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 #include "gmxpre.h"
  36
  37 #include "pme_loadbal.h"
  38
  39 #include "config.h"
  40
  41 #include "gromacs/legacyheaders/calcgrid.h"
  42 #include "gromacs/legacyheaders/domdec.h"
  43 #include "gromacs/legacyheaders/force.h"
  44 #include "gromacs/legacyheaders/macros.h"
  45 #include "gromacs/legacyheaders/md_logging.h"
  46 #include "gromacs/legacyheaders/network.h"
  47 #include "gromacs/legacyheaders/pme.h"
  48 #include "gromacs/legacyheaders/sim_util.h"
  49 #include "gromacs/legacyheaders/types/commrec.h"
  50 #include "gromacs/math/vec.h"
  51 #include "gromacs/mdlib/nbnxn_cuda/nbnxn_cuda_data_mgmt.h"
  52 #include "gromacs/pbcutil/pbc.h"
  53 #include "gromacs/utility/cstringutil.h"
  54 #include "gromacs/utility/smalloc.h"
  55
  56 /* Parameters and setting for one PP-PME setup */
  57 typedef struct {
  58     real      rcut_coulomb;    /* Coulomb cut-off                              */
  59     real      rlist;           /* pair-list cut-off                            */
  60     real      rlistlong;       /* LR pair-list cut-off                         */
  61     int       nstcalclr;       /* frequency of evaluating long-range forces for group scheme */
  62     real      spacing;         /* (largest) PME grid spacing                   */
  63     ivec      grid;            /* the PME grid dimensions                      */
  64     real      grid_efficiency; /* ineffiency factor for non-uniform grids <= 1 */
  65     real      ewaldcoeff_q;    /* Electrostatic Ewald coefficient            */
  66     real      ewaldcoeff_lj;   /* LJ Ewald coefficient, only for the call to send_switchgrid */
  67     gmx_pme_t pmedata;         /* the data structure used in the PME code      */
  68     int       count;           /* number of times this setup has been timed    */
  69     double    cycles;          /* the fastest time for this setup in cycles    */
  70 } pme_setup_t;
  71
  72 /* In the initial scan, step by grids that are at least a factor 0.8 coarser */
  73 #define PME_LB_GRID_SCALE_FAC  0.8
  74 /* In the initial scan, try to skip grids with uneven x/y/z spacing,
  75  * checking if the "efficiency" is more than 5% worse than the previous grid.
  76  */
  77 #define PME_LB_GRID_EFFICIENCY_REL_FAC  1.05
  78 /* Rerun up till 12% slower setups than the fastest up till now */
  79 #define PME_LB_SLOW_FAC  1.12
  80 /* If setups get more than 2% faster, do another round to avoid
  81  * choosing a slower setup due to acceleration or fluctuations.
  82  */
  83 #define PME_LB_ACCEL_TOL 1.02
  84
  85 enum {
  86     epmelblimNO, epmelblimBOX, epmelblimDD, epmelblimPMEGRID, epmelblimNR
  87 };
  88
  89 const char *pmelblim_str[epmelblimNR] =
  90 { "no", "box size", "domain decompostion", "PME grid restriction" };
  91
  92 struct pme_load_balancing {
  93     int          nstage;             /* the current maximum number of stages */
  94
  95     real         cut_spacing;        /* the minimum cutoff / PME grid spacing ratio */
  96     real         rcut_vdw;           /* Vdw cutoff (does not change) */
  97     real         rcut_coulomb_start; /* Initial electrostatics cutoff */
  98     int          nstcalclr_start;    /* Initial electrostatics cutoff */
  99     real         rbuf_coulomb;       /* the pairlist buffer size */
 100     real         rbuf_vdw;           /* the pairlist buffer size */
 101     matrix       box_start;          /* the initial simulation box */
 102     int          n;                  /* the count of setup as well as the allocation size */
 103     pme_setup_t *setup;              /* the PME+cutoff setups */
 104     int          cur;                /* the current setup */
 105     int          fastest;            /* fastest setup up till now */
 106     int          start;              /* start of setup range to consider in stage>0 */
 107     int          end;                /* end   of setup range to consider in stage>0 */
 108     int          elimited;           /* was the balancing limited, uses enum above */
