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 "config.h"
  36
  37 #include "types/commrec.h"
  38 #include "network.h"
  39 #include "calcgrid.h"
  40 #include "pme.h"
  41 #include "domdec.h"
  42 #include "nbnxn_cuda_data_mgmt.h"
  43 #include "force.h"
  44 #include "macros.h"
  45 #include "md_logging.h"
  46 #include "pme_loadbal.h"
  47
  48 #include "gromacs/math/vec.h"
  49 #include "gromacs/pbcutil/pbc.h"
  50 #include "gromacs/utility/cstringutil.h"
  51 #include "gromacs/utility/smalloc.h"
  52
  53 /* Parameters and setting for one PP-PME setup */
  54 typedef struct {
  55     real      rcut_coulomb;    /* Coulomb cut-off                              */
  56     real      rlist;           /* pair-list cut-off                            */
  57     real      rlistlong;       /* LR pair-list cut-off                         */
  58     int       nstcalclr;       /* frequency of evaluating long-range forces for group scheme */
  59     real      spacing;         /* (largest) PME grid spacing                   */
  60     ivec      grid;            /* the PME grid dimensions                      */
  61     real      grid_efficiency; /* ineffiency factor for non-uniform grids <= 1 */
  62     real      ewaldcoeff_q;    /* Electrostatic Ewald coefficient            */
  63     real      ewaldcoeff_lj;   /* LJ Ewald coefficient, only for the call to send_switchgrid */
  64     gmx_pme_t pmedata;         /* the data structure used in the PME code      */
  65     int       count;           /* number of times this setup has been timed    */
  66     double    cycles;          /* the fastest time for this setup in cycles    */
  67 } pme_setup_t;
  68
  69 /* In the initial scan, step by grids that are at least a factor 0.8 coarser */
  70 #define PME_LB_GRID_SCALE_FAC  0.8
  71 /* In the initial scan, try to skip grids with uneven x/y/z spacing,
  72  * checking if the "efficiency" is more than 5% worse than the previous grid.
  73  */
  74 #define PME_LB_GRID_EFFICIENCY_REL_FAC  1.05
  75 /* Rerun up till 12% slower setups than the fastest up till now */
  76 #define PME_LB_SLOW_FAC  1.12
  77 /* If setups get more than 2% faster, do another round to avoid
  78  * choosing a slower setup due to acceleration or fluctuations.
  79  */
  80 #define PME_LB_ACCEL_TOL 1.02
  81
  82 enum {
  83     epmelblimNO, epmelblimBOX, epmelblimDD, epmelblimPMEGRID, epmelblimNR
  84 };
  85
  86 const char *pmelblim_str[epmelblimNR] =
  87 { "no", "box size", "domain decompostion", "PME grid restriction" };
  88
  89 struct pme_load_balancing {
  90     int          nstage;             /* the current maximum number of stages */
  91
  92     real         cut_spacing;        /* the minimum cutoff / PME grid spacing ratio */
  93     real         rcut_vdw;           /* Vdw cutoff (does not change) */
  94     real         rcut_coulomb_start; /* Initial electrostatics cutoff */
  95     int          nstcalclr_start;    /* Initial electrostatics cutoff */
  96     real         rbuf_coulomb;       /* the pairlist buffer size */
  97     real         rbuf_vdw;           /* the pairlist buffer size */
  98     matrix       box_start;          /* the initial simulation box */
  99     int          n;                  /* the count of setup as well as the allocation size */
 100     pme_setup_t *setup;              /* the PME+cutoff setups */
 101     int          cur;                /* the current setup */
 102     int          fastest;            /* fastest setup up till now */
 103     int          start;              /* start of setup range to consider in stage>0 */
 104     int          end;                /* end   of setup range to consider in stage>0 */
 105     int          elimited;           /* was the balancing limited, uses enum above */
 106     int          cutoff_scheme;      /* Verlet or group cut-offs */
 107
 108     int          stage;              /* the current stage */
