup with next larger cut-off to the list */
 219     pme_lb->n++;
 220     srenew(pme_lb->setup, pme_lb->n);
 221     set          = &pme_lb->setup[pme_lb->n-1];
 222     set->pmedata = NULL;
 223
 224     get_pme_nnodes(dd, &npmenodes_x, &npmenodes_y);
 225
 226     fac = 1;
 227     do
 228     {
 229         /* Avoid infinite while loop, which can occur at the minimum grid size.
 230          * Note that in practice load balancing will stop before this point.
 231          * The factor 2.1 allows for the extreme case in which only grids
 232          * of powers of 2 are allowed (the current code supports more grids).
 233          */
 234         if (fac > 2.1)
 235         {
 236             pme_lb->n--;
 237
 238             return FALSE;
 239         }
 240
 241         fac *= 1.01;
 242         clear_ivec(set->grid);
 243         sp = calc_grid(NULL, pme_lb->box_start,
 244                        fac*pme_lb->setup[pme_lb->cur].spacing,
 245                        &set->grid[XX],
 246                        &set->grid[YY],
 247                        &set->grid[ZZ]);
 248
 249         /* As here we can't easily check if one of the PME nodes
 250          * uses threading, we do a conservative grid check.
 251          * This means we can't use pme_order or less grid lines
 252          * per PME node along x, which is not a strong restriction.
 253          */
 254         gmx_pme_check_restrictions(pme_order,
 255                                    set->grid[XX], set->grid[YY], set->grid[ZZ],
 256                                    npmenodes_x, npmenodes_y,
 257                                    TRUE,
 258                                    FALSE,
 259                                    &grid_ok);
 260     }
 261     while (sp <= 1.001*pme_lb->setup[pme_lb->cur].spacing || !grid_ok);
 262
 263     set->rcut_coulomb = pme_lb->cut_spacing*sp;
 264     if (set->rcut_coulomb < pme_lb->rcut_coulomb_start)
 265     {
 266         /* This is unlikely, but can happen when e.g. continuing from
 267          * a checkpoint after equilibration where the box shrank a lot.
 268          * We want to avoid rcoulomb getting smaller than rvdw
 269          * and there might be more issues with decreasing rcoulomb.
 270          */
 271         set->rcut_coulomb = pme_lb->rcut_coulomb_start;
 272     }
 273
 274     if (pme_lb->cutoff_scheme == ecutsVERLET)
 275     {
 276         set->rlist        = set->rcut_coulomb + pme_lb->rbuf_coulomb;
 277         /* We dont use LR lists with Verlet, but this avoids if-statements in further checks */
 278         set->rlistlong    = set->rlist;
 279     }
 280     else
 281     {
 282         tmpr_coulomb          = set->rcut_coulomb + pme_lb->rbuf_coulomb;
 283         tmpr_vdw              = pme_lb->rcut_vdw + pme_lb->rbuf_vdw;
 284         set->rlist            = min(tmpr_coulomb, tmpr_vdw);
 285         set->rlistlong        = max(tmpr_coulomb, tmpr_vdw);
 286
 287         /* Set the long-range update frequency */
 288         if (set->rlist == set->rlistlong)
 289         {
 290             /* No long-range interactions if the short-/long-range cutoffs are identical */
 291             set->nstcalclr = 0;
 292         }
 293         else if (pme_lb->nstcalclr_start == 0 || pme_lb->nstcalclr_start == 1)
 294         {
 295             /* We were not doing long-range before, but now we are since rlist!=rlistlong */
 296             set->nstcalclr = 1;
 297         }
 298         else
 299         {
 300             /* We were already doing long-range interactions from the start */
 301             if (pme_lb->rcut_vdw > pme_lb->rcut_coulomb_start)
 302             {
 303                 /* We were originally doing long-range VdW-only interactions.
 304                  * If rvdw is still longer than rcoulomb we keep the original nstcalclr,
 305                  * but if the coulomb cutoff has become longer we should update the long-range
 306                  * part every step.