 109     int          cutoff_scheme;      /* Verlet or group cut-offs */
 110
 111     int          stage;              /* the current stage */
 112 };
 113
 114 void pme_loadbal_init(pme_load_balancing_t *pme_lb_p,
 115                       const t_inputrec *ir, matrix box,
 116                       const interaction_const_t *ic,
 117                       gmx_pme_t pmedata)
 118 {
 119     pme_load_balancing_t pme_lb;
 120     real                 spm, sp;
 121     int                  d;
 122
 123     snew(pme_lb, 1);
 124
 125     /* Any number of stages >= 2 is supported */
 126     pme_lb->nstage   = 2;
 127
 128     pme_lb->cutoff_scheme = ir->cutoff_scheme;
 129
 130     if (pme_lb->cutoff_scheme == ecutsVERLET)
 131     {
 132         pme_lb->rbuf_coulomb = ic->rlist - ic->rcoulomb;
 133         pme_lb->rbuf_vdw     = pme_lb->rbuf_coulomb;
 134     }
 135     else
 136     {
 137         if (ic->rcoulomb > ic->rlist)
 138         {
 139             pme_lb->rbuf_coulomb = ic->rlistlong - ic->rcoulomb;
 140         }
 141         else
 142         {
 143             pme_lb->rbuf_coulomb = ic->rlist - ic->rcoulomb;
 144         }
 145         if (ic->rvdw > ic->rlist)
 146         {
 147             pme_lb->rbuf_vdw = ic->rlistlong - ic->rvdw;
 148         }
 149         else
 150         {
 151             pme_lb->rbuf_vdw = ic->rlist - ic->rvdw;
 152         }
 153     }
 154
 155     copy_mat(box, pme_lb->box_start);
 156     if (ir->ePBC == epbcXY && ir->nwall == 2)
 157     {
 158         svmul(ir->wall_ewald_zfac, pme_lb->box_start[ZZ], pme_lb->box_start[ZZ]);
 159     }
 160
 161     pme_lb->n = 1;
 162     snew(pme_lb->setup, pme_lb->n);
 163
 164     pme_lb->rcut_vdw                 = ic->rvdw;
 165     pme_lb->rcut_coulomb_start       = ir->rcoulomb;
 166     pme_lb->nstcalclr_start          = ir->nstcalclr;
 167
 168     pme_lb->cur                      = 0;
 169     pme_lb->setup[0].rcut_coulomb    = ic->rcoulomb;
 170     pme_lb->setup[0].rlist           = ic->rlist;
 171     pme_lb->setup[0].rlistlong       = ic->rlistlong;
 172     pme_lb->setup[0].nstcalclr       = ir->nstcalclr;
 173     pme_lb->setup[0].grid[XX]        = ir->nkx;
 174     pme_lb->setup[0].grid[YY]        = ir->nky;
 175     pme_lb->setup[0].grid[ZZ]        = ir->nkz;
 176     pme_lb->setup[0].ewaldcoeff_q    = ic->ewaldcoeff_q;
 177     pme_lb->setup[0].ewaldcoeff_lj   = ic->ewaldcoeff_lj;
 178
 179     pme_lb->setup[0].pmedata  = pmedata;
 180
 181     spm = 0;
 182     for (d = 0; d < DIM; d++)
 183     {
 184         sp = norm(pme_lb->box_start[d])/pme_lb->setup[0].grid[d];
 185         if (sp > spm)
 186         {
 187             spm = sp;
 188         }
 189     }
 190     pme_lb->setup[0].spacing = spm;
 191
 192     if (ir->fourier_spacing > 0)
 193     {
 194         pme_lb->cut_spacing = ir->rcoulomb/ir->fourier_spacing;
 195     }
 196     else
 197     {
 198         pme_lb->cut_spacing = ir->rcoulomb/pme_lb->setup[0].spacing;
 199     }
 200
 201     pme_lb->stage = 0;
 202
 203     pme_lb->fastest  = 0;
 204     pme_lb->start    = 0;
 205     pme_lb->end      = 0;
 206     pme_lb->elimited = epmelblimNO;
 207
 208     *pme_lb_p = pme_lb;
 209 }
 210
 211 static gmx_bool pme_loadbal_increase_cutoff(pme_load_balancing_t  pme_lb,
 212                                             int                   pme_order,
 213                                             const gmx_domdec_t   *dd)
 214 {
 215     pme_setup_t *set;
 216     int          npmenodes_x, npmenodes_y;
 217     real         fac, sp;
 218     real         tmpr_coulomb, tmpr_vdw;
 219     int          d;
 220     gmx_bool     grid_ok;
 221
 222     /* Try to add a new setup with next larger cut-off to the list */
 223     pme_lb->n++;
 224     srenew(pme_lb->setup, pme_lb->n);
 225     set          = &pme_lb->setup[pme_lb->n-1];
 226     set->pmedata = NULL;
 227
 228     get_pme_nnodes(dd, &npmenodes_x, &npmenodes_y);
 229
 230     fac = 1;
 231     do
 232     {
 233         /* Avoid infinite while loop, which can occur at the minimum grid size.