 109 };
 110
 111 void pme_loadbal_init(pme_load_balancing_t *pme_lb_p,
 112                       const t_inputrec *ir, matrix box,
 113                       const interaction_const_t *ic,
 114                       gmx_pme_t pmedata)
 115 {
 116     pme_load_balancing_t pme_lb;
 117     real                 spm, sp;
 118     int                  d;
 119
 120     snew(pme_lb, 1);
 121
 122     /* Any number of stages >= 2 is supported */
 123     pme_lb->nstage   = 2;
 124
 125     pme_lb->cutoff_scheme = ir->cutoff_scheme;
 126
 127     if (pme_lb->cutoff_scheme == ecutsVERLET)
 128     {
 129         pme_lb->rbuf_coulomb = ic->rlist - ic->rcoulomb;
 130         pme_lb->rbuf_vdw     = pme_lb->rbuf_coulomb;
 131     }
 132     else
 133     {
 134         if (ic->rcoulomb > ic->rlist)
 135         {
 136             pme_lb->rbuf_coulomb = ic->rlistlong - ic->rcoulomb;
 137         }
 138         else
 139         {
 140             pme_lb->rbuf_coulomb = ic->rlist - ic->rcoulomb;
 141         }
 142         if (ic->rvdw > ic->rlist)
 143         {
 144             pme_lb->rbuf_vdw = ic->rlistlong - ic->rvdw;
 145         }
 146         else
 147         {
 148             pme_lb->rbuf_vdw = ic->rlist - ic->rvdw;
 149         }
 150     }
 151
 152     copy_mat(box, pme_lb->box_start);
 153     if (ir->ePBC == epbcXY && ir->nwall == 2)
 154     {
 155         svmul(ir->wall_ewald_zfac, pme_lb->box_start[ZZ], pme_lb->box_start[ZZ]);
 156     }
 157
 158     pme_lb->n = 1;
 159     snew(pme_lb->setup, pme_lb->n);
 160
 161     pme_lb->rcut_vdw                 = ic->rvdw;
 162     pme_lb->rcut_coulomb_start       = ir->rcoulomb;
 163     pme_lb->nstcalclr_start          = ir->nstcalclr;
 164
 165     pme_lb->cur                      = 0;
 166     pme_lb->setup[0].rcut_coulomb    = ic->rcoulomb;
 167     pme_lb->setup[0].rlist           = ic->rlist;
 168     pme_lb->setup[0].rlistlong       = ic->rlistlong;
 169     pme_lb->setup[0].nstcalclr       = ir->nstcalclr;
 170     pme_lb->setup[0].grid[XX]        = ir->nkx;
 171     pme_lb->setup[0].grid[YY]        = ir->nky;
 172     pme_lb->setup[0].grid[ZZ]        = ir->nkz;
 173     pme_lb->setup[0].ewaldcoeff_q    = ic->ewaldcoeff_q;
 174     pme_lb->setup[0].ewaldcoeff_lj   = ic->ewaldcoeff_lj;
 175
 176     pme_lb->setup[0].pmedata  = pmedata;
 177
 178     spm = 0;
 179     for (d = 0; d < DIM; d++)
 180     {
 181         sp = norm(pme_lb->box_start[d])/pme_lb->setup[0].grid[d];
 182         if (sp > spm)
 183         {
 184             spm = sp;
 185         }
 186     }
 187     pme_lb->setup[0].spacing = spm;
 188
 189     if (ir->fourier_spacing > 0)
 190     {
 191         pme_lb->cut_spacing = ir->rcoulomb/ir->fourier_spacing;
 192     }
 193     else
 194     {
 195         pme_lb->cut_spacing = ir->rcoulomb/pme_lb->setup[0].spacing;
 196     }
 197
 198     pme_lb->stage = 0;
 199
 200     pme_lb->fastest  = 0;
 201     pme_lb->start    = 0;
 202     pme_lb->end      = 0;
 203     pme_lb->elimited = epmelblimNO;
 204
 205     *pme_lb_p = pme_lb;
 206 }
 207
 208 static gmx_bool pme_loadbal_increase_cutoff(pme_load_balancing_t  pme_lb,
 209                                             int                   pme_order,
 210                                             const gmx_domdec_t   *dd)
 211 {
 212     pme_setup_t *set;
 213     int          npmenodes_x, npmenodes_y;
 214     real         fac, sp;
 215     real         tmpr_coulomb, tmpr_vdw;
 216     int          d;
 217     gmx_bool     grid_ok;
 218
 219     /* Try to add a new setup with next larger cut-off to the list */
 220     pme_lb->n++;
 221     srenew(pme_lb->setup, pme_lb->n);
 222     set          = &pme_lb->setup[pme_lb->n-1];
 223     set->pmedata = NULL;
 224
 225     get_pme_nnodes(dd, &npmenodes_x, &npmenodes_y);
 226
 227     fac = 1;
 228     do
 229     {
 230         /* Avoid infinite while loop, which can occur at the minimum grid size.
