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 "smalloc.h"
  40 #include "network.h"
  41 #include "calcgrid.h"
  42 #include "pme.h"
  43 #include "vec.h"
  44 #include "domdec.h"
  45 #include "nbnxn_cuda_data_mgmt.h"
  46 #include "force.h"
  47 #include "macros.h"
  48 #include "md_logging.h"
  49 #include "pme_loadbal.h"
  50
  51 /* Parameters and setting for one PP-PME setup */
  52 typedef struct {
  53     real      rcut_coulomb;    /* Coulomb cut-off                              */
  54     real      rlist;           /* pair-list cut-off                            */
  55     real      rlistlong;       /* LR pair-list cut-off                         */
  56     int       nstcalclr;       /* frequency of evaluating long-range forces for group scheme */
  57     real      spacing;         /* (largest) PME grid spacing                   */
  58     ivec      grid;            /* the PME grid dimensions                      */
  59     real      grid_efficiency; /* ineffiency factor for non-uniform grids <= 1 */
  60     real      ewaldcoeff_q;    /* Electrostatic Ewald coefficient            */
  61     real      ewaldcoeff_lj;   /* LJ Ewald coefficient, only for the call to send_switchgrid */
  62     gmx_pme_t pmedata;         /* the data structure used in the PME code      */
  63     int       count;           /* number of times this setup has been timed    */
  64     double    cycles;          /* the fastest time for this setup in cycles    */
  65 } pme_setup_t;
  66
  67 /* In the initial scan, step by grids that are at least a factor 0.8 coarser */
  68 #define PME_LB_GRID_SCALE_FAC  0.8
  69 /* In the initial scan, try to skip grids with uneven x/y/z spacing,
  70  * checking if the "efficiency" is more than 5% worse than the previous grid.
  71  */
  72 #define PME_LB_GRID_EFFICIENCY_REL_FAC  1.05
  73 /* Rerun up till 12% slower setups than the fastest up till now */
  74 #define PME_LB_SLOW_FAC  1.12
  75 /* If setups get more than 2% faster, do another round to avoid
  76  * choosing a slower setup due to acceleration or fluctuations.
  77  */
  78 #define PME_LB_ACCEL_TOL 1.02
  79
  80 enum {
  81     epmelblimNO, epmelblimBOX, epmelblimDD, epmelblimPMEGRID, epmelblimNR
  82 };
  83
  84 const char *pmelblim_str[epmelblimNR] =
  85 { "no", "box size", "domain decompostion", "PME grid restriction" };
  86
  87 struct pme_load_balancing {
  88     int          nstage;             /* the current maximum number of stages */
  89
  90     real         cut_spacing;        /* the minimum cutoff / PME grid spacing ratio */
  91     real         rcut_vdw;           /* Vdw cutoff (does not change) */
  92     real         rcut_coulomb_start; /* Initial electrostatics cutoff */
  93     int          nstcalclr_start;    /* Initial electrostatics cutoff */
  94     real         rbuf_coulomb;       /* the pairlist buffer size */
  95     real         rbuf_vdw;           /* the pairlist buffer size */
  96     matrix       box_start;          /* the initial simulation box */
  97     int          n;                  /* the count of setup as well as the allocation size */
  98     pme_setup_t *setup;              /* the PME+cutoff setups */
  99     int          cur;                /* the current setup */
 100     int          fastest;            /* fastest setup up till now */
 101     int          start;              /* start of setup range to consider in stage>0 */
 102     int          end;                /* end   of setup range to consider in stage>0 */
 103     int          elimited;           /* was the balancing limited, uses enum above */
 104     int          cutoff_scheme;      /* Verlet or group cut-offs */
 105
 106     int          stage;              /* the current stage */
 107 };
 108
 109 void pme_loadbal_init(pme_load_balancing_t *pme_lb_p,
 110                       const t_inputrec *ir, matrix box,
