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
 269     if (pme_lb->cutoff_scheme == ecutsVERLET)
 270     {
 271         set->rlist        = set->rcut_coulomb + pme_lb->rbuf_coulomb;
 272         /* We dont use LR lists with Verlet, but this avoids if-statements in further checks */
 273         set->rlistlong    = set->rlist;
 274     }
 275     else
 276     {
 277         tmpr_coulomb          = set->rcut_coulomb + pme_lb->rbuf_coulomb;
 278         tmpr_vdw              = pme_lb->rcut_vdw + pme_lb->rbuf_vdw;
 279         set->rlist            = min(tmpr_coulomb, tmpr_vdw);
 280         set->rlistlong        = max(tmpr_coulomb, tmpr_vdw);
 281
 282         /* Set the long-range update frequency */
 283         if (set->rlist == set->rlistlong)
 284         {
 285             /* No long-range interactions if the short-/long-range cutoffs are identical */
 286             set->nstcalclr = 0;
 287         }
 288         else if (pme_lb->nstcalclr_start == 0 || pme_lb->nstcalclr_start == 1)
 289         {
 290             /* We were not doing long-range before, but now we are since rlist!=rlistlong */
 291             set->nstcalclr = 1;
 292         }
 293         else
 294         {
 295             /* We were already doing long-range interactions from the start */
 296             if (pme_lb->rcut_vdw > pme_lb->rcut_coulomb_start)
 297             {
 298                 /* We were originally doing long-range VdW-only interactions.
 299                  * If rvdw is still longer than rcoulomb we keep the original nstcalclr,
 300                  * but if the coulomb cutoff has become longer we should update the long-range
 301                  * part every step.
 302                  */
 303                 set->nstcalclr = (tmpr_vdw > tmpr_coulomb) ? pme_lb->nstcalclr_start : 1;
 304             }
 305             else
 306             {
 307                 /* We were not doing any long-range interaction from the start,
 308                  * since it is not possible to do twin-range coulomb for the PME interaction.
 309                  */
 310                 set->nstcalclr = 1;
 311             }
 312         }
 313     }
 314
 315     set->spacing      = sp;
 316     /* The grid efficiency is the size wrt a grid with uniform x/y/z spacing */
 317     set->grid_efficiency = 1;
 318     for (d = 0; d < DIM; d++)
 319     {
 320         set->grid_efficiency *= (set->grid[d]*sp)/norm(pme_lb->box_start[d]);
 321     }
 322     /* The Ewald coefficient is inversly proportional to the cut-off */
 323     set->ewaldcoeff_q =
 324         pme_lb->setup[0].ewaldcoeff_q*pme_lb->setup[0].rcut_coulomb/set->rcut_coulomb;
 325     /* We set ewaldcoeff_lj in set, even when LJ-PME is not used */
 326     set->ewaldcoeff_lj =
 327         pme_lb->setup[0].ewaldcoeff_lj*pme_lb->setup[0].rcut_coulomb/set->rcut_coulomb;
 328
 329     set->count   = 0;
 330     set->cycles  = 0;
 331
 332     if (debug)
 333     {
 334         fprintf(debug, "PME loadbal: grid %d %d %d, coulomb cutoff %f\n",
 335                 set->grid[XX], set->grid[YY], set->grid[ZZ], set->rcut_coulomb);
 336     }
 337     return TRUE;
 338 }
 339
 340 static void print_grid(FILE *fp_err, FILE *fp_log,
 341                        const char *pre,
 342                        const char *desc,
 343                        const pme_setup_t *set,
 344                        double cycles)
 345 {
 346     char buf[STRLEN], buft[STRLEN];
 347
 348     if (cycles >= 0)
 349     {
 350         sprintf(buft, ": %.1f M-cycles", cycles*1e-6);
 351     }
 352     else
 353     {
 354         buft[0] = '\0';
 355     }
 356     sprintf(buf, "%-11s%10s pme grid %d %d %d, coulomb cutoff %.3f%s",
 357             pre,
 358             desc, set->grid[XX], set->grid[YY], set->grid[ZZ], set->rcut_coulomb,