 307                  */
 308                 set->nstcalclr = (tmpr_vdw > tmpr_coulomb) ? pme_lb->nstcalclr_start : 1;
 309             }
 310             else
 311             {
 312                 /* We were not doing any long-range interaction from the start,
 313                  * since it is not possible to do twin-range coulomb for the PME interaction.
 314                  */
 315                 set->nstcalclr = 1;
 316             }
 317         }
 318     }
 319
 320     set->spacing      = sp;
 321     /* The grid efficiency is the size wrt a grid with uniform x/y/z spacing */
 322     set->grid_efficiency = 1;
 323     for (d = 0; d < DIM; d++)
 324     {
 325         set->grid_efficiency *= (set->grid[d]*sp)/norm(pme_lb->box_start[d]);
 326     }
 327     /* The Ewald coefficient is inversly proportional to the cut-off */
 328     set->ewaldcoeff_q =
 329         pme_lb->setup[0].ewaldcoeff_q*pme_lb->setup[0].rcut_coulomb/set->rcut_coulomb;
 330     /* We set ewaldcoeff_lj in set, even when LJ-PME is not used */
 331     set->ewaldcoeff_lj =
 332         pme_lb->setup[0].ewaldcoeff_lj*pme_lb->setup[0].rcut_coulomb/set->rcut_coulomb;
 333
 334     set->count   = 0;
 335     set->cycles  = 0;
 336
 337     if (debug)
 338     {
 339         fprintf(debug, "PME loadbal: grid %d %d %d, coulomb cutoff %f\n",
 340                 set->grid[XX], set->grid[YY], set->grid[ZZ], set->rcut_coulomb);
 341     }
 342     return TRUE;
 343 }
 344
 345 static void print_grid(FILE *fp_err, FILE *fp_log,
 346                        const char *pre,
 347                        const char *desc,
 348                        const pme_setup_t *set,
 349                        double cycles)
 350 {
 351     char buf[STRLEN], buft[STRLEN];
 352
 353     if (cycles >= 0)
 354     {
 355         sprintf(buft, ": %.1f M-cycles", cycles*1e-6);
 356     }
 357     else
 358     {
 359         buft[0] = '\0';
 360     }
 361     sprintf(buf, "%-11s%10s pme grid %d %d %d, coulomb cutoff %.3f%s",
 362             pre,
 363             desc, set->grid[XX], set->grid[YY], set->grid[ZZ], set->rcut_coulomb,
 364             buft);
 365     if (fp_err != NULL)
 366     {
 367         fprintf(fp_err, "\r%s\n", buf);
 368     }
 369     if (fp_log != NULL)
 370     {
 371         fprintf(fp_log, "%s\n", buf);
 372     }
 373 }
 374
 375 static int pme_loadbal_end(pme_load_balancing_t pme_lb)
 376 {
 377     /* In the initial stage only n is set; end is not set yet */
 378     if (pme_lb->end > 0)
 379     {
 380         return pme_lb->end;
 381     }
 382     else
 383     {
 384         return pme_lb->n;
 385     }
 386 }
 387
 388 static void print_loadbal_limited(FILE *fp_err, FILE *fp_log,
 389                                   gmx_int64_t step,
 390                                   pme_load_balancing_t pme_lb)
 391 {
 392     char buf[STRLEN], sbuf[22];
 393
 394     sprintf(buf, "step %4s: the %s limits the PME load balancing to a coulomb cut-off of %.3f",
 395             gmx_step_str(step, sbuf),
 396             pmelblim_str[pme_lb->elimited],
 397             pme_lb->setup[pme_loadbal_end(pme_lb)-1].rcut_coulomb);
 398     if (fp_err != NULL)
 399     {