 111                       const interaction_const_t *ic,
 112                       gmx_pme_t pmedata)
 113 {
 114     pme_load_balancing_t pme_lb;
 115     real                 spm, sp;
 116     int                  d;
 117
 118     snew(pme_lb, 1);
 119
 120     /* Any number of stages >= 2 is supported */
 121     pme_lb->nstage   = 2;
 122
 123     pme_lb->cutoff_scheme = ir->cutoff_scheme;
 124
 125     if (pme_lb->cutoff_scheme == ecutsVERLET)
 126     {
 127         pme_lb->rbuf_coulomb = ic->rlist - ic->rcoulomb;
 128         pme_lb->rbuf_vdw     = pme_lb->rbuf_coulomb;
 129     }
 130     else
 131     {
 132         if (ic->rcoulomb > ic->rlist)
 133         {
 134             pme_lb->rbuf_coulomb = ic->rlistlong - ic->rcoulomb;
 135         }
 136         else
 137         {
 138             pme_lb->rbuf_coulomb = ic->rlist - ic->rcoulomb;
 139         }
 140         if (ic->rvdw > ic->rlist)
 141         {
 142             pme_lb->rbuf_vdw = ic->rlistlong - ic->rvdw;
 143         }
 144         else
 145         {
 146             pme_lb->rbuf_vdw = ic->rlist - ic->rvdw;
 147         }
 148     }
 149
 150     copy_mat(box, pme_lb->box_start);
 151     if (ir->ePBC == epbcXY && ir->nwall == 2)
 152     {
 153         svmul(ir->wall_ewald_zfac, pme_lb->box_start[ZZ], pme_lb->box_start[ZZ]);
 154     }
 155
 156     pme_lb->n = 1;
 157     snew(pme_lb->setup, pme_lb->n);
 158
 159     pme_lb->rcut_vdw                 = ic->rvdw;
 160     pme_lb->rcut_coulomb_start       = ir->rcoulomb;
 161     pme_lb->nstcalclr_start          = ir->nstcalclr;
 162
 163     pme_lb->cur                      = 0;
 164     pme_lb->setup[0].rcut_coulomb    = ic->rcoulomb;
 165     pme_lb->setup[0].rlist           = ic->rlist;
 166     pme_lb->setup[0].rlistlong       = ic->rlistlong;
 167     pme_lb->setup[0].nstcalclr       = ir->nstcalclr;
 168     pme_lb->setup[0].grid[XX]        = ir->nkx;
 169     pme_lb->setup[0].grid[YY]        = ir->nky;
 170     pme_lb->setup[0].grid[ZZ]        = ir->nkz;
 171     pme_lb->setup[0].ewaldcoeff_q    = ic->ewaldcoeff_q;
 172     pme_lb->setup[0].ewaldcoeff_lj   = ic->ewaldcoeff_lj;
 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                                             const gmx_domdec_t   *dd)
 209 {
 210     pme_setup_t *set;
 211     int          npmenodes_x, npmenodes_y;
 212     real         fac, sp;
 213     real         tmpr_coulomb, tmpr_vdw;
 214     int          d;
 215     gmx_bool     grid_ok;
 216
 217     /* Try to add a new setup with next larger cut-off to the list */
 218     pme_lb->n++;
 219     srenew(pme_lb->setup, pme_lb->n);
 220     set          = &pme_lb->setup[pme_lb->n-1];
 221     set->pmedata = NULL;
 222
 223     get_pme_nnodes(dd, &npmenodes_x, &npmenodes_y);
 224
 225     fac = 1;
 226     do
 227     {
 228         /* Avoid infinite while loop, which can occur at the minimum grid size.
 229          * Note that in practice load balancing will stop before this point.
 230          * The factor 2.1 allows for the extreme case in which only grids
 231          * of powers of 2 are allowed (the current code supports more grids).
 232          */
 233         if (fac > 2.1)
 234         {
 235             pme_lb->n--;
 236
 237             return FALSE;
 238         }
 239
 240         fac *= 1.01;
 241         clear_ivec(set->grid);
 242         sp = calc_grid(NULL, pme_lb->box_start,
 243                        fac*pme_lb->setup[pme_lb->cur].spacing,
 244                        &set->grid[XX],
 245                        &set->grid[YY],
 246                        &set->grid[ZZ]);
 247
 248         /* As here we can't easily check if one of the PME nodes
 249          * uses threading, we do a conservative grid check.
 250          * This means we can't use pme_order or less grid lines
 251          * per PME node along x, which is not a strong restriction.