 231          * Note that in practice load balancing will stop before this point.
 232          * The factor 2.1 allows for the extreme case in which only grids
 233          * of powers of 2 are allowed (the current code supports more grids).
 234          */
 235         if (fac > 2.1)
 236         {
 237             pme_lb->n--;
 238
 239             return FALSE;
 240         }
 241
 242         fac *= 1.01;
 243         clear_ivec(set->grid);
 244         sp = calc_grid(NULL, pme_lb->box_start,
 245                        fac*pme_lb->setup[pme_lb->cur].spacing,
 246                        &set->grid[XX],
 247                        &set->grid[YY],
 248                        &set->grid[ZZ]);
 249
 250         /* As here we can't easily check if one of the PME nodes
 251          * uses threading, we do a conservative grid check.
 252          * This means we can't use pme_order or less grid lines
 253          * per PME node along x, which is not a strong restriction.
 254          */
 255         gmx_pme_check_restrictions(pme_order,
 256                                    set->grid[XX], set->grid[YY], set->grid[ZZ],
 257                                    npmenodes_x, npmenodes_y,
 258                                    TRUE,
 259                                    FALSE,
 260                                    &grid_ok);
 261     }
 262     while (sp <= 1.001*pme_lb->setup[pme_lb->cur].spacing || !grid_ok);
 263
 264     set->rcut_coulomb = pme_lb->cut_spacing*sp;
 265
 266     if (pme_lb->cutoff_scheme == ecutsVERLET)
 267     {
 268         set->rlist        = set->rcut_coulomb + pme_lb->rbuf_coulomb;
 269         /* We dont use LR lists with Verlet, but this avoids if-statements in further checks */
 270         set->rlistlong    = set->rlist;
 271     }
 272     else
 273     {
 274         tmpr_coulomb          = set->rcut_coulomb + pme_lb->rbuf_coulomb;
 275         tmpr_vdw              = pme_lb->rcut_vdw + pme_lb->rbuf_vdw;
 276         set->rlist            = min(tmpr_coulomb, tmpr_vdw);
 277         set->rlistlong        = max(tmpr_coulomb, tmpr_vdw);
 278
 279         /* Set the long-range update frequency */
 280         if (set->rlist == set->rlistlong)
 281         {
 282             /* No long-range interactions if the short-/long-range cutoffs are identical */
 283             set->nstcalclr = 0;
 284         }
 285         else if (pme_lb->nstcalclr_start == 0 || pme_lb->nstcalclr_start == 1)
 286         {
 287             /* We were not doing long-range before, but now we are since rlist!=rlistlong */
 288             set->nstcalclr = 1;
 289         }
 290         else
 291         {
 292             /* We were already doing long-range interactions from the start */
 293             if (pme_lb->rcut_vdw > pme_lb->rcut_coulomb_start)
 294             {
 295                 /* We were originally doing long-range VdW-only interactions.