 400         fprintf(fp_err, "\r%s\n", buf);
 401     }
 402     if (fp_log != NULL)
 403     {
 404         fprintf(fp_log, "%s\n", buf);
 405     }
 406 }
 407
 408 static void switch_to_stage1(pme_load_balancing_t pme_lb)
 409 {
 410     pme_lb->start = 0;
 411     while (pme_lb->start+1 < pme_lb->n &&
 412            (pme_lb->setup[pme_lb->start].count == 0 ||
 413             pme_lb->setup[pme_lb->start].cycles >
 414             pme_lb->setup[pme_lb->fastest].cycles*PME_LB_SLOW_FAC))
 415     {
 416         pme_lb->start++;
 417     }
 418     while (pme_lb->start > 0 && pme_lb->setup[pme_lb->start-1].cycles == 0)
 419     {
 420         pme_lb->start--;
 421     }
 422
 423     pme_lb->end = pme_lb->n;
 424     if (pme_lb->setup[pme_lb->end-1].count > 0 &&
 425         pme_lb->setup[pme_lb->end-1].cycles >
 426         pme_lb->setup[pme_lb->fastest].cycles*PME_LB_SLOW_FAC)
 427     {
 428         pme_lb->end--;
 429     }
 430
 431     pme_lb->stage = 1;
 432
 433     /* Next we want to choose setup pme_lb->start, but as we will increase
 434      * pme_ln->cur by one right after returning, we subtract 1 here.
 435      */
 436     pme_lb->cur = pme_lb->start - 1;
 437 }
 438
 439 gmx_bool pme_load_balance(pme_load_balancing_t pme_lb,
 440                           t_commrec           *cr,
 441                           FILE                *fp_err,
 442                           FILE                *fp_log,
 443                           t_inputrec          *ir,
 444                           t_state             *state,
 445                           double               cycles,
 446                           interaction_const_t *ic,
 447                           nonbonded_verlet_t  *nbv,
 448                           gmx_pme_t           *pmedata,
 449                           gmx_int64_t          step)
 450 {
 451     gmx_bool     OK;
 452     pme_setup_t *set;
 453     double       cycles_fast;
 454     char         buf[STRLEN], sbuf[22];
 455     real         rtab;
 456     gmx_bool     bUsesSimpleTables = TRUE;
 457
 458     if (pme_lb->stage == pme_lb->nstage)
 459     {
 460         return FALSE;
 461     }
 462
 463     if (PAR(cr))
 464     {
 465         gmx_sumd(1, &cycles, cr);
 466         cycles /= cr->nnodes;
 467     }
 468
 469     set = &pme_lb->setup[pme_lb->cur];
 470     set->count++;
 471
 472     rtab = ir->rlistlong + ir->tabext;
 473
 474     if (set->count % 2 == 1)
 475     {
 476         /* Skip the first cycle, because the first step after a switch
 477          * is much slower due to allocation and/or caching effects.
 478          */
 479         return TRUE;
 480     }
 481
 482     sprintf(buf, "step %4s: ", gmx_step_str(step, sbuf));
 483     print_grid(fp_err, fp_log, buf, "timed with", set, cycles);
 484
 485     if (set->count <= 2)
 486     {
 487         set->cycles = cycles;
 488     }
 489     else
 490     {
 491         if (cycles*PME_LB_ACCEL_TOL < set->cycles &&
 492             pme_lb->stage == pme_lb->nstage - 1)
 493         {
 494             /* The performance went up a lot (due to e.g. DD load balancing).
 495              * Add a stage, keep the minima, but rescan all setups.