 252          */
 253         gmx_pme_check_restrictions(pme_order,
 254                                    set->grid[XX], set->grid[YY], set->grid[ZZ],
 255                                    npmenodes_x, npmenodes_y,
 256                                    TRUE,
 257                                    FALSE,
 258                                    &grid_ok);
 259     }
 260     while (sp <= 1.001*pme_lb->setup[pme_lb->cur].spacing || !grid_ok);
 261
 262     set->rcut_coulomb = pme_lb->cut_spacing*sp;
 263
 264     if (pme_lb->cutoff_scheme == ecutsVERLET)
 265     {
 266         set->rlist        = set->rcut_coulomb + pme_lb->rbuf_coulomb;
 267         /* We dont use LR lists with Verlet, but this avoids if-statements in further checks */
 268         set->rlistlong    = set->rlist;
 269     }
 270     else
 271     {
 272         tmpr_coulomb          = set->rcut_coulomb + pme_lb->rbuf_coulomb;
 273         tmpr_vdw              = pme_lb->rcut_vdw + pme_lb->rbuf_vdw;
 274         set->rlist            = min(tmpr_coulomb, tmpr_vdw);
 275         set->rlistlong        = max(tmpr_coulomb, tmpr_vdw);
 276
 277         /* Set the long-range update frequency */
 278         if (set->rlist == set->rlistlong)
 279         {
 280             /* No long-range interactions if the short-/long-range cutoffs are identical */
 281             set->nstcalclr = 0;
 282         }
 283         else if (pme_lb->nstcalclr_start == 0 || pme_lb->nstcalclr_start == 1)
 284         {
 285             /* We were not doing long-range before, but now we are since rlist!=rlistlong */
 286             set->nstcalclr = 1;
 287         }
 288         else
 289         {
 290             /* We were already doing long-range interactions from the start */
 291             if (pme_lb->rcut_vdw > pme_lb->rcut_coulomb_start)
 292             {
 293                 /* We were originally doing long-range VdW-only interactions.
 294                  * If rvdw is still longer than rcoulomb we keep the original nstcalclr,
 295                  * but if the coulomb cutoff has become longer we should update the long-range
 296                  * part every step.
 297                  */
 298                 set->nstcalclr = (tmpr_vdw > tmpr_coulomb) ? pme_lb->nstcalclr_start : 1;
 299             }
 300             else
 301             {
 302                 /* We were not doing any long-range interaction from the start,
 303                  * since it is not possible to do twin-range coulomb for the PME interaction.
 304                  */
 305                 set->nstcalclr = 1;
 306             }
 307         }
 308     }
 309
 310     set->spacing      = sp;
 311     /* The grid efficiency is the size wrt a grid with uniform x/y/z spacing */
 312     set->grid_efficiency = 1;
 313     for (d = 0; d < DIM; d++)
 314     {
 315         set->grid_efficiency *= (set->grid[d]*sp)/norm(pme_lb->box_start[d]);
 316     }
 317     /* The Ewald coefficient is inversly proportional to the cut-off */
 318     set->ewaldcoeff_q =
 319         pme_lb->setup[0].ewaldcoeff_q*pme_lb->setup[0].rcut_coulomb/set->rcut_coulomb;
 320     /* We set ewaldcoeff_lj in set, even when LJ-PME is not used */
 321     set->ewaldcoeff_lj =
 322         pme_lb->setup[0].ewaldcoeff_lj*pme_lb->setup[0].rcut_coulomb/set->rcut_coulomb;
 323
 324     set->count   = 0;
 325     set->cycles  = 0;
 326
 327     if (debug)
 328     {
 329         fprintf(debug, "PME loadbal: grid %d %d %d, coulomb cutoff %f\n",
 330                 set->grid[XX], set->grid[YY], set->grid[ZZ], set->rcut_coulomb);
 331     }
 332     return TRUE;
 333 }
 334
 335 static void print_grid(FILE *fp_err, FILE *fp_log,
 336                        const char *pre,
 337                        const char *desc,
 338                        const pme_setup_t *set,
 339                        double cycles)
 340 {
 341     char buf[STRLEN], buft[STRLEN];
 342
 343     if (cycles >= 0)
 344     {