 234          * Note that in practice load balancing will stop before this point.
 235          * The factor 2.1 allows for the extreme case in which only grids
 236          * of powers of 2 are allowed (the current code supports more grids).
 237          */
 238         if (fac > 2.1)
 239         {
 240             pme_lb->n--;
 241
 242             return FALSE;
 243         }
 244
 245         fac *= 1.01;
 246         clear_ivec(set->grid);
 247         sp = calc_grid(NULL, pme_lb->box_start,
 248                        fac*pme_lb->setup[pme_lb->cur].spacing,
 249                        &set->grid[XX],
 250                        &set->grid[YY],
 251                        &set->grid[ZZ]);
 252
 253         /* As here we can't easily check if one of the PME nodes
 254          * uses threading, we do a conservative grid check.
 255          * This means we can't use pme_order or less grid lines
 256          * per PME node along x, which is not a strong restriction.
 257          */
 258         gmx_pme_check_restrictions(pme_order,
 259                                    set->grid[XX], set->grid[YY], set->grid[ZZ],
 260                                    npmenodes_x, npmenodes_y,
 261                                    TRUE,
 262                                    FALSE,
 263                                    &grid_ok);
 264     }
 265     while (sp <= 1.001*pme_lb->setup[pme_lb->cur].spacing || !grid_ok);
 266
 267     set->rcut_coulomb = pme_lb->cut_spacing*sp;
 268     if (set->rcut_coulomb < pme_lb->rcut_coulomb_start)
 269     {
 270         /* This is unlikely, but can happen when e.g. continuing from
 271          * a checkpoint after equilibration where the box shrank a lot.
 272          * We want to avoid rcoulomb getting smaller than rvdw
 273          * and there might be more issues with decreasing rcoulomb.
 274          */
 275         set->rcut_coulomb = pme_lb->rcut_coulomb_start;
 276     }
 277
 278     if (pme_lb->cutoff_scheme == ecutsVERLET)
 279     {
 280         set->rlist        = set->rcut_coulomb + pme_lb->rbuf_coulomb;
 281         /* We dont use LR lists with Verlet, but this avoids if-statements in further checks */
 282         set->rlistlong    = set->rlist;
 283     }
 284     else
 285     {
 286         tmpr_coulomb          = set->rcut_coulomb + pme_lb->rbuf_coulomb;
 287         tmpr_vdw              = pme_lb->rcut_vdw + pme_lb->rbuf_vdw;
 288         set->rlist            = min(tmpr_coulomb, tmpr_vdw);
 289         set->rlistlong        = max(tmpr_coulomb, tmpr_vdw);
 290
 291         /* Set the long-range update frequency */
 292         if (set->rlist == set->rlistlong)
 293         {
 294             /* No long-range interactions if the short-/long-range cutoffs are identical */
 295             set->nstcalclr = 0;
 296         }
 297         else if (pme_lb->nstcalclr_start == 0 || pme_lb->nstcalclr_start == 1)
 298         {
 299             /* We were not doing long-range before, but now we are since rlist!=rlistlong */
 300             set->nstcalclr = 1;
 301         }
 302         else
 303         {
 304             /* We were already doing long-range interactions from the start */
 305             if (pme_lb->rcut_vdw > pme_lb->rcut_coulomb_start)
 306             {
 307                 /* We were originally doing long-range VdW-only interactions.
 308                  * If rvdw is still longer than rcoulomb we keep the original nstcalclr,
 309                  * but if the coulomb cutoff has become longer we should update the long-range
 310                  * part every step.