 359             buft);
 360     if (fp_err != NULL)
 361     {
 362         fprintf(fp_err, "\r%s\n", buf);
 363     }
 364     if (fp_log != NULL)
 365     {
 366         fprintf(fp_log, "%s\n", buf);
 367     }
 368 }
 369
 370 static int pme_loadbal_end(pme_load_balancing_t pme_lb)
 371 {
 372     /* In the initial stage only n is set; end is not set yet */
 373     if (pme_lb->end > 0)
 374     {
 375         return pme_lb->end;
 376     }
 377     else
 378     {
 379         return pme_lb->n;
 380     }
 381 }
 382
 383 static void print_loadbal_limited(FILE *fp_err, FILE *fp_log,
 384                                   gmx_int64_t step,
 385                                   pme_load_balancing_t pme_lb)
 386 {
 387     char buf[STRLEN], sbuf[22];
 388
 389     sprintf(buf, "step %4s: the %s limits the PME load balancing to a coulomb cut-off of %.3f",
 390             gmx_step_str(step, sbuf),
 391             pmelblim_str[pme_lb->elimited],
 392             pme_lb->setup[pme_loadbal_end(pme_lb)-1].rcut_coulomb);
 393     if (fp_err != NULL)
 394     {
 395         fprintf(fp_err, "\r%s\n", buf);
 396     }
 397     if (fp_log != NULL)
 398     {
 399         fprintf(fp_log, "%s\n", buf);
 400     }
 401 }
 402
 403 static void switch_to_stage1(pme_load_balancing_t pme_lb)
 404 {
 405     pme_lb->start = 0;
 406     while (pme_lb->start+1 < pme_lb->n &&
 407            (pme_lb->setup[pme_lb->start].count == 0 ||
 408             pme_lb->setup[pme_lb->start].cycles >
 409             pme_lb->setup[pme_lb->fastest].cycles*PME_LB_SLOW_FAC))
 410     {
 411         pme_lb->start++;
 412     }
 413     while (pme_lb->start > 0 && pme_lb->setup[pme_lb->start-1].cycles == 0)
 414     {
 415         pme_lb->start--;
 416     }
 417
 418     pme_lb->end = pme_lb->n;
 419     if (pme_lb->setup[pme_lb->end-1].count > 0 &&
 420         pme_lb->setup[pme_lb->end-1].cycles >
 421         pme_lb->setup[pme_lb->fastest].cycles*PME_LB_SLOW_FAC)
 422     {
 423         pme_lb->end--;
 424     }
 425
 426     pme_lb->stage = 1;
 427
 428     /* Next we want to choose setup pme_lb->start, but as we will increase
 429      * pme_ln->cur by one right after returning, we subtract 1 here.
 430      */
 431     pme_lb->cur = pme_lb->start - 1;
 432 }
 433
 434 gmx_bool pme_load_balance(pme_load_balancing_t        pme_lb,
 435                           t_commrec                  *cr,
 436                           FILE                       *fp_err,
 437                           FILE                       *fp_log,
 438                           t_inputrec                 *ir,
 439                           t_state                    *state,
 440                           double                      cycles,
 441                           interaction_const_t        *ic,
 442                           struct nonbonded_verlet_t  *nbv,
 443                           gmx_pme_t                  *pmedata,
 444                           gmx_int64_t                 step)
 445 {
 446     gmx_bool     OK;
 447     pme_setup_t *set;
 448     double       cycles_fast;
 449     char         buf[STRLEN], sbuf[22];
 450     real         rtab;
 451     gmx_bool     bUsesSimpleTables = TRUE;
 452
 453     if (pme_lb->stage == pme_lb->nstage)
 454     {
 455         return FALSE;
 456     }
 457
 458     if (PAR(cr))
 459     {
 460         gmx_sumd(1, &cycles, cr);
 461         cycles /= cr->nnodes;
 462     }
 463
 464     set = &pme_lb->setup[pme_lb->cur];
 465     set->count++;
 466
 467     rtab = ir->rlistlong + ir->tabext;
 468
 469     if (set->count % 2 == 1)
 470     {
 471         /* Skip the first cycle, because the first step after a switch
 472          * is much slower due to allocation and/or caching effects.