 345         sprintf(buft, ": %.1f M-cycles", cycles*1e-6);
 346     }
 347     else
 348     {
 349         buft[0] = '\0';
 350     }
 351     sprintf(buf, "%-11s%10s pme grid %d %d %d, coulomb cutoff %.3f%s",
 352             pre,
 353             desc, set->grid[XX], set->grid[YY], set->grid[ZZ], set->rcut_coulomb,
 354             buft);
 355     if (fp_err != NULL)
 356     {
 357         fprintf(fp_err, "\r%s\n", buf);
 358     }
 359     if (fp_log != NULL)
 360     {
 361         fprintf(fp_log, "%s\n", buf);
 362     }
 363 }
 364
 365 static int pme_loadbal_end(pme_load_balancing_t pme_lb)
 366 {
 367     /* In the initial stage only n is set; end is not set yet */
 368     if (pme_lb->end > 0)
 369     {
 370         return pme_lb->end;
 371     }
 372     else
 373     {
 374         return pme_lb->n;
 375     }
 376 }
 377
 378 static void print_loadbal_limited(FILE *fp_err, FILE *fp_log,
 379                                   gmx_int64_t step,
 380                                   pme_load_balancing_t pme_lb)
 381 {
 382     char buf[STRLEN], sbuf[22];
 383
 384     sprintf(buf, "step %4s: the %s limits the PME load balancing to a coulomb cut-off of %.3f",
 385             gmx_step_str(step, sbuf),
 386             pmelblim_str[pme_lb->elimited],
 387             pme_lb->setup[pme_loadbal_end(pme_lb)-1].rcut_coulomb);
 388     if (fp_err != NULL)
 389     {
 390         fprintf(fp_err, "\r%s\n", buf);
 391     }
 392     if (fp_log != NULL)
 393     {
 394         fprintf(fp_log, "%s\n", buf);
 395     }
 396 }
 397
 398 static void switch_to_stage1(pme_load_balancing_t pme_lb)
 399 {
 400     pme_lb->start = 0;
 401     while (pme_lb->start+1 < pme_lb->n &&
 402            (pme_lb->setup[pme_lb->start].count == 0 ||
 403             pme_lb->setup[pme_lb->start].cycles >
 404             pme_lb->setup[pme_lb->fastest].cycles*PME_LB_SLOW_FAC))
 405     {
 406         pme_lb->start++;
 407     }
 408     while (pme_lb->start > 0 && pme_lb->setup[pme_lb->start-1].cycles == 0)
 409     {
 410         pme_lb->start--;
 411     }
 412
 413     pme_lb->end = pme_lb->n;
 414     if (pme_lb->setup[pme_lb->end-1].count > 0 &&
 415         pme_lb->setup[pme_lb->end-1].cycles >
 416         pme_lb->setup[pme_lb->fastest].cycles*PME_LB_SLOW_FAC)
 417     {
 418         pme_lb->end--;
 419     }
 420
 421     pme_lb->stage = 1;
 422
 423     /* Next we want to choose setup pme_lb->start, but as we will increase
 424      * pme_ln->cur by one right after returning, we subtract 1 here.
 425      */
 426     pme_lb->cur = pme_lb->start - 1;
 427 }
 428
 429 gmx_bool pme_load_balance(pme_load_balancing_t pme_lb,
 430                           t_commrec           *cr,
 431                           FILE                *fp_err,
 432                           FILE                *fp_log,
 433                           t_inputrec          *ir,
 434                           t_state             *state,
 435                           double               cycles,
 436                           interaction_const_t *ic,
 437                           nonbonded_verlet_t  *nbv,
 438                           gmx_pme_t           *pmedata,
 439                           gmx_int64_t          step)
 440 {
 441     gmx_bool     OK;
 442     pme_setup_t *set;
 443     double       cycles_fast;
 444     char         buf[STRLEN], sbuf[22];
 445     real         rtab;
 446     gmx_bool     bUsesSimpleTables = TRUE;
 447
 448     if (pme_lb->stage == pme_lb->nstage)
 449     {
 450         return FALSE;
 451     }
 452
 453     if (PAR(cr))
 454     {
 455         gmx_sumd(1, &cycles, cr);
 456         cycles /= cr->nnodes;
 457     }
 458
 459     set = &pme_lb->setup[pme_lb->cur];
 460     set->count++;
 461
 462     rtab = ir->rlistlong + ir->tabext;
 463
 464     if (set->count % 2 == 1)
 465     {
 466         /* Skip the first cycle, because the first step after a switch
 467          * is much slower due to allocation and/or caching effects.
