src/programs/mdrun/pme_loadbal.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013,2014, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 #ifdef HAVE_CONFIG_H
  36 #include <config.h>
  37 #endif
  38
  39 #include "gromacs/utility/smalloc.h"
  40 #include "types/commrec.h"
  41 #include "network.h"
  42 #include "calcgrid.h"
  43 #include "pme.h"
  44 #include "vec.h"
  45 #include "domdec.h"
  46 #include "nbnxn_cuda_data_mgmt.h"
  47 #include "force.h"
  48 #include "macros.h"
  49 #include "md_logging.h"
  50 #include "pme_loadbal.h"
  51
  52 /* Parameters and setting for one PP-PME setup */
  53 typedef struct {
  54     real      rcut_coulomb;    /* Coulomb cut-off                              */
  55     real      rlist;           /* pair-list cut-off                            */
  56     real      rlistlong;       /* LR pair-list cut-off                         */
  57     int       nstcalclr;       /* frequency of evaluating long-range forces for group scheme */
  58     real      spacing;         /* (largest) PME grid spacing                   */
  59     ivec      grid;            /* the PME grid dimensions                      */
  60     real      grid_efficiency; /* ineffiency factor for non-uniform grids <= 1 */
  61     real      ewaldcoeff_q;    /* Electrostatic Ewald coefficient            */
  62     real      ewaldcoeff_lj;   /* LJ Ewald coefficient, only for the call to send_switchgrid */
  63     gmx_pme_t pmedata;         /* the data structure used in the PME code      */
  64     int       count;           /* number of times this setup has been timed    */
  65     double    cycles;          /* the fastest time for this setup in cycles    */
  66 } pme_setup_t;
  67
  68 /* In the initial scan, step by grids that are at least a factor 0.8 coarser */
  69 #define PME_LB_GRID_SCALE_FAC  0.8
  70 /* In the initial scan, try to skip grids with uneven x/y/z spacing,
  71  * checking if the "efficiency" is more than 5% worse than the previous grid.
  72  */
  73 #define PME_LB_GRID_EFFICIENCY_REL_FAC  1.05
  74 /* Rerun up till 12% slower setups than the fastest up till now */
  75 #define PME_LB_SLOW_FAC  1.12
  76 /* If setups get more than 2% faster, do another round to avoid
  77  * choosing a slower setup due to acceleration or fluctuations.
  78  */
  79 #define PME_LB_ACCEL_TOL 1.02
  80
  81 enum {
  82     epmelblimNO, epmelblimBOX, epmelblimDD, epmelblimPMEGRID, epmelblimNR
  83 };
  84
  85 const char *pmelblim_str[epmelblimNR] =
  86 { "no", "box size", "domain decompostion", "PME grid restriction" };
  87
  88 struct pme_load_balancing {
  89     int          nstage;             /* the current maximum number of stages */
  90
  91     real         cut_spacing;        /* the minimum cutoff / PME grid spacing ratio */
  92     real         rcut_vdw;           /* Vdw cutoff (does not change) */
  93     real         rcut_coulomb_start; /* Initial electrostatics cutoff */
  94     int          nstcalclr_start;    /* Initial electrostatics cutoff */
  95     real         rbuf_coulomb;       /* the pairlist buffer size */
  96     real         rbuf_vdw;           /* the pairlist buffer size */
  97     matrix       box_start;          /* the initial simulation box */
  98     int          n;                  /* the count of setup as well as the allocation size */
  99     pme_setup_t *setup;              /* the PME+cutoff setups */
 100     int          cur;                /* the current setup */
 101     int          fastest;            /* fastest setup up till now */
 102     int          start;              /* start of setup range to consider in stage>0 */
 103     int          end;                /* end   of setup range to consider in stage>0 */
 104     int          elimited;           /* was the balancing limited, uses enum above */
 105     int          cutoff_scheme;      /* Verlet or group cut-offs */
 106
 107     int          stage;              /* the current stage */