 473          */
 474         return TRUE;
 475     }
 476
 477     sprintf(buf, "step %4s: ", gmx_step_str(step, sbuf));
 478     print_grid(fp_err, fp_log, buf, "timed with", set, cycles);
 479
 480     if (set->count <= 2)
 481     {
 482         set->cycles = cycles;
 483     }
 484     else
 485     {
 486         if (cycles*PME_LB_ACCEL_TOL < set->cycles &&
 487             pme_lb->stage == pme_lb->nstage - 1)
 488         {
 489             /* The performance went up a lot (due to e.g. DD load balancing).
 490              * Add a stage, keep the minima, but rescan all setups.
 491              */
 492             pme_lb->nstage++;
 493
 494             if (debug)
 495             {
 496                 fprintf(debug, "The performance for grid %d %d %d went from %.3f to %.1f M-cycles, this is more than %f\n"
 497                         "Increased the number stages to %d"
 498                         " and ignoring the previous performance\n",
 499                         set->grid[XX], set->grid[YY], set->grid[ZZ],
 500                         cycles*1e-6, set->cycles*1e-6, PME_LB_ACCEL_TOL,
 501                         pme_lb->nstage);
 502             }
 503         }
 504         set->cycles = min(set->cycles, cycles);
 505     }
 506
 507     if (set->cycles < pme_lb->setup[pme_lb->fastest].cycles)
 508     {
 509         pme_lb->fastest = pme_lb->cur;
 510
 511         if (DOMAINDECOMP(cr))
 512         {
 513             /* We found a new fastest setting, ensure that with subsequent
 514              * shorter cut-off's the dynamic load balancing does not make
 515              * the use of the current cut-off impossible. This solution is
 516              * a trade-off, as the PME load balancing and DD domain size
 517              * load balancing can interact in complex ways.
 518              * With the Verlet kernels, DD load imbalance will usually be
 519              * mainly due to bonded interaction imbalance, which will often
 520              * quickly push the domain boundaries beyond the limit for the
 521              * optimal, PME load balanced, cut-off. But it could be that
 522              * better overal performance can be obtained with a slightly
 523              * shorter cut-off and better DD load balancing.
 524              */
 525             change_dd_dlb_cutoff_limit(cr);
 526         }
 527     }
 528     cycles_fast = pme_lb->setup[pme_lb->fastest].cycles;
 529
 530     /* Check in stage 0 if we should stop scanning grids.
 531      * Stop when the time is more than SLOW_FAC longer than the fastest.
 532      */
 533     if (pme_lb->stage == 0 && pme_lb->cur > 0 &&
 534         cycles > pme_lb->setup[pme_lb->fastest].cycles*PME_LB_SLOW_FAC)
 535     {
 536         pme_lb->n = pme_lb->cur + 1;
 537         /* Done with scanning, go to stage 1 */
 538         switch_to_stage1(pme_lb);
 539     }
 540
 541     if (pme_lb->stage == 0)
 542     {
 543         int gridsize_start;
 544
 545         gridsize_start = set->grid[XX]*set->grid[YY]*set->grid[ZZ];
 546
 547         do
 548         {
 549             if (pme_lb->cur+1 < pme_lb->n)
 550             {
 551                 /* We had already generated the next setup */
 552                 OK = TRUE;
 553             }
 554             else
 555             {
 556                 /* Find the next setup */
 557                 OK = pme_loadbal_increase_cutoff(pme_lb, ir->pme_order, cr->dd);
 558
 559                 if (!OK)
 560                 {
 561                     pme_lb->elimited = epmelblimPMEGRID;
 562                 }
 563             }
 564
 565             if (OK && ir->ePBC != epbcNONE)
 566             {
 567                 OK = (sqr(pme_lb->setup[pme_lb->cur+1].rlistlong)
 568                       <= max_cutoff2(ir->ePBC, state->box));
 569                 if (!OK)
 570                 {
 571                     pme_lb->elimited = epmelblimBOX;
 572                 }
 573             }
 574
 575             if (OK)
 576             {
 577                 pme_lb->cur++;
 578
 579                 if (DOMAINDECOMP(cr))
 580                 {
 581                     OK = change_dd_cutoff(cr, state, ir,
 582                                           pme_lb->setup[pme_lb->cur].rlistlong);
 583                     if (!OK)
 584                     {
 585                         /* Failed: do not use this setup */
 586                         pme_lb->cur--;
 587                         pme_lb->elimited = epmelblimDD;
 588                     }
 589                 }
 590             }
 591             if (!OK)
 592             {
 593                 /* We hit the upper limit for the cut-off,
 594                  * the setup should not go further than cur.