 468          */
 469         return TRUE;
 470     }
 471
 472     sprintf(buf, "step %4s: ", gmx_step_str(step, sbuf));
 473     print_grid(fp_err, fp_log, buf, "timed with", set, cycles);
 474
 475     if (set->count <= 2)
 476     {
 477         set->cycles = cycles;
 478     }
 479     else
 480     {
 481         if (cycles*PME_LB_ACCEL_TOL < set->cycles &&
 482             pme_lb->stage == pme_lb->nstage - 1)
 483         {
 484             /* The performance went up a lot (due to e.g. DD load balancing).
 485              * Add a stage, keep the minima, but rescan all setups.
 486              */
 487             pme_lb->nstage++;
 488
 489             if (debug)
 490             {
 491                 fprintf(debug, "The performance for grid %d %d %d went from %.3f to %.1f M-cycles, this is more than %f\n"
 492                         "Increased the number stages to %d"
 493                         " and ignoring the previous performance\n",
 494                         set->grid[XX], set->grid[YY], set->grid[ZZ],
 495                         cycles*1e-6, set->cycles*1e-6, PME_LB_ACCEL_TOL,
 496                         pme_lb->nstage);
 497             }
 498         }
 499         set->cycles = min(set->cycles, cycles);
 500     }
 501
 502     if (set->cycles < pme_lb->setup[pme_lb->fastest].cycles)
 503     {
 504         pme_lb->fastest = pme_lb->cur;
 505
 506         if (DOMAINDECOMP(cr))
 507         {
 508             /* We found a new fastest setting, ensure that with subsequent
 509              * shorter cut-off's the dynamic load balancing does not make
 510              * the use of the current cut-off impossible. This solution is
 511              * a trade-off, as the PME load balancing and DD domain size
 512              * load balancing can interact in complex ways.
 513              * With the Verlet kernels, DD load imbalance will usually be
 514              * mainly due to bonded interaction imbalance, which will often
 515              * quickly push the domain boundaries beyond the limit for the
 516              * optimal, PME load balanced, cut-off. But it could be that
 517              * better overal performance can be obtained with a slightly
 518              * shorter cut-off and better DD load balancing.
 519              */
 520             change_dd_dlb_cutoff_limit(cr);
 521         }
 522     }
 523     cycles_fast = pme_lb->setup[pme_lb->fastest].cycles;
 524
 525     /* Check in stage 0 if we should stop scanning grids.
 526      * Stop when the time is more than SLOW_FAC longer than the fastest.
 527      */
 528     if (pme_lb->stage == 0 && pme_lb->cur > 0 &&
 529         cycles > pme_lb->setup[pme_lb->fastest].cycles*PME_LB_SLOW_FAC)
 530     {
 531         pme_lb->n = pme_lb->cur + 1;
 532         /* Done with scanning, go to stage 1 */
 533         switch_to_stage1(pme_lb);
 534     }
 535
 536     if (pme_lb->stage == 0)
 537     {
 538         int gridsize_start;
 539
 540         gridsize_start = set->grid[XX]*set->grid[YY]*set->grid[ZZ];
 541
 542         do
 543         {
 544             if (pme_lb->cur+1 < pme_lb->n)
 545             {
 546                 /* We had already generated the next setup */
 547                 OK = TRUE;
 548             }
 549             else
 550             {
 551                 /* Find the next setup */
 552                 OK = pme_loadbal_increase_cutoff(pme_lb, ir->pme_order, cr->dd);
 553
 554                 if (!OK)
 555                 {
 556                     pme_lb->elimited = epmelblimPMEGRID;
 557                 }
 558             }
 559
 560             if (OK && ir->ePBC != epbcNONE)
 561             {
 562                 OK = (sqr(pme_lb->setup[pme_lb->cur+1].rlistlong)
 563                       <= max_cutoff2(ir->ePBC, state->box));
 564                 if (!OK)
 565                 {
 566                     pme_lb->elimited = epmelblimBOX;
 567                 }
 568             }
 569
 570             if (OK)
 571             {
 572                 pme_lb->cur++;
 573
 574                 if (DOMAINDECOMP(cr))
 575                 {
 576                     OK = change_dd_cutoff(cr, state, ir,
 577                                           pme_lb->setup[pme_lb->cur].rlistlong);
 578                     if (!OK)
 579                     {
 580                         /* Failed: do not use this setup */
 581                         pme_lb->cur--;
 582                         pme_lb->elimited = epmelblimDD;
 583                     }
 584                 }
 585             }
 586             if (!OK)
 587             {
 588                 /* We hit the upper limit for the cut-off,
 589                  * the setup should not go further than cur.