 311                  */
 312                 set->nstcalclr = (tmpr_vdw > tmpr_coulomb) ? pme_lb->nstcalclr_start : 1;
 313             }
 314             else
 315             {
 316                 /* We were not doing any long-range interaction from the start,
 317                  * since it is not possible to do twin-range coulomb for the PME interaction.
 318                  */
 319                 set->nstcalclr = 1;
 320             }
 321         }
 322     }
 323
 324     set->spacing      = sp;
 325     /* The grid efficiency is the size wrt a grid with uniform x/y/z spacing */
 326     set->grid_efficiency = 1;
 327     for (d = 0; d < DIM; d++)
 328     {
 329         set->grid_efficiency *= (set->grid[d]*sp)/norm(pme_lb->box_start[d]);
 330     }
 331     /* The Ewald coefficient is inversly proportional to the cut-off */
 332     set->ewaldcoeff_q =
 333         pme_lb->setup[0].ewaldcoeff_q*pme_lb->setup[0].rcut_coulomb/set->rcut_coulomb;
 334     /* We set ewaldcoeff_lj in set, even when LJ-PME is not used */
 335     set->ewaldcoeff_lj =
 336         pme_lb->setup[0].ewaldcoeff_lj*pme_lb->setup[0].rcut_coulomb/set->rcut_coulomb;
 337
 338     set->count   = 0;
 339     set->cycles  = 0;
 340
 341     if (debug)
 342     {
 343         fprintf(debug, "PME loadbal: grid %d %d %d, coulomb cutoff %f\n",
 344                 set->grid[XX], set->grid[YY], set->grid[ZZ], set->rcut_coulomb);
 345     }
 346     return TRUE;
 347 }
 348
 349 static void print_grid(FILE *fp_err, FILE *fp_log,
 350                        const char *pre,
 351                        const char *desc,
 352                        const pme_setup_t *set,
 353                        double cycles)
 354 {
 355     char buf[STRLEN], buft[STRLEN];
 356
 357     if (cycles >= 0)
 358     {
 359         sprintf(buft, ": %.1f M-cycles", cycles*1e-6);
 360     }
 361     else
 362     {
 363         buft[0] = '\0';
 364     }
 365     sprintf(buf, "%-11s%10s pme grid %d %d %d, coulomb cutoff %.3f%s",
 366             pre,
 367             desc, set->grid[XX], set->grid[YY], set->grid[ZZ], set->rcut_coulomb,
 368             buft);
 369     if (fp_err != NULL)
 370     {
 371         fprintf(fp_err, "\r%s\n", buf);
 372     }
 373     if (fp_log != NULL)
 374     {
 375         fprintf(fp_log, "%s\n", buf);
 376     }
 377 }
 378
 379 static int pme_loadbal_end(pme_load_balancing_t pme_lb)
 380 {
 381     /* In the initial stage only n is set; end is not set yet */
 382     if (pme_lb->end > 0)
 383     {
 384         return pme_lb->end;
 385     }
 386     else
 387     {
 388         return pme_lb->n;
 389     }
 390 }
 391
 392 static void print_loadbal_limited(FILE *fp_err, FILE *fp_log,
 393                                   gmx_int64_t step,
 394                                   pme_load_balancing_t pme_lb)
 395 {
 396     char buf[STRLEN], sbuf[22];
 397
 398     sprintf(buf, "step %4s: the %s limits the PME load balancing to a coulomb cut-off of %.3f",
 399             gmx_step_str(step, sbuf),
 400             pmelblim_str[pme_lb->elimited],
 401             pme_lb->setup[pme_loadbal_end(pme_lb)-1].rcut_coulomb);
 402     if (fp_err != NULL)
 403     {
 404         fprintf(fp_err, "\r%s\n", buf);
 405     }
 406     if (fp_log != NULL)
 407     {
 408         fprintf(fp_log, "%s\n", buf);
 409     }
 410 }
 411
 412 static void switch_to_stage1(pme_load_balancing_t pme_lb)
 413 {
 414     pme_lb->start = 0;
 415     while (pme_lb->start+1 < pme_lb->n &&
 416            (pme_lb->setup[pme_lb->start].count == 0 ||
 417             pme_lb->setup[pme_lb->start].cycles >
 418             pme_lb->setup[pme_lb->fastest].cycles*PME_LB_SLOW_FAC))
 419     {
 420         pme_lb->start++;
 421     }
 422     while (pme_lb->start > 0 && pme_lb->setup[pme_lb->start-1].cycles == 0)
 423     {
 424         pme_lb->start--;
 425     }
 426
 427     pme_lb->end = pme_lb->n;
 428     if (pme_lb->setup[pme_lb->end-1].count > 0 &&
 429         pme_lb->setup[pme_lb->end-1].cycles >
 430         pme_lb->setup[pme_lb->fastest].cycles*PME_LB_SLOW_FAC)
 431     {
 432         pme_lb->end--;
 433     }
 434
 435     pme_lb->stage = 1;
 436
 437     /* Next we want to choose setup pme_lb->start, but as we will increase
 438      * pme_ln->cur by one right after returning, we subtract 1 here.