 496              */
 497             pme_lb->nstage++;
 498
 499             if (debug)
 500             {
 501                 fprintf(debug, "The performance for grid %d %d %d went from %.3f to %.1f M-cycles, this is more than %f\n"
 502                         "Increased the number stages to %d"
 503                         " and ignoring the previous performance\n",
 504                         set->grid[XX], set->grid[YY], set->grid[ZZ],
 505                         cycles*1e-6, set->cycles*1e-6, PME_LB_ACCEL_TOL,
 506                         pme_lb->nstage);
 507             }
 508         }
 509         set->cycles = min(set->cycles, cycles);
 510     }
 511
 512     if (set->cycles < pme_lb->setup[pme_lb->fastest].cycles)
 513     {
 514         pme_lb->fastest = pme_lb->cur;
 515
 516         if (DOMAINDECOMP(cr))
 517         {
 518             /* We found a new fastest setting, ensure that with subsequent
 519              * shorter cut-off's the dynamic load balancing does not make
 520              * the use of the current cut-off impossible. This solution is
 521              * a trade-off, as the PME load balancing and DD domain size
 522              * load balancing can interact in complex ways.
 523              * With the Verlet kernels, DD load imbalance will usually be
 524              * mainly due to bonded interaction imbalance, which will often
 525              * quickly push the domain boundaries beyond the limit for the
 526              * optimal, PME load balanced, cut-off. But it could be that
 527              * better overal performance can be obtained with a slightly
 528              * shorter cut-off and better DD load balancing.
 529              */
 530             change_dd_dlb_cutoff_limit(cr);
 531         }
 532     }
 533     cycles_fast = pme_lb->setup[pme_lb->fastest].cycles;
 534
 535     /* Check in stage 0 if we should stop scanning grids.
 536      * Stop when the time is more than SLOW_FAC longer than the fastest.
 537      */
 538     if (pme_lb->stage == 0 && pme_lb->cur > 0 &&
 539         cycles > pme_lb->setup[pme_lb->fastest].cycles*PME_LB_SLOW_FAC)
 540     {
 541         pme_lb->n = pme_lb->cur + 1;
 542         /* Done with scanning, go to stage 1 */
 543         switch_to_stage1(pme_lb);
 544     }
 545
 546     if (pme_lb->stage == 0)
 547     {
 548         int gridsize_start;
 549
 550         gridsize_start = set->grid[XX]*set->grid[YY]*set->grid[ZZ];
 551
 552         do
 553         {
 554             if (pme_lb->cur+1 < pme_lb->n)
 555             {
 556                 /* We had already generated the next setup */
 557                 OK = TRUE;
 558             }
 559             else
 560             {
 561                 /* Find the next setup */
 562                 OK = pme_loadbal_increase_cutoff(pme_lb, ir->pme_order, cr->dd);
 563
 564                 if (!OK)
 565                 {
 566                     pme_lb->elimited = epmelblimPMEGRID;
 567                 }
 568             }
 569
 570             if (OK && ir->ePBC != epbcNONE)
 571             {
 572                 OK = (sqr(pme_lb->setup[pme_lb->cur+1].rlistlong)
 573                       <= max_cutoff2(ir->ePBC, state->box));
 574                 if (!OK)
 575                 {
 576                     pme_lb->elimited = epmelblimBOX;
 577                 }
 578             }
 579
 580             if (OK)
 581             {
 582                 pme_lb->cur++;
 583
 584                 if (DOMAINDECOMP(cr))
 585                 {
 586                     OK = change_dd_cutoff(cr, state, ir,
 587                                           pme_lb->setup[pme_lb->cur].rlistlong);
 588                     if (!OK)
 589                     {
 590                         /* Failed: do not use this setup */
 591                         pme_lb->cur--;
 592                         pme_lb->elimited = epmelblimDD;
 593                     }
 594                 }
 595             }
 596             if (!OK)
 597             {
 598                 /* We hit the upper limit for the cut-off,
 599                  * the setup should not go further than cur.