 108 };
 109
 110 void pme_loadbal_init(pme_load_balancing_t *pme_lb_p,
 111                       const t_inputrec *ir, matrix box,
 112                       const interaction_const_t *ic,
 113                       gmx_pme_t pmedata)
 114 {
 115     pme_load_balancing_t pme_lb;
 116     real                 spm, sp;
 117     int                  d;
 118
 119     snew(pme_lb, 1);
 120
 121     /* Any number of stages >= 2 is supported */
 122     pme_lb->nstage   = 2;
 123
 124     pme_lb->cutoff_scheme = ir->cutoff_scheme;
 125
 126     if (pme_lb->cutoff_scheme == ecutsVERLET)
 127     {
 128         pme_lb->rbuf_coulomb = ic->rlist - ic->rcoulomb;
 129         pme_lb->rbuf_vdw     = pme_lb->rbuf_coulomb;
 130     }
 131     else
 132     {
 133         if (ic->rcoulomb > ic->rlist)
 134         {
 135             pme_lb->rbuf_coulomb = ic->rlistlong - ic->rcoulomb;
 136         }
 137         else
 138         {
 139             pme_lb->rbuf_coulomb = ic->rlist - ic->rcoulomb;
 140         }
 141         if (ic->rvdw > ic->rlist)
 142         {
 143             pme_lb->rbuf_vdw = ic->rlistlong - ic->rvdw;
 144         }
 145         else
 146         {
 147             pme_lb->rbuf_vdw = ic->rlist - ic->rvdw;
 148         }
 149     }
 150
 151     copy_mat(box, pme_lb->box_start);
 152     if (ir->ePBC == epbcXY && ir->nwall == 2)
 153     {
 154         svmul(ir->wall_ewald_zfac, pme_lb->box_start[ZZ], pme_lb->box_start[ZZ]);
 155     }
 156
 157     pme_lb->n = 1;
 158     snew(pme_lb->setup, pme_lb->n);
 159
 160     pme_lb->rcut_vdw                 = ic->rvdw;
 161     pme_lb->rcut_coulomb_start       = ir->rcoulomb;
 162     pme_lb->nstcalclr_start          = ir->nstcalclr;
 163
 164     pme_lb->cur                      = 0;
 165     pme_lb->setup[0].rcut_coulomb    = ic->rcoulomb;
 166     pme_lb->setup[0].rlist           = ic->rlist;
 167     pme_lb->setup[0].rlistlong       = ic->rlistlong;
 168     pme_lb->setup[0].nstcalclr       = ir->nstcalclr;
 169     pme_lb->setup[0].grid[XX]        = ir->nkx;
 170     pme_lb->setup[0].grid[YY]        = ir->nky;
 171     pme_lb->setup[0].grid[ZZ]        = ir->nkz;
 172     pme_lb->setup[0].ewaldcoeff_q    = ic->ewaldcoeff_q;
 173     pme_lb->setup[0].ewaldcoeff_lj   = ic->ewaldcoeff_lj;
 174
 175     pme_lb->setup[0].pmedata  = pmedata;
 176
 177     spm = 0;
 178     for (d = 0; d < DIM; d++)
 179     {
 180         sp = norm(pme_lb->box_start[d])/pme_lb->setup[0].grid[d];
 181         if (sp > spm)
 182         {
 183             spm = sp;
 184         }
 185     }
 186     pme_lb->setup[0].spacing = spm;
 187
 188     if (ir->fourier_spacing > 0)
 189     {
 190         pme_lb->cut_spacing = ir->rcoulomb/ir->fourier_spacing;
 191     }
 192     else
 193     {
 194         pme_lb->cut_spacing = ir->rcoulomb/pme_lb->setup[0].spacing;
 195     }
 196
 197     pme_lb->stage = 0;
 198
 199     pme_lb->fastest  = 0;
 200     pme_lb->start    = 0;
 201     pme_lb->end      = 0;
 202     pme_lb->elimited = epmelblimNO;
 203
 204     *pme_lb_p = pme_lb;
 205 }
 206
 207 static gmx_bool pme_loadbal_increase_cutoff(pme_load_balancing_t  pme_lb,
 208                                             int                   pme_order,
 209                                             const gmx_domdec_t   *dd)
 210 {
 211     pme_setup_t *set;
 212     int          npmenodes_x, npmenodes_y;
 213     real         fac, sp;
 214     real         tmpr_coulomb, tmpr_vdw;
 215     int          d;
 216     gmx_bool     grid_ok;
 217
 218     /* Try to add a new setup with next larger cut-off to the list */
 219     pme_lb->n++;
 220     srenew(pme_lb->setup, pme_lb->n);
 221     set          = &pme_lb->setup[pme_lb->n-1];
 222     set->pmedata = NULL;
 223
 224     get_pme_nnodes(dd, &npmenodes_x, &npmenodes_y);
 225
 226     fac = 1;
 227     do
 228     {
 229         /* Avoid infinite while loop, which can occur at the minimum grid size.