 439      */
 440     pme_lb->cur = pme_lb->start - 1;
 441 }
 442
 443 gmx_bool pme_load_balance(pme_load_balancing_t        pme_lb,
 444                           t_commrec                  *cr,
 445                           FILE                       *fp_err,
 446                           FILE                       *fp_log,
 447                           t_inputrec                 *ir,
 448                           t_state                    *state,
 449                           double                      cycles,
 450                           interaction_const_t        *ic,
 451                           struct nonbonded_verlet_t  *nbv,
 452                           gmx_pme_t                  *pmedata,
 453                           gmx_int64_t                 step)
 454 {
 455     gmx_bool     OK;
 456     pme_setup_t *set;
 457     double       cycles_fast;
 458     char         buf[STRLEN], sbuf[22];
 459     real         rtab;
 460     gmx_bool     bUsesSimpleTables = TRUE;
 461
 462     if (pme_lb->stage == pme_lb->nstage)
 463     {
 464         return FALSE;
 465     }
 466
 467     if (PAR(cr))
 468     {
 469         gmx_sumd(1, &cycles, cr);
 470         cycles /= cr->nnodes;
 471     }
 472
 473     set = &pme_lb->setup[pme_lb->cur];
 474     set->count++;
 475
 476     rtab = ir->rlistlong + ir->tabext;
 477
 478     if (set->count % 2 == 1)
 479     {
 480         /* Skip the first cycle, because the first step after a switch
 481          * is much slower due to allocation and/or caching effects.
 482          */
 483         return TRUE;
 484     }
 485
 486     sprintf(buf, "step %4s: ", gmx_step_str(step, sbuf));
 487     print_grid(fp_err, fp_log, buf, "timed with", set, cycles);
 488
 489     if (set->count <= 2)
 490     {
 491         set->cycles = cycles;
 492     }
 493     else
 494     {
 495         if (cycles*PME_LB_ACCEL_TOL < set->cycles &&
 496             pme_lb->stage == pme_lb->nstage - 1)
 497         {
 498             /* The performance went up a lot (due to e.g. DD load balancing).
 499              * Add a stage, keep the minima, but rescan all setups.
 500              */
 501             pme_lb->nstage++;
 502
 503             if (debug)
 504             {
 505                 fprintf(debug, "The performance for grid %d %d %d went from %.3f to %.1f M-cycles, this is more than %f\n"
 506                         "Increased the number stages to %d"
 507                         " and ignoring the previous performance\n",
 508                         set->grid[XX], set->grid[YY], set->grid[ZZ],
 509                         cycles*1e-6, set->cycles*1e-6, PME_LB_ACCEL_TOL,
 510                         pme_lb->nstage);
 511             }
 512         }
 513         set->cycles = min(set->cycles, cycles);
 514     }
 515
 516     if (set->cycles < pme_lb->setup[pme_lb->fastest].cycles)
 517     {
 518         pme_lb->fastest = pme_lb->cur;
 519
 520         if (DOMAINDECOMP(cr))
 521         {
 522             /* We found a new fastest setting, ensure that with subsequent
 523              * shorter cut-off's the dynamic load balancing does not make
 524              * the use of the current cut-off impossible. This solution is
 525              * a trade-off, as the PME load balancing and DD domain size
 526              * load balancing can interact in complex ways.