 590                  */
 591                 pme_lb->n = pme_lb->cur + 1;
 592                 print_loadbal_limited(fp_err, fp_log, step, pme_lb);
 593                 /* Switch to the next stage */
 594                 switch_to_stage1(pme_lb);
 595             }
 596         }
 597         while (OK &&
 598                !(pme_lb->setup[pme_lb->cur].grid[XX]*
 599                  pme_lb->setup[pme_lb->cur].grid[YY]*
 600                  pme_lb->setup[pme_lb->cur].grid[ZZ] <
 601                  gridsize_start*PME_LB_GRID_SCALE_FAC
 602                  &&
 603                  pme_lb->setup[pme_lb->cur].grid_efficiency <
 604                  pme_lb->setup[pme_lb->cur-1].grid_efficiency*PME_LB_GRID_EFFICIENCY_REL_FAC));
 605     }
 606
 607     if (pme_lb->stage > 0 && pme_lb->end == 1)
 608     {
 609         pme_lb->cur   = 0;
 610         pme_lb->stage = pme_lb->nstage;
 611     }
 612     else if (pme_lb->stage > 0 && pme_lb->end > 1)
 613     {
 614         /* If stage = nstage-1:
 615          *   scan over all setups, rerunning only those setups
 616          *   which are not much slower than the fastest
 617          * else:
 618          *   use the next setup
 619          */
 620         do
 621         {
 622             pme_lb->cur++;
 623             if (pme_lb->cur == pme_lb->end)
 624             {
 625                 pme_lb->stage++;
 626                 pme_lb->cur = pme_lb->start;
 627             }
 628         }
 629         while (pme_lb->stage == pme_lb->nstage - 1 &&
 630                pme_lb->setup[pme_lb->cur].count > 0 &&
 631                pme_lb->setup[pme_lb->cur].cycles > cycles_fast*PME_LB_SLOW_FAC);
 632
 633         if (pme_lb->stage == pme_lb->nstage)
 634         {
 635             /* We are done optimizing, use the fastest setup we found */
 636             pme_lb->cur = pme_lb->fastest;
 637         }
 638     }
 639
 640     if (DOMAINDECOMP(cr) && pme_lb->stage > 0)
 641     {
 642         OK = change_dd_cutoff(cr, state, ir, pme_lb->setup[pme_lb->cur].rlistlong);
 643         if (!OK)
 644         {
 645             /* Failsafe solution */
 646             if (pme_lb->cur > 1 && pme_lb->stage == pme_lb->nstage)
 647             {
 648                 pme_lb->stage--;
 649             }
 650             pme_lb->fastest  = 0;
 651             pme_lb->start    = 0;
 652             pme_lb->end      = pme_lb->cur;
 653             pme_lb->cur      = pme_lb->start;
 654             pme_lb->elimited = epmelblimDD;
 655             print_loadbal_limited(fp_err, fp_log, step, pme_lb);
 656         }
 657     }
 658
 659     /* Change the Coulomb cut-off and the PME grid */
 660
 661     set = &pme_lb->setup[pme_lb->cur];
 662
 663     ic->rcoulomb     = set->rcut_coulomb;
 664     ic->rlist        = set->rlist;
 665     ic->rlistlong    = set->rlistlong;
 666     ir->nstcalclr    = set->nstcalclr;
 667     ic->ewaldcoeff_q = set->ewaldcoeff_q;
 668     /* TODO: centralize the code that sets the potentials shifts */
 669     if (ic->coulomb_modifier == eintmodPOTSHIFT)
 670     {
 671         ic->sh_ewald = gmx_erfc(ic->ewaldcoeff_q*ic->rcoulomb);
 672     }
 673     if (EVDW_PME(ic->vdwtype))
 674     {
 675         /* We have PME for both Coulomb and VdW, set rvdw equal to rcoulomb */
 676         ic->rvdw            = set->rcut_coulomb;
 677         ic->ewaldcoeff_lj   = set->ewaldcoeff_lj;
 678         if (ic->vdw_modifier == eintmodPOTSHIFT)
 679         {