 230          * Note that in practice load balancing will stop before this point.
 231          * The factor 2.1 allows for the extreme case in which only grids
 232          * of powers of 2 are allowed (the current code supports more grids).
 233          */
 234         if (fac > 2.1)
 235         {
 236             pme_lb->n--;
 237
 238             return FALSE;
 239         }
 240
 241         fac *= 1.01;
 242         clear_ivec(set->grid);
 243         sp = calc_grid(NULL, pme_lb->box_start,
 244                        fac*pme_lb->setup[pme_lb->cur].spacing,
 245                        &set->grid[XX],
 246                        &set->grid[YY],
 247                        &set->grid[ZZ]);
 248
 249         /* As here we can't easily check if one of the PME nodes
 250          * uses threading, we do a conservative grid check.
 251          * This means we can't use pme_order or less grid lines
 252          * per PME node along x, which is not a strong restriction.
 253          */
 254         gmx_pme_check_restrictions(pme_order,
 255                                    set->grid[XX], set->grid[YY], set->grid[ZZ],
 256                                    npmenodes_x, npmenodes_y,
 257                                    TRUE,
 258                                    FALSE,
 259                                    &grid_ok);
 260     }
 261     while (sp <= 1.001*pme_lb->setup[pme_lb->cur].spacing || !grid_ok);
 262
 263     set->rcut_coulomb = pme_lb->cut_spacing*sp;
 264
 265     if (pme_lb->cutoff_scheme == ecutsVERLET)
 266     {
 267         set->rlist        = set->rcut_coulomb + pme_lb->rbuf_coulomb;
 268         /* We dont use LR lists with Verlet, but this avoids if-statements in further checks */
 269         set->rlistlong    = set->rlist;
 270     }
 271     else
 272     {
 273         tmpr_coulomb          = set->rcut_coulomb + pme_lb->rbuf_coulomb;
 274         tmpr_vdw              = pme_lb->rcut_vdw + pme_lb->rbuf_vdw;
 275         set->rlist            = min(tmpr_coulomb, tmpr_vdw);
 276         set->rlistlong        = max(tmpr_coulomb, tmpr_vdw);
 277
 278         /* Set the long-range update frequency */
 279         if (set->rlist == set->rlistlong)
 280         {
 281             /* No long-range interactions if the short-/long-range cutoffs are identical */
 282             set->nstcalclr = 0;
 283         }
 284         else if (pme_lb->nstcalclr_start == 0 || pme_lb->nstcalclr_start == 1)
 285         {
 286             /* We were not doing long-range before, but now we are since rlist!=rlistlong */
 287             set->nstcalclr = 1;
 288         }
 289         else
 290         {
 291             /* We were already doing long-range interactions from the start */
 292             if (pme_lb->rcut_vdw > pme_lb->rcut_coulomb_start)
 293             {
 294                 /* We were originally doing long-range VdW-only interactions.
 295                  * If rvdw is still longer than rcoulomb we keep the original nstcalclr,
 296                  * but if the coulomb cutoff has become longer we should update the long-range
 297                  * part every step.
 298                  */
 299                 set->nstcalclr = (tmpr_vdw > tmpr_coulomb) ? pme_lb->nstcalclr_start : 1;
 300             }
 301             else
 302             {
 303                 /* We were not doing any long-range interaction from the start,
 304                  * since it is not possible to do twin-range coulomb for the PME interaction.