 527              * With the Verlet kernels, DD load imbalance will usually be
 528              * mainly due to bonded interaction imbalance, which will often
 529              * quickly push the domain boundaries beyond the limit for the
 530              * optimal, PME load balanced, cut-off. But it could be that
 531              * better overal performance can be obtained with a slightly
 532              * shorter cut-off and better DD load balancing.
 533              */
 534             change_dd_dlb_cutoff_limit(cr);
 535         }
 536     }
 537     cycles_fast = pme_lb->setup[pme_lb->fastest].cycles;
 538
 539     /* Check in stage 0 if we should stop scanning grids.
 540      * Stop when the time is more than SLOW_FAC longer than the fastest.
 541      */
 542     if (pme_lb->stage == 0 && pme_lb->cur > 0 &&
 543         cycles > pme_lb->setup[pme_lb->fastest].cycles*PME_LB_SLOW_FAC)
 544     {
 545         pme_lb->n = pme_lb->cur + 1;
 546         /* Done with scanning, go to stage 1 */
 547         switch_to_stage1(pme_lb);
 548     }
 549
 550     if (pme_lb->stage == 0)
 551     {
 552         int gridsize_start;
 553
 554         gridsize_start = set->grid[XX]*set->grid[YY]*set->grid[ZZ];
 555
 556         do
 557         {
 558             if (pme_lb->cur+1 < pme_lb->n)
 559             {
 560                 /* We had already generated the next setup */
 561                 OK = TRUE;
 562             }
 563             else
 564             {
 565                 /* Find the next setup */
 566                 OK = pme_loadbal_increase_cutoff(pme_lb, ir->pme_order, cr->dd);
 567
 568                 if (!OK)
 569                 {
 570                     pme_lb->elimited = epmelblimPMEGRID;
 571                 }
 572             }
 573
 574             if (OK && ir->ePBC != epbcNONE)
 575             {
 576                 OK = (sqr(pme_lb->setup[pme_lb->cur+1].rlistlong)
 577                       <= max_cutoff2(ir->ePBC, state->box));
 578                 if (!OK)
 579                 {
 580                     pme_lb->elimited = epmelblimBOX;
 581                 }
 582             }
 583
 584             if (OK)
 585             {
 586                 pme_lb->cur++;
 587
 588                 if (DOMAINDECOMP(cr))
 589                 {
 590                     OK = change_dd_cutoff(cr, state, ir,
 591                                           pme_lb->setup[pme_lb->cur].rlistlong);
 592                     if (!OK)
 593                     {
 594                         /* Failed: do not use this setup */
 595                         pme_lb->cur--;
 596                         pme_lb->elimited = epmelblimDD;
 597                     }
 598                 }
 599             }
 600             if (!OK)
 601             {
 602                 /* We hit the upper limit for the cut-off,
 603                  * the setup should not go further than cur.
 604                  */
 605                 pme_lb->n = pme_lb->cur + 1;
 606                 print_loadbal_limited(fp_err, fp_log, step, pme_lb);
 607                 /* Switch to the next stage */
 608                 switch_to_stage1(pme_lb);
 609             }
 610         }
 611         while (OK &&
 612                !(pme_lb->setup[pme_lb->cur].grid[XX]*
 613                  pme_lb->setup[pme_lb->cur].grid[YY]*
 614                  pme_lb->setup[pme_lb->cur].grid[ZZ] <
 615                  gridsize_start*PME_LB_GRID_SCALE_FAC
 616                  &&
 617                  pme_lb->setup[pme_lb->cur].grid_efficiency <
 618                  pme_lb->setup[pme_lb->cur-1].grid_efficiency*PME_LB_GRID_EFFICIENCY_REL_FAC));
 619     }
 620
 621     if (pme_lb->stage > 0 && pme_lb->end == 1)
 622     {
 623         pme_lb->cur   = 0;
 624         pme_lb->stage = pme_lb->nstage;
 625     }
 626     else if (pme_lb->stage > 0 && pme_lb->end > 1)
 627     {
 628         /* If stage = nstage-1:
 629          *   scan over all setups, rerunning only those setups