 600                  */
 601                 pme_lb->n = pme_lb->cur + 1;
 602                 print_loadbal_limited(fp_err, fp_log, step, pme_lb);
 603                 /* Switch to the next stage */
 604                 switch_to_stage1(pme_lb);
 605             }
 606         }
 607         while (OK &&
 608                !(pme_lb->setup[pme_lb->cur].grid[XX]*
 609                  pme_lb->setup[pme_lb->cur].grid[YY]*
 610                  pme_lb->setup[pme_lb->cur].grid[ZZ] <
 611                  gridsize_start*PME_LB_GRID_SCALE_FAC
 612                  &&
 613                  pme_lb->setup[pme_lb->cur].grid_efficiency <
 614                  pme_lb->setup[pme_lb->cur-1].grid_efficiency*PME_LB_GRID_EFFICIENCY_REL_FAC));
 615     }
 616
 617     if (pme_lb->stage > 0 && pme_lb->end == 1)
 618     {
 619         pme_lb->cur   = 0;
 620         pme_lb->stage = pme_lb->nstage;
 621     }
 622     else if (pme_lb->stage > 0 && pme_lb->end > 1)
 623     {
 624         /* If stage = nstage-1:
 625          *   scan over all setups, rerunning only those setups
 626          *   which are not much slower than the fastest
 627          * else:
 628          *   use the next setup
 629          */
 630         do
 631         {
 632             pme_lb->cur++;
 633             if (pme_lb->cur == pme_lb->end)
 634             {
 635                 pme_lb->stage++;
 636                 pme_lb->cur = pme_lb->start;
 637             }
 638         }
 639         while (pme_lb->stage == pme_lb->nstage - 1 &&
 640                pme_lb->setup[pme_lb->cur].count > 0 &&
 641                pme_lb->setup[pme_lb->cur].cycles > cycles_fast*PME_LB_SLOW_FAC);
 642
 643         if (pme_lb->stage == pme_lb->nstage)
 644         {
 645             /* We are done optimizing, use the fastest setup we found */
 646             pme_lb->cur = pme_lb->fastest;
 647         }
 648     }
 649
 650     if (DOMAINDECOMP(cr) && pme_lb->stage > 0)
 651     {
 652         OK = change_dd_cutoff(cr, state, ir, pme_lb->setup[pme_lb->cur].rlistlong);
 653         if (!OK)
 654         {
 655             /* Failsafe solution */
 656             if (pme_lb->cur > 1 && pme_lb->stage == pme_lb->nstage)
 657             {
 658                 pme_lb->stage--;
 659             }
 660             pme_lb->fastest  = 0;
 661             pme_lb->start    = 0;
 662             pme_lb->end      = pme_lb->cur;
 663             pme_lb->cur      = pme_lb->start;
 664             pme_lb->elimited = epmelblimDD;
 665             print_loadbal_limited(fp_err, fp_log, step, pme_lb);
 666         }
 667     }
 668
 669     /* Change the Coulomb cut-off and the PME grid */
 670
 671     set = &pme_lb->setup[pme_lb->cur];
 672
 673     ic->rcoulomb     = set->rcut_coulomb;
 674     ic->rlist        = set->rlist;
 675     ic->rlistlong    = set->rlistlong;
 676     ir->nstcalclr    = set->nstcalclr;
 677     ic->ewaldcoeff_q = set->ewaldcoeff_q;
 678     /* TODO: centralize the code that sets the potentials shifts */
 679     if (ic->coulomb_modifier == eintmodPOTSHIFT)
 680     {
 681         ic->sh_ewald = gmx_erfc(ic->ewaldcoeff_q*ic->rcoulomb);
 682     }
 683     if (EVDW_PME(ic->vdwtype))
 684     {
 685         /* We have PME for both Coulomb and VdW, set rvdw equal to rcoulomb */
 686         ic->rvdw            = set->rcut_coulomb;
 687         ic->ewaldcoeff_lj   = set->ewaldcoeff_lj;
 688         if (ic->vdw_modifier == eintmodPOTSHIFT)
 689         {
 690             real crc2;
 691
 692             ic->dispersion_shift.cpot = -pow(ic->rvdw, -6.0);
 693             ic->repulsion_shift.cpot  = -pow(ic->rvdw, -12.0);
 694             ic->sh_invrc6             = -ic->dispersion_shift.cpot;
 695             crc2                      = sqr(ic->ewaldcoeff_lj*ic->rvdw);
 696             ic->sh_lj_ewald           = (exp(-crc2)*(1 + crc2 + 0.5*crc2*crc2) - 1)*pow(ic->rvdw, -6.0);
 697         }
 698     }
 699
 700     bUsesSimpleTables = uses_simple_tables(ir->cutoff_scheme, nbv, 0);
 701     if (pme_lb->cutoff_scheme == ecutsVERLET &&
 702         nbv->grp[0].kernel_type == nbnxnk8x8x8_CUDA)
 703     {
 704         nbnxn_cuda_pme_loadbal_update_param(nbv->cu_nbv, ic);
 705
 706         /* With tMPI + GPUs some ranks may be sharing GPU(s) and therefore
 707          * also sharing texture references. To keep the code simple, we don't
 708          * treat texture references as shared resources, but this means that
 709          * the coulomb_tab texture ref will get updated by multiple threads.