 305                  */
 306                 set->nstcalclr = 1;
 307             }
 308         }
 309     }
 310
 311     set->spacing      = sp;
 312     /* The grid efficiency is the size wrt a grid with uniform x/y/z spacing */
 313     set->grid_efficiency = 1;
 314     for (d = 0; d < DIM; d++)
 315     {
 316         set->grid_efficiency *= (set->grid[d]*sp)/norm(pme_lb->box_start[d]);
 317     }
 318     /* The Ewald coefficient is inversly proportional to the cut-off */
 319     set->ewaldcoeff_q =
 320         pme_lb->setup[0].ewaldcoeff_q*pme_lb->setup[0].rcut_coulomb/set->rcut_coulomb;
 321     /* We set ewaldcoeff_lj in set, even when LJ-PME is not used */
 322     set->ewaldcoeff_lj =
 323         pme_lb->setup[0].ewaldcoeff_lj*pme_lb->setup[0].rcut_coulomb/set->rcut_coulomb;
 324
 325     set->count   = 0;
 326     set->cycles  = 0;
 327
 328     if (debug)
 329     {
 330         fprintf(debug, "PME loadbal: grid %d %d %d, coulomb cutoff %f\n",
 331                 set->grid[XX], set->grid[YY], set->grid[ZZ], set->rcut_coulomb);
 332     }
 333     return TRUE;
 334 }
 335
 336 static void print_grid(FILE *fp_err, FILE *fp_log,
 337                        const char *pre,
 338                        const char *desc,
 339                        const pme_setup_t *set,
 340                        double cycles)
 341 {
 342     char buf[STRLEN], buft[STRLEN];
 343
 344     if (cycles >= 0)
 345     {
 346         sprintf(buft, ": %.1f M-cycles", cycles*1e-6);
 347     }
 348     else
 349     {
 350         buft[0] = '\0';
 351     }
 352     sprintf(buf, "%-11s%10s pme grid %d %d %d, coulomb cutoff %.3f%s",
 353             pre,
 354             desc, set->grid[XX], set->grid[YY], set->grid[ZZ], set->rcut_coulomb,
 355             buft);
 356     if (fp_err != NULL)
 357     {
 358         fprintf(fp_err, "\r%s\n", buf);
 359     }
 360     if (fp_log != NULL)
 361     {
 362         fprintf(fp_log, "%s\n", buf);
 363     }
 364 }
 365
 366 static int pme_loadbal_end(pme_load_balancing_t pme_lb)
 367 {
 368     /* In the initial stage only n is set; end is not set yet */
 369     if (pme_lb->end > 0)
 370     {
 371         return pme_lb->end;
 372     }
 373     else
 374     {
 375         return pme_lb->n;
 376     }
 377 }
 378
 379 static void print_loadbal_limited(FILE *fp_err, FILE *fp_log,
 380                                   gmx_int64_t step,
 381                                   pme_load_balancing_t pme_lb)
 382 {
 383     char buf[STRLEN], sbuf[22];
 384
 385     sprintf(buf, "step %4s: the %s limits the PME load balancing to a coulomb cut-off of %.3f",
 386             gmx_step_str(step, sbuf),
 387             pmelblim_str[pme_lb->elimited],
 388             pme_lb->setup[pme_loadbal_end(pme_lb)-1].rcut_coulomb);
 389     if (fp_err != NULL)
 390     {
 391         fprintf(fp_err, "\r%s\n", buf);
 392     }
 393     if (fp_log != NULL)
 394     {
 395         fprintf(fp_log, "%s\n", buf);
 396     }
 397 }
 398
 399 static void switch_to_stage1(pme_load_balancing_t pme_lb)
 400 {
 401     pme_lb->start = 0;
 402     while (pme_lb->start+1 < pme_lb->n &&
 403            (pme_lb->setup[pme_lb->start].count == 0 ||
 404             pme_lb->setup[pme_lb->start].cycles >
 405             pme_lb->setup[pme_lb->fastest].cycles*PME_LB_SLOW_FAC))
 406     {
 407         pme_lb->start++;
 408     }
 409     while (pme_lb->start > 0 && pme_lb->setup[pme_lb->start-1].cycles == 0)
 410     {
 411         pme_lb->start--;
 412     }
 413
 414     pme_lb->end = pme_lb->n;
 415     if (pme_lb->setup[pme_lb->end-1].count > 0 &&
 416         pme_lb->setup[pme_lb->end-1].cycles >
 417         pme_lb->setup[pme_lb->fastest].cycles*PME_LB_SLOW_FAC)
 418     {
 419         pme_lb->end--;
 420     }
 421
 422     pme_lb->stage = 1;
 423
 424     /* Next we want to choose setup pme_lb->start, but as we will increase
 425      * pme_ln->cur by one right after returning, we subtract 1 here.