 595                  */
 596                 pme_lb->n = pme_lb->cur + 1;
 597                 print_loadbal_limited(fp_err, fp_log, step, pme_lb);
 598                 /* Switch to the next stage */
 599                 switch_to_stage1(pme_lb);
 600             }
 601         }
 602         while (OK &&
 603                !(pme_lb->setup[pme_lb->cur].grid[XX]*
 604                  pme_lb->setup[pme_lb->cur].grid[YY]*
 605                  pme_lb->setup[pme_lb->cur].grid[ZZ] <
 606                  gridsize_start*PME_LB_GRID_SCALE_FAC
 607                  &&
 608                  pme_lb->setup[pme_lb->cur].grid_efficiency <
 609                  pme_lb->setup[pme_lb->cur-1].grid_efficiency*PME_LB_GRID_EFFICIENCY_REL_FAC));
 610     }
 611
 612     if (pme_lb->stage > 0 && pme_lb->end == 1)
 613     {
 614         pme_lb->cur   = 0;
 615         pme_lb->stage = pme_lb->nstage;
 616     }
 617     else if (pme_lb->stage > 0 && pme_lb->end > 1)
 618     {
 619         /* If stage = nstage-1:
 620          *   scan over all setups, rerunning only those setups
 621          *   which are not much slower than the fastest
 622          * else:
 623          *   use the next setup
 624          */
 625         do
 626         {
 627             pme_lb->cur++;
 628             if (pme_lb->cur == pme_lb->end)
 629             {
 630                 pme_lb->stage++;
 631                 pme_lb->cur = pme_lb->start;
 632             }
 633         }
 634         while (pme_lb->stage == pme_lb->nstage - 1 &&
 635                pme_lb->setup[pme_lb->cur].count > 0 &&
 636                pme_lb->setup[pme_lb->cur].cycles > cycles_fast*PME_LB_SLOW_FAC);
 637
 638         if (pme_lb->stage == pme_lb->nstage)
 639         {
 640             /* We are done optimizing, use the fastest setup we found */
 641             pme_lb->cur = pme_lb->fastest;
 642         }
 643     }
 644
 645     if (DOMAINDECOMP(cr) && pme_lb->stage > 0)
 646     {
 647         OK = change_dd_cutoff(cr, state, ir, pme_lb->setup[pme_lb->cur].rlistlong);
 648         if (!OK)
 649         {
 650             /* Failsafe solution */
 651             if (pme_lb->cur > 1 && pme_lb->stage == pme_lb->nstage)
 652             {
 653                 pme_lb->stage--;
 654             }
 655             pme_lb->fastest  = 0;
 656             pme_lb->start    = 0;
 657             pme_lb->end      = pme_lb->cur;
 658             pme_lb->cur      = pme_lb->start;
 659             pme_lb->elimited = epmelblimDD;
 660             print_loadbal_limited(fp_err, fp_log, step, pme_lb);
 661         }
 662     }
 663
 664     /* Change the Coulomb cut-off and the PME grid */
 665
 666     set = &pme_lb->setup[pme_lb->cur];
 667
 668     ic->rcoulomb     = set->rcut_coulomb;
 669     ic->rlist        = set->rlist;
 670     ic->rlistlong    = set->rlistlong;
 671     ir->nstcalclr    = set->nstcalclr;
 672     ic->ewaldcoeff_q = set->ewaldcoeff_q;
 673     /* TODO: centralize the code that sets the potentials shifts */
 674     if (ic->coulomb_modifier == eintmodPOTSHIFT)
 675     {
 676         ic->sh_ewald = gmx_erfc(ic->ewaldcoeff_q*ic->rcoulomb);
 677     }
 678     if (EVDW_PME(ic->vdwtype))
 679     {
 680         /* We have PME for both Coulomb and VdW, set rvdw equal to rcoulomb */
 681         ic->rvdw            = set->rcut_coulomb;