 710          * Hence, to ensure that the non-bonded kernels don't start before all
 711          * texture binding operations are finished, we need to wait for all ranks
 712          * to arrive here before continuing.
 713          *
 714          * Note that we could omit this barrier if GPUs are not shared (or
 715          * texture objects are used), but as this is initialization code, there
 716          * is not point in complicating things.
 717          */
 718 #ifdef GMX_THREAD_MPI
 719         if (PAR(cr))
 720         {
 721             gmx_barrier(cr);
 722         }
 723 #endif  /* GMX_THREAD_MPI */
 724     }
 725
 726     /* Usually we won't need the simple tables with GPUs.
 727      * But we do with hybrid acceleration and with free energy.
 728      * To avoid bugs, we always re-initialize the simple tables here.
 729      */
 730     init_interaction_const_tables(NULL, ic, bUsesSimpleTables, rtab);
 731
 732     if (cr->duty & DUTY_PME)
 733     {
 734         if (pme_lb->setup[pme_lb->cur].pmedata == NULL)
 735         {
 736             /* Generate a new PME data structure,
 737              * copying part of the old pointers.
 738              */
 739             gmx_pme_reinit(&set->pmedata,
 740                            cr, pme_lb->setup[0].pmedata, ir,
 741                            set->grid);
 742         }
 743         *pmedata = set->pmedata;
 744     }
 745     else
 746     {
 747         /* Tell our PME-only node to switch grid */
 748         gmx_pme_send_switchgrid(cr, set->grid, set->ewaldcoeff_q, set->ewaldcoeff_lj);
 749     }
 750
 751     if (debug)
 752     {
 753         print_grid(NULL, debug, "", "switched to", set, -1);
 754     }
 755
 756     if (pme_lb->stage == pme_lb->nstage)
 757     {
 758         print_grid(fp_err, fp_log, "", "optimal", set, -1);
 759     }
 760
 761     return TRUE;
 762 }
 763
 764 void restart_pme_loadbal(pme_load_balancing_t pme_lb, int n)
 765 {
 766     pme_lb->nstage += n;
 767 }
 768
 769 static int pme_grid_points(const pme_setup_t *setup)
 770 {
 771     return setup->grid[XX]*setup->grid[YY]*setup->grid[ZZ];
 772 }
 773
 774 static real pme_loadbal_rlist(const pme_setup_t *setup)
 775 {
 776     /* With the group cut-off scheme we can have twin-range either
 777      * for Coulomb or for VdW, so we need a check here.
 778      * With the Verlet cut-off scheme rlist=rlistlong.