 296                  * If rvdw is still longer than rcoulomb we keep the original nstcalclr,
 297                  * but if the coulomb cutoff has become longer we should update the long-range
 298                  * part every step.
 299                  */
 300                 set->nstcalclr = (tmpr_vdw > tmpr_coulomb) ? pme_lb->nstcalclr_start : 1;
 301             }
 302             else
 303             {
 304                 /* We were not doing any long-range interaction from the start,
 305                  * since it is not possible to do twin-range coulomb for the PME interaction.
 306                  */
 307                 set->nstcalclr = 1;
 308             }
 309         }
 310     }
 311
 312     set->spacing      = sp;
 313     /* The grid efficiency is the size wrt a grid with uniform x/y/z spacing */
 314     set->grid_efficiency = 1;
 315     for (d = 0; d < DIM; d++)
 316     {
 317         set->grid_efficiency *= (set->grid[d]*sp)/norm(pme_lb->box_start[d]);
 318     }
 319     /* The Ewald coefficient is inversly proportional to the cut-off */
 320     set->ewaldcoeff_q =
 321         pme_lb->setup[0].ewaldcoeff_q*pme_lb->setup[0].rcut_coulomb/set->rcut_coulomb;
 322     /* We set ewaldcoeff_lj in set, even when LJ-PME is not used */
 323     set->ewaldcoeff_lj =
 324         pme_lb->setup[0].ewaldcoeff_lj*pme_lb->setup[0].rcut_coulomb/set->rcut_coulomb;
 325
 326     set->count   = 0;
 327     set->cycles  = 0;
 328
 329     if (debug)
 330     {
 331         fprintf(debug, "PME loadbal: grid %d %d %d, coulomb cutoff %f\n",
 332                 set->grid[XX], set->grid[YY], set->grid[ZZ], set->rcut_coulomb);
 333     }
 334     return TRUE;
 335 }
 336
 337 static void print_grid(FILE *fp_err, FILE *fp_log,
 338                        const char *pre,
 339                        const char *desc,
 340                        const pme_setup_t *set,
 341                        double cycles)
 342 {
 343     char buf[STRLEN], buft[STRLEN];
 344
 345     if (cycles >= 0)
 346     {
 347         sprintf(buft, ": %.1f M-cycles", cycles*1e-6);
 348     }
 349     else
 350     {
 351         buft[0] = '\0';
 352     }
 353     sprintf(buf, "%-11s%10s pme grid %d %d %d, coulomb cutoff %.3f%s",
 354             pre,
 355             desc, set->grid[XX], set->grid[YY], set->grid[ZZ], set->rcut_coulomb,
 356             buft);
 357     if (fp_err != NULL)
 358     {
 359         fprintf(fp_err, "\r%s\n", buf);
 360     }
 361     if (fp_log != NULL)
 362     {
 363         fprintf(fp_log, "%s\n", buf);
 364     }
 365 }
 366
 367 static int pme_loadbal_end(pme_load_balancing_t pme_lb)
 368 {
 369     /* In the initial stage only n is set; end is not set yet */
 370     if (pme_lb->end > 0)
 371     {
 372         return pme_lb->end;
 373     }
 374     else
 375     {
 376         return pme_lb->n;
 377     }
 378 }
 379
 380 static void print_loadbal_limited(FILE *fp_err, FILE *fp_log,
 381                                   gmx_int64_t step,
 382                                   pme_load_balancing_t pme_lb)
 383 {
 384     char buf[STRLEN], sbuf[22];
 385
 386     sprintf(buf, "step %4s: the %s limits the PME load balancing to a coulomb cut-off of %.3f",
 387             gmx_step_str(step, sbuf),
 388             pmelblim_str[pme_lb->elimited],
 389             pme_lb->setup[pme_loadbal_end(pme_lb)-1].rcut_coulomb);
 390     if (fp_err != NULL)
 391     {
 392         fprintf(fp_err, "\r%s\n", buf);
 393     }
 394     if (fp_log != NULL)
 395     {
 396         fprintf(fp_log, "%s\n", buf);
 397     }
 398 }
 399
 400 static void switch_to_stage1(pme_load_balancing_t pme_lb)