 682         ic->ewaldcoeff_lj   = set->ewaldcoeff_lj;
 683         if (ic->vdw_modifier == eintmodPOTSHIFT)
 684         {
 685             real crc2;
 686
 687             ic->dispersion_shift.cpot = -pow(ic->rvdw, -6.0);
 688             ic->repulsion_shift.cpot  = -pow(ic->rvdw, -12.0);
 689             ic->sh_invrc6             = -ic->dispersion_shift.cpot;
 690             crc2                      = sqr(ic->ewaldcoeff_lj*ic->rvdw);
 691             ic->sh_lj_ewald           = (exp(-crc2)*(1 + crc2 + 0.5*crc2*crc2) - 1)*pow(ic->rvdw, -6.0);
 692         }
 693     }
 694
 695     bUsesSimpleTables = uses_simple_tables(ir->cutoff_scheme, nbv, 0);
 696     nbnxn_cuda_pme_loadbal_update_param(nbv, ic);
 697
 698     /* With tMPI + GPUs some ranks may be sharing GPU(s) and therefore
 699      * also sharing texture references. To keep the code simple, we don't
 700      * treat texture references as shared resources, but this means that
 701      * the coulomb_tab texture ref will get updated by multiple threads.
 702      * Hence, to ensure that the non-bonded kernels don't start before all
 703      * texture binding operations are finished, we need to wait for all ranks
 704      * to arrive here before continuing.
 705      *
 706      * Note that we could omit this barrier if GPUs are not shared (or
 707      * texture objects are used), but as this is initialization code, there
 708      * is not point in complicating things.
 709      */
 710 #ifdef GMX_THREAD_MPI
 711     if (PAR(cr) && use_GPU(nbv))
 712     {
 713         gmx_barrier(cr);
 714     }
 715 #endif  /* GMX_THREAD_MPI */
 716
 717     /* Usually we won't need the simple tables with GPUs.
 718      * But we do with hybrid acceleration and with free energy.
 719      * To avoid bugs, we always re-initialize the simple tables here.
 720      */
 721     init_interaction_const_tables(NULL, ic, bUsesSimpleTables, rtab);
 722
 723     if (cr->duty & DUTY_PME)
 724     {
 725         if (pme_lb->setup[pme_lb->cur].pmedata == NULL)
 726         {
 727             /* Generate a new PME data structure,
 728              * copying part of the old pointers.
 729              */
 730             gmx_pme_reinit(&set->pmedata,
 731                            cr, pme_lb->setup[0].pmedata, ir,
 732                            set->grid);
 733         }
 734         *pmedata = set->pmedata;
 735     }
 736     else
 737     {
 738         /* Tell our PME-only node to switch grid */
 739         gmx_pme_send_switchgrid(cr, set->grid, set->ewaldcoeff_q, set->ewaldcoeff_lj);
 740     }
 741
 742     if (debug)
 743     {
 744         print_grid(NULL, debug, "", "switched to", set, -1);
 745     }
 746
 747     if (pme_lb->stage == pme_lb->nstage)
 748     {
 749         print_grid(fp_err, fp_log, "", "optimal", set, -1);
 750     }
 751
 752     return TRUE;
 753 }
 754
 755 void restart_pme_loadbal(pme_load_balancing_t pme_lb, int n)
 756 {
 757     pme_lb->nstage += n;
 758 }
 759
 760 static int pme_grid_points(const pme_setup_t *setup)
 761 {
 762     return setup->grid[XX]*setup->grid[YY]*setup->grid[ZZ];
 763 }
 764
 765 static real pme_loadbal_rlist(const pme_setup_t *setup)
 766 {
 767     /* With the group cut-off scheme we can have twin-range either
 768      * for Coulomb or for VdW, so we need a check here.
 769      * With the Verlet cut-off scheme rlist=rlistlong.