 426      */
 427     pme_lb->cur = pme_lb->start - 1;
 428 }
 429
 430 gmx_bool pme_load_balance(pme_load_balancing_t pme_lb,
 431                           t_commrec           *cr,
 432                           FILE                *fp_err,
 433                           FILE                *fp_log,
 434                           t_inputrec          *ir,
 435                           t_state             *state,
 436                           double               cycles,
 437                           interaction_const_t *ic,
 438                           nonbonded_verlet_t  *nbv,
 439                           gmx_pme_t           *pmedata,
 440                           gmx_int64_t          step)
 441 {
 442     gmx_bool     OK;
 443     pme_setup_t *set;
 444     double       cycles_fast;
 445     char         buf[STRLEN], sbuf[22];
 446     real         rtab;
 447     gmx_bool     bUsesSimpleTables = TRUE;
 448
 449     if (pme_lb->stage == pme_lb->nstage)
 450     {
 451         return FALSE;
 452     }
 453
 454     if (PAR(cr))
 455     {
 456         gmx_sumd(1, &cycles, cr);
 457         cycles /= cr->nnodes;
 458     }
 459
 460     set = &pme_lb->setup[pme_lb->cur];
 461     set->count++;
 462
 463     rtab = ir->rlistlong + ir->tabext;
 464
 465     if (set->count % 2 == 1)
 466     {
 467         /* Skip the first cycle, because the first step after a switch
 468          * is much slower due to allocation and/or caching effects.
 469          */
 470         return TRUE;
 471     }
 472
 473     sprintf(buf, "step %4s: ", gmx_step_str(step, sbuf));
 474     print_grid(fp_err, fp_log, buf, "timed with", set, cycles);
 475
 476     if (set->count <= 2)
 477     {
 478         set->cycles = cycles;
 479     }
 480     else
 481     {
 482         if (cycles*PME_LB_ACCEL_TOL < set->cycles &&
 483             pme_lb->stage == pme_lb->nstage - 1)
 484         {
 485             /* The performance went up a lot (due to e.g. DD load balancing).
 486              * Add a stage, keep the minima, but rescan all setups.
 487              */
 488             pme_lb->nstage++;
 489
 490             if (debug)
 491             {
 492                 fprintf(debug, "The performance for grid %d %d %d went from %.3f to %.1f M-cycles, this is more than %f\n"
 493                         "Increased the number stages to %d"
 494                         " and ignoring the previous performance\n",
 495                         set->grid[XX], set->grid[YY], set->grid[ZZ],
 496                         cycles*1e-6, set->cycles*1e-6, PME_LB_ACCEL_TOL,
 497                         pme_lb->nstage);
 498             }
 499         }
 500         set->cycles = min(set->cycles, cycles);
 501     }
 502
 503     if (set->cycles < pme_lb->setup[pme_lb->fastest].cycles)
 504     {
 505         pme_lb->fastest = pme_lb->cur;
 506
 507         if (DOMAINDECOMP(cr))
 508         {
 509             /* We found a new fastest setting, ensure that with subsequent
 510              * shorter cut-off's the dynamic load balancing does not make
 511              * the use of the current cut-off impossible. This solution is
 512              * a trade-off, as the PME load balancing and DD domain size
 513              * load balancing can interact in complex ways.
 514              * With the Verlet kernels, DD load imbalance will usually be
 515              * mainly due to bonded interaction imbalance, which will often
 516              * quickly push the domain boundaries beyond the limit for the
 517              * optimal, PME load balanced, cut-off. But it could be that
 518              * better overal performance can be obtained with a slightly
 519              * shorter cut-off and better DD load balancing.