 630          *   which are not much slower than the fastest
 631          * else:
 632          *   use the next setup
 633          */
 634         do
 635         {
 636             pme_lb->cur++;
 637             if (pme_lb->cur == pme_lb->end)
 638             {
 639                 pme_lb->stage++;
 640                 pme_lb->cur = pme_lb->start;
 641             }
 642         }
 643         while (pme_lb->stage == pme_lb->nstage - 1 &&
 644                pme_lb->setup[pme_lb->cur].count > 0 &&
 645                pme_lb->setup[pme_lb->cur].cycles > cycles_fast*PME_LB_SLOW_FAC);
 646
 647         if (pme_lb->stage == pme_lb->nstage)
 648         {
 649             /* We are done optimizing, use the fastest setup we found */
 650             pme_lb->cur = pme_lb->fastest;
 651         }
 652     }
 653
 654     if (DOMAINDECOMP(cr) && pme_lb->stage > 0)
 655     {
 656         OK = change_dd_cutoff(cr, state, ir, pme_lb->setup[pme_lb->cur].rlistlong);
 657         if (!OK)
 658         {
 659             /* Failsafe solution */
 660             if (pme_lb->cur > 1 && pme_lb->stage == pme_lb->nstage)
 661             {
 662                 pme_lb->stage--;
 663             }
 664             pme_lb->fastest  = 0;
 665             pme_lb->start    = 0;
 666             pme_lb->end      = pme_lb->cur;
 667             pme_lb->cur      = pme_lb->start;
 668             pme_lb->elimited = epmelblimDD;
 669             print_loadbal_limited(fp_err, fp_log, step, pme_lb);
 670         }
 671     }
 672
 673     /* Change the Coulomb cut-off and the PME grid */
 674
 675     set = &pme_lb->setup[pme_lb->cur];
 676
 677     ic->rcoulomb     = set->rcut_coulomb;
 678     ic->rlist        = set->rlist;
 679     ic->rlistlong    = set->rlistlong;
 680     ir->nstcalclr    = set->nstcalclr;
 681     ic->ewaldcoeff_q = set->ewaldcoeff_q;
 682     /* TODO: centralize the code that sets the potentials shifts */
 683     if (ic->coulomb_modifier == eintmodPOTSHIFT)
 684     {
 685         ic->sh_ewald = gmx_erfc(ic->ewaldcoeff_q*ic->rcoulomb);
 686     }
 687     if (EVDW_PME(ic->vdwtype))
 688     {
 689         /* We have PME for both Coulomb and VdW, set rvdw equal to rcoulomb */
 690         ic->rvdw            = set->rcut_coulomb;
 691         ic->ewaldcoeff_lj   = set->ewaldcoeff_lj;
 692         if (ic->vdw_modifier == eintmodPOTSHIFT)
 693         {
 694             real crc2;
 695
 696             ic->dispersion_shift.cpot = -pow(ic->rvdw, -6.0);
 697             ic->repulsion_shift.cpot  = -pow(ic->rvdw, -12.0);
 698             ic->sh_invrc6             = -ic->dispersion_shift.cpot;
 699             crc2                      = sqr(ic->ewaldcoeff_lj*ic->rvdw);
 700             ic->sh_lj_ewald           = (exp(-crc2)*(1 + crc2 + 0.5*crc2*crc2) - 1)*pow(ic->rvdw, -6.0);
 701         }
 702     }
 703
 704     bUsesSimpleTables = uses_simple_tables(ir->cutoff_scheme, nbv, 0);
 705     nbnxn_cuda_pme_loadbal_update_param(nbv, ic);
 706
 707     /* With tMPI + GPUs some ranks may be sharing GPU(s) and therefore
 708      * also sharing texture references. To keep the code simple, we don't
 709      * treat texture references as shared resources, but this means that
 710      * the coulomb_tab texture ref will get updated by multiple threads.
 711      * Hence, to ensure that the non-bonded kernels don't start before all
 712      * texture binding operations are finished, we need to wait for all ranks
 713      * to arrive here before continuing.
 714      *
 715      * Note that we could omit this barrier if GPUs are not shared (or
 716      * texture objects are used), but as this is initialization code, there
 717      * is not point in complicating things.
 718      */
 719 #ifdef GMX_THREAD_MPI
 720     if (PAR(cr) && use_GPU(nbv))
 721     {
 722         gmx_barrier(cr);
 723     }
 724 #endif  /* GMX_THREAD_MPI */
 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 }