 401 {
 402     pme_lb->start = 0;
 403     while (pme_lb->start+1 < pme_lb->n &&
 404            (pme_lb->setup[pme_lb->start].count == 0 ||
 405             pme_lb->setup[pme_lb->start].cycles >
 406             pme_lb->setup[pme_lb->fastest].cycles*PME_LB_SLOW_FAC))
 407     {
 408         pme_lb->start++;
 409     }
 410     while (pme_lb->start > 0 && pme_lb->setup[pme_lb->start-1].cycles == 0)
 411     {
 412         pme_lb->start--;
 413     }
 414
 415     pme_lb->end = pme_lb->n;
 416     if (pme_lb->setup[pme_lb->end-1].count > 0 &&
 417         pme_lb->setup[pme_lb->end-1].cycles >
 418         pme_lb->setup[pme_lb->fastest].cycles*PME_LB_SLOW_FAC)
 419     {
 420         pme_lb->end--;
 421     }
 422
 423     pme_lb->stage = 1;
 424
 425     /* Next we want to choose setup pme_lb->start, but as we will increase
 426      * pme_ln->cur by one right after returning, we subtract 1 here.
 427      */
 428     pme_lb->cur = pme_lb->start - 1;
 429 }
 430
 431 gmx_bool pme_load_balance(pme_load_balancing_t pme_lb,
 432                           t_commrec           *cr,
 433                           FILE                *fp_err,
 434                           FILE                *fp_log,
 435                           t_inputrec          *ir,
 436                           t_state             *state,
 437                           double               cycles,
 438                           interaction_const_t *ic,
 439                           nonbonded_verlet_t  *nbv,
 440                           gmx_pme_t           *pmedata,
 441                           gmx_int64_t          step)
 442 {
 443     gmx_bool     OK;
 444     pme_setup_t *set;
 445     double       cycles_fast;
 446     char         buf[STRLEN], sbuf[22];
 447     real         rtab;
 448     gmx_bool     bUsesSimpleTables = TRUE;
 449
 450     if (pme_lb->stage == pme_lb->nstage)
 451     {
 452         return FALSE;
 453     }
 454
 455     if (PAR(cr))
 456     {
 457         gmx_sumd(1, &cycles, cr);
 458         cycles /= cr->nnodes;
 459     }
 460
 461     set = &pme_lb->setup[pme_lb->cur];
 462     set->count++;
 463
 464     rtab = ir->rlistlong + ir->tabext;
 465
 466     if (set->count % 2 == 1)
 467     {
 468         /* Skip the first cycle, because the first step after a switch
 469          * is much slower due to allocation and/or caching effects.
 470          */
 471         return TRUE;
 472     }
 473
 474     sprintf(buf, "step %4s: ", gmx_step_str(step, sbuf));
 475     print_grid(fp_err, fp_log, buf, "timed with", set, cycles);
 476
 477     if (set->count <= 2)
 478     {
 479         set->cycles = cycles;
 480     }
 481     else
 482     {
 483         if (cycles*PME_LB_ACCEL_TOL < set->cycles &&
 484             pme_lb->stage == pme_lb->nstage - 1)
 485         {
 486             /* The performance went up a lot (due to e.g. DD load balancing).
 487              * Add a stage, keep the minima, but rescan all setups.
 488              */
 489             pme_lb->nstage++;
 490
 491             if (debug)
 492             {
 493                 fprintf(debug, "The performance for grid %d %d %d went from %.3f to %.1f M-cycles, this is more than %f\n"
 494                         "Increased the number stages to %d"
 495                         " and ignoring the previous performance\n",
 496                         set->grid[XX], set->grid[YY], set->grid[ZZ],
 497                         cycles*1e-6, set->cycles*1e-6, PME_LB_ACCEL_TOL,
 498                         pme_lb->nstage);
 499             }
 500         }
 501         set->cycles = min(set