 680             real crc2;
 681
 682             ic->dispersion_shift.cpot = -pow(ic->rvdw, -6.0);
 683             ic->repulsion_shift.cpot  = -pow(ic->rvdw, -12.0);
 684             ic->sh_invrc6             = -ic->dispersion_shift.cpot;
 685             crc2                      = sqr(ic->ewaldcoeff_lj*ic->rvdw);
 686             ic->sh_lj_ewald           = (exp(-crc2)*(1 + crc2 + 0.5*crc2*crc2) - 1)*pow(ic->rvdw, -6.0);
 687         }
 688     }
 689
 690     bUsesSimpleTables = uses_simple_tables(ir->cutoff_scheme, nbv, 0);
 691     if (pme_lb->cutoff_scheme == ecutsVERLET &&
 692         nbv->grp[0].kernel_type == nbnxnk8x8x8_CUDA)
 693     {
 694         nbnxn_cuda_pme_loadbal_update_param(nbv->cu_nbv, ic);
 695
 696         /* With tMPI + GPUs some ranks may be sharing GPU(s) and therefore
 697          * also sharing texture references. To keep the code simple, we don't
 698          * treat texture references as shared resources, but this means that
 699          * the coulomb_tab texture ref will get updated by multiple threads.
 700          * Hence, to ensure that the non-bonded kernels don't start before all
 701          * texture binding operations are finished, we need to wait for all ranks
 702          * to arrive here before continuing.
 703          *
 704          * Note that we could omit this barrier if GPUs are not shared (or
 705          * texture objects are used), but as this is initialization code, there
 706          * is not point in complicating things.
 707          */
 708 #ifdef GMX_THREAD_MPI
 709         if (PAR(cr))
 710         {
 711             gmx_barrier(cr);
 712         }
 713 #endif  /* GMX_THREAD_MPI */
 714     }
 715
 716     /* Usually we won't need the simple tables with GPUs.
 717      * But we do with hybrid acceleration and with free energy.
 718      * To avoid bugs, we always re-initialize the simple tables here.
 719      */
 720     init_interaction_const_tables(NULL, ic, bUsesSimpleTables, rtab);
 721
 722     if (cr->duty & DUTY_PME)
 723     {
 724         if (pme_lb->setup[pme_lb->cur].pmedata == NULL)
 725         {
 726             /* Generate a new PME data structure,
 727              * copying part of the old pointers.
 728              */
 729             gmx_pme_reinit(&set->pmedata,
 730                            cr, pme_lb->setup[0].pmedata, ir,
 731                            set->grid);
 732         }
 733         *pmedata = set->pmedata;
 734     }
 735     else
 736     {
 737         /* Tell our PME-only node to switch grid */
 738         gmx_pme_send_switchgrid(cr, set->grid, set->ewaldcoeff_q, set->ewaldcoeff_lj);
 739     }
 740
 741     if (debug)
 742     {
 743         print_grid(NULL, debug, "", "switched to", set, -1);
 744     }
 745
 746     if (pme_lb->stage == pme_lb->nstage)
 747     {
 748         print_grid(fp_err, fp_log, "", "optimal", set, -1);
 749     }
 750
 751     return TRUE;
 752 }
 753
 754 void restart_pme_loadbal(pme_load_balancing_t pme_lb, int n)
 755 {
 756     pme_lb->nstage += n;
 757 }
 758
 759 static int pme_grid_points(const pme_setup_t *setup)
 760 {
 761     return setup->grid[XX]*setup->grid[YY]*setup->grid[ZZ];
 762 }
 763
 764 static real pme_loadbal_rlist(const pme_setup_t *setup)
 765 {
 766     /* With the group cut-off scheme we can have twin-range either
 767      * for Coulomb or for VdW, so we need a check here.
 768      * With the Verlet cut-off scheme rlist=rlistlong.