 770      */
 771     if (setup->rcut_coulomb > setup->rlist)
 772     {
 773         return setup->rlistlong;
 774     }
 775     else
 776     {
 777         return setup->rlist;
 778     }
 779 }
 780
 781 static void print_pme_loadbal_setting(FILE              *fplog,
 782                                       char              *name,
 783                                       const pme_setup_t *setup)
 784 {
 785     fprintf(fplog,
 786             "   %-7s %6.3f nm %6.3f nm     %3d %3d %3d   %5.3f nm  %5.3f nm\n",
 787             name,
 788             setup->rcut_coulomb, pme_loadbal_rlist(setup),
 789             setup->grid[XX], setup->grid[YY], setup->grid[ZZ],
 790             setup->spacing, 1/setup->ewaldcoeff_q);
 791 }
 792
 793 static void print_pme_loadbal_settings(pme_load_balancing_t pme_lb,
 794                                        t_commrec           *cr,
 795                                        FILE                *fplog,
 796                                        gmx_bool             bNonBondedOnGPU)
 797 {
 798     double pp_ratio, grid_ratio;
 799
 800     pp_ratio   = pow(pme_loadbal_rlist(&pme_lb->setup[pme_lb->cur])/pme_loadbal_rlist(&pme_lb->setup[0]), 3.0);
 801     grid_ratio = pme_grid_points(&pme_lb->setup[pme_lb->cur])/
 802         (double)pme_grid_points(&pme_lb->setup[0]);
 803
 804     fprintf(fplog, "\n");
 805     fprintf(fplog, "       P P   -   P M E   L O A D   B A L A N C I N G\n");
 806     fprintf(fplog, "\n");
 807     /* Here we only warn when the optimal setting is the last one */
 808     if (pme_lb->elimited != epmelblimNO &&
 809         pme_lb->cur == pme_loadbal_end(pme_lb)-1)
 810     {
 811         fprintf(fplog, " NOTE: The PP/PME load balancing was limited by the %s,\n",
 812                 pmelblim_str[pme_lb->elimited]);
 813         fprintf(fplog, "       you might not have reached a good load balance.\n");
 814         if (pme_lb->elimited == epmelblimDD)
 815         {
 816             fprintf(fplog, "       Try different mdrun -dd settings or lower the -dds value.\n");
 817         }
 818         fprintf(fplog, "\n");
 819     }
 820     fprintf(fplog, " PP/PME load balancing changed the cut-off and PME settings:\n");
 821     fprintf(fplog, "           particle-particle                    PME\n");
 822     fprintf(fplog, "            rcoulomb  rlist            grid      spacing   1/beta\n");
 823     print_pme_loadbal_setting(fplog, "initial", &pme_lb->setup[0]);
 824     print_pme_loadbal_setting(fplog, "final", &pme_lb->setup[pme_lb->cur]);
 825     fprintf(fplog, " cost-ratio           %4.2f             %4.2f\n",
 826             pp_ratio, grid_ratio);
 827     fprintf(fplog, " (note that these numbers concern only part of the total PP and PME load)\n");
 828
 829     if (pp_ratio > 1.5 && !bNonBondedOnGPU)
 830     {
 831         md_print_warn(cr, fplog,
 832                       "NOTE: PME load balancing increased the non-bonded workload by more than 50%%.\n"
 833                       "      For better performance, use (more) PME ranks (mdrun -npme),\n"
 834                       "      or if you are beyond the scaling limit, use fewer total ranks (or nodes).\n");
 835     }
 836     else
 837     {
 838         fprintf(fplog, "\n");
 839     }
 840 }
 841
 842 void pme_loadbal_done(pme_load_balancing_t pme_lb,
 843                       t_commrec *cr, FILE *fplog,
 844                       gmx_bool bNonBondedOnGPU)
 845 {
 846     if (fplog != NULL && (pme_lb->cur > 0 || pme_lb->elimited != epmelblimNO))
 847     {
 848         print_pme_loadbal_settings(pme_lb, cr, fplog, bNonBondedOnGPU);
 849     }
 850
 851     /* TODO: Here we should free all pointers in pme_lb,
 852      * but as it contains pme data structures,
 853      * we need to first make pme.c free all data.
 854      */
 855 }