 520              */
 521             change_dd_dlb_cutoff_limit(cr);
 522         }
 523     }
 524     cycles_fast = pme_lb->setup[pme_lb->fastest].cycles;
 525
 526     /* Check in stage 0 if we should stop scanning grids.
 527      * Stop when the time is more than SLOW_FAC longer than the fastest.
 528      */
 529     if (pme_lb->stage == 0 && pme_lb->cur > 0 &&
 530         cycles > pme_lb->setup[pme_lb->fastest].cycles*PME_LB_SLOW_FAC)
 531     {
 532         pme_lb->n = pme_lb->cur + 1;
 533         /* Done with scanning, go to stage 1 */
 534         switch_to_stage1(pme_lb);
 535     }
 536
 537     if (pme_lb->stage == 0)
 538     {
 539         int gridsize_start;
 540
 541         gridsize_start = set->grid[XX]*set->grid[YY]*set->grid[ZZ];
 542
 543         do
 544         {
 545             if (pme_lb->cur+1 < pme_lb->n)
 546             {
 547                 /* We had already generated the next setup */
 548                 OK = TRUE;
 549             }
 550             else
 551             {
 552                 /* Find the next setup */
 553                 OK = pme_loadbal_increase_cutoff(pme_lb, ir->pme_order, cr->dd);
 554
 555                 if (!OK)
 556                 {
 557                     pme_lb->elimited = epmelblimPMEGRID;
 558                 }
 559             }
 560
 561             if (OK && ir->ePBC != epbcNONE)
 562             {
 563                 OK = (sqr(pme_lb->setup[pme_lb->cur+1].rlistlong)
 564                       <= max_cutoff2(ir->ePBC, state->box));
 565                 if (!OK)
 566                 {
 567                     pme_lb->elimited = epmelblimBOX;
 568                 }
 569             }
 570
 571             if (OK)
 572             {
 573                 pme_lb->cur++;
 574
 575                 if (DOMAINDECOMP(cr))
 576                 {
 577                     OK = change_dd_cutoff(cr, state, ir,
 578                                           pme_lb->setup[pme_lb->cur].rlistlong);
 579                     if (!OK)
 580                     {
 581                         /* Failed: do not use this setup */
 582                         pme_lb->cur--;
 583                         pme_lb->elimited = epmelblimDD;
 584                     }
 585                 }
 586             }
 587             if (!OK)
 588             {
 589                 /* We hit the upper limit for the cut-off,
 590                  * the setup should not go further than cur.
 591                  */
 592                 pme_lb->n = pme_lb->cur + 1;
 593                 print_loadbal_limited(fp_err, fp_log, step, pme_lb);
 594                 /* Switch to the next stage */
 595                 switch_to_stage1(pme_lb);
 596             }
 597         }
 598         while (OK &&
 599                !(pme_lb->setup[pme_lb->cur].grid[XX]*
 600                  pme_lb->setup[pme_lb->cur].grid[YY]*
 601                  pme_lb->setup[pme_lb->cur].grid[ZZ] <
 602                  gridsize_start*PME_LB_GRID_SCALE_FAC
 603                  &&
 604                  pme_lb->setup[pme_lb->cur].grid_efficiency <
 605                  pme_lb->setup[pme_lb->cur-1].grid_efficiency*PME_LB_GRID_EFFICIENCY_REL_FAC));
 606     }
 607
 608     if (pme_lb->stage > 0 && pme_lb->end == 1)
 609     {
 610         pme_lb->cur   = 0;
 611         pme_lb->stage = pme_lb->nstage;
 612     }
 613     else if (pme_lb->stage > 0 && pme_lb->end > 1)
 614     {
 615         /* If stage = nstage-1:
 616          *   scan over all setups, rerunning only those setups
 617          *   which are not much slower than the fastest
 618          * else:
 619          *   use the next setup
 620          */
 621         do
 622         {
 623             pme_lb->cur++;
 624             if (pme_lb->cur == pme_lb->end)
 625             {
 626                 pme_lb->stage++;