 779      */
 780     if (setup->rcut_coulomb > setup->rlist)
 781     {
 782         return setup->rlistlong;
 783     }
 784     else
 785     {
 786         return setup->rlist;
 787     }
 788 }
 789
 790 static void print_pme_loadbal_setting(FILE              *fplog,
 791                                       char              *name,
 792                                       const pme_setup_t *setup)
 793 {
 794     fprintf(fplog,
 795             "   %-7s %6.3f nm %6.3f nm     %3d %3d %3d   %5.3f nm  %5.3f nm\n",
 796             name,
 797             setup->rcut_coulomb, pme_loadbal_rlist(setup),
 798             setup->grid[XX], setup->grid[YY], setup->grid[ZZ],
 799             setup->spacing, 1/setup->ewaldcoeff_q);
 800 }
 801
 802 static void print_pme_loadbal_settings(pme_load_balancing_t pme_lb,
 803                                        t_commrec           *cr,
 804                                        FILE                *fplog,
 805                                        gmx_bool             bNonBondedOnGPU)
 806 {
 807     double pp_ratio, grid_ratio;
 808
 809     pp_ratio   = pow(pme_loadbal_rlist(&pme_lb->setup[pme_lb->cur])/pme_loadbal_rlist(&pme_lb->setup[0]), 3.0);
 810     grid_ratio = pme_grid_points(&pme_lb->setup[pme_lb->cur])/
 811         (double)pme_grid_points(&pme_lb->setup[0]);
 812
 813     fprintf(fplog, "\n");
 814     fprintf(fplog, "       P P   -   P M E   L O A D   B A L A N C I N G\n");
 815     fprintf(fplog, "\n");
 816     /* Here we only warn when the optimal setting is the last one */
 817     if (pme_lb->elimited != epmelblimNO &&
 818         pme_lb->cur == pme_loadbal_end(pme_lb)-1)
 819     {
 820         fprintf(fplog, " NOTE: The PP/PME load balancing was limited by the %s,\n",
 821                 pmelblim_str[pme_lb->elimited]);
 822         fprintf(fplog, "       you might not have reached a good load balance.\n");
 823         if (pme_lb->elimited == epmelblimDD)
 824         {
 825             fprintf(fplog, "       Try different mdrun -dd settings or lower the -dds value.\n");
 826         }
 827         fprintf(fplog, "\n");
 828     }
 829     fprintf(fplog, " PP/PME load balancing changed the cut-off and PME settings:\n");
 830     fprintf(fplog, "           particle-particle                    PME\n");
 831     fprintf(fplog, "            rcoulomb  rlist            grid      spacing   1/beta\n");
 832     print_pme_loadbal_setting(fplog, "initial", &pme_lb->setup[0]);
 833     print_pme_loadbal_setting(fplog, "final", &pme_lb->setup[pme_lb->cur]);
 834     fprintf(fplog, " cost-ratio           %4.2f             %4.2f\n",
 835             pp_ratio, grid_ratio);
 836     fprintf(fplog, " (note that these numbers concern only part of the total PP and PME load)\n");
 837
 838     if (pp_ratio > 1.5 && !bNonBondedOnGPU)
 839     {
 840         md_print_warn(cr, fplog,
 841                       "NOTE: PME load balancing increased the non-bonded workload by more than 50%%.\n"
 842                       "      For better performance, use (more) PME ranks (mdrun -npme),\n"
 843                       "      or if you are beyond the scaling limit, use fewer total ranks (or nodes).\n");
 844     }
 845     else
 846     {
 847         fprintf(fplog, "\n");
 848     }
 849 }
 850
 851 void pme_loadbal_done(pme_load_balancing_t pme_lb,
 852                       t_commrec *cr, FILE *fplog,
 853                       gmx_bool bNonBondedOnGPU)
 854 {
 855     if (fplog != NULL && (pme_lb->cur > 0 || pme_lb->elimited != epmelblimNO))
 856     {
 857         print_pme_loadbal_settings(pme_lb, cr, fplog, bNonBondedOnGPU);
 858     }
 859
 860     /* TODO: Here we should free all pointers in pme_lb,
 861      * but as it contains pme data structures,
 862      * we need to first make pme.c free all data.
 863      */
 864 }