 769      */
 770     if (setup->rcut_coulomb > setup->rlist)
 771     {
 772         return setup->rlistlong;
 773     }
 774     else
 775     {
 776         return setup->rlist;
 777     }
 778 }
 779
 780 static void print_pme_loadbal_setting(FILE              *fplog,
 781                                       char              *name,
 782                                       const pme_setup_t *setup)
 783 {
 784     fprintf(fplog,
 785             "   %-7s %6.3f nm %6.3f nm     %3d %3d %3d   %5.3f nm  %5.3f nm\n",
 786             name,
 787             setup->rcut_coulomb, pme_loadbal_rlist(setup),
 788             setup->grid[XX], setup->grid[YY], setup->grid[ZZ],
 789             setup->spacing, 1/setup->ewaldcoeff_q);
 790 }
 791
 792 static void print_pme_loadbal_settings(pme_load_balancing_t pme_lb,
 793                                        t_commrec           *cr,
 794                                        FILE                *fplog,
 795                                        gmx_bool             bNonBondedOnGPU)
 796 {
 797     double pp_ratio, grid_ratio;
 798
 799     pp_ratio   = pow(pme_loadbal_rlist(&pme_lb->setup[pme_lb->cur])/pme_loadbal_rlist(&pme_lb->setup[0]), 3.0);
 800     grid_ratio = pme_grid_points(&pme_lb->setup[pme_lb->cur])/
 801         (double)pme_grid_points(&pme_lb->setup[0]);
 802
 803     fprintf(fplog, "\n");
 804     fprintf(fplog, "       P P   -   P M E   L O A D   B A L A N C I N G\n");
 805     fprintf(fplog, "\n");
 806     /* Here we only warn when the optimal setting is the last one */
 807     if (pme_lb->elimited != epmelblimNO &&
 808         pme_lb->cur == pme_loadbal_end(pme_lb)-1)
 809     {
 810         fprintf(fplog, " NOTE: The PP/PME load balancing was limited by the %s,\n",
 811                 pmelblim_str[pme_lb->elimited]);
 812         fprintf(fplog, "       you might not have reached a good load balance.\n");
 813         if (pme_lb->elimited == epmelblimDD)
 814         {
 815             fprintf(fplog, "       Try different mdrun -dd settings or lower the -dds value.\n");
 816         }
 817         fprintf(fplog, "\n");
 818     }
 819     fprintf(fplog, " PP/PME load balancing changed the cut-off and PME settings:\n");
 820     fprintf(fplog, "           particle-particle                    PME\n");
 821     fprintf(fplog, "            rcoulomb  rlist            grid      spacing   1/beta\n");
 822     print_pme_loadbal_setting(fplog, "initial", &pme_lb->setup[0]);
 823     print_pme_loadbal_setting(fplog, "final", &pme_lb->setup[pme_lb->cur]);
 824     fprintf(fplog, " cost-ratio           %4.2f             %4.2f\n",
 825             pp_ratio, grid_ratio);
 826     fprintf(fplog, " (note that these numbers concern only part of the total PP and PME load)\n");
 827
 828     if (pp_ratio > 1.5 && !bNonBondedOnGPU)
 829     {
 830         md_print_warn(cr, fplog,
 831                       "NOTE: PME load balancing increased the non-bonded workload by more than 50%%.\n"
 832                       "      For better performance use (more) PME nodes (mdrun -npme),\n"
 833                       "      or in case you are beyond the scaling limit, use less nodes in total.\n");
 834     }
 835     else
 836     {
 837         fprintf(fplog, "\n");
 838     }
 839 }
 840
 841 void pme_loadbal_done(pme_load_balancing_t pme_lb,
 842                       t_commrec *cr, FILE *fplog,
 843                       gmx_bool bNonBondedOnGPU)
 844 {
 845     if (fplog != NULL && (pme_lb->cur > 0 || pme_lb->elimited != epmelblimNO))
 846     {
 847         print_pme_loadbal_settings(pme_lb, cr, fplog, bNonBondedOnGPU);
 848     }
 849
 850     /* TODO: Here we should free all pointers in pme_lb,
 851      * but as it contains pme data structures,
 852      * we need to first make pme.c free all data.
 853      */
 854 }