 627                 pme_lb->cur = pme_lb->start;
 628             }
 629         }
 630         while (pme_lb->stage == pme_lb->nstage - 1 &&
 631                pme_lb->setup[pme_lb->cur].count > 0 &&
 632                pme_lb->setup[pme_lb->cur].cycles > cycles_fast*PME_LB_SLOW_FAC);
 633
 634         if (pme_lb->stage == pme_lb->nstage)
 635         {
 636             /* We are done optimizing, use the fastest setup we found */
 637             pme_lb->cur = pme_lb->fastest;
 638         }
 639     }
 640
 641     if (DOMAINDECOMP(cr) && pme_lb->stage > 0)
 642     {
 643         OK = change_dd_cutoff(cr, state, ir, pme_lb->setup[pme_lb->cur].rlistlong);
 644         if (!OK)
 645         {
 646             /* Failsafe solution */
 647             if (pme_lb->cur > 1 && pme_lb->stage == pme_lb->nstage)
 648             {
 649                 pme_lb->stage--;
 650             }
 651             pme_lb->fastest  = 0;
 652             pme_lb->start    = 0;
 653             pme_lb->end      = pme_lb->cur;
 654             pme_lb->cur      = pme_lb->start;
 655             pme_lb->elimited = epmelblimDD;
 656             print_loadbal_limited(fp_err, fp_log, step, pme_lb);
 657         }
 658     }
 659
 660     /* Change the Coulomb cut-off and the PME grid */
 661
 662     set = &pme_lb->setup[pme_lb->cur];
 663
 664     ic->rcoulomb     = set->rcut_coulomb;
 665     ic->rlist        = set->rlist;
 666     ic->rlistlong    = set->rlistlong;
 667     ir->nstcalclr    = set->nstcalclr;
 668     ic->ewaldcoeff_q = set->ewaldcoeff_q;
 669     /* TODO: centralize the code that sets the potentials shifts */
 670     if (ic->coulomb_modifier == eintmodPOTSHIFT)
 671     {
 672         ic->sh_ewald = gmx_erfc(ic->ewaldcoeff_q*ic->rcoulomb);
 673     }
 674     if (EVDW_PME(ic->vdwtype))
 675     {
 676         /* We have PME for both Coulomb and VdW, set rvdw equal to rcoulomb */
 677         ic->rvdw            = set->rcut_coulomb;
 678         ic->ewaldcoeff_lj   = set->ewaldcoeff_lj;
 679         if (ic->vdw_modifier == eintmodPOTSHIFT)
 680         {
 681             real crc2;
 682
 683             ic->dispersion_shift.cpot = -pow(ic->rvdw, -6.0);
 684             ic->repulsion_shift.cpot  = -pow(ic->rvdw, -12.0);
 685             ic->sh_invrc6             = -ic->dispersion_shift.cpot;
 686             crc2                      = sqr(ic->ewaldcoeff_lj*ic->rvdw);
 687             ic->sh_lj_ewald           = (exp(-crc2)*(1 + crc2 + 0.5*crc2*crc2) - 1)*pow(ic->rvdw, -6.0);
 688         }
 689     }
 690
 691     bUsesSimpleTables = uses_simple_tables(ir->cutoff_scheme, nbv, 0);
 692     if (pme_lb->cutoff_scheme == ecutsVERLET &&
 693         nbv->grp[0].kernel_type == nbnxnk8x8x8_CUDA)
 694     {
 695         nbnxn_cuda_pme_loadbal_update_param(nbv->cu_nbv, ic);
 696
 697         /* With tMPI + GPUs some ranks may be sharing GPU(s) and therefore
 698          * also sharing texture references. To keep the code simple, we don't
 699          * treat texture references as shared resources, but this means that
 700          * the coulomb_tab texture ref will get updated by multiple threads.
 701          * Hence, to ensure that the non-bonded kernels don't start before all
 702          * texture binding operations are finished, we need to wait for all ranks
 703          * to arrive here before continuing.
 704          *
 705          * Note that we could omit this barrier if GPUs are not shared (or
 706          * texture objects are used), but as this is initialization code, there
 707          * is not point in complicating things.
 708          */
 709 #ifdef GMX_THREAD_MPI
 710         if (PAR(cr))
 711         {
 712             gmx_barrier(cr);
 713         }
 714 #endif  /* GMX_THREAD_MPI */
 715     }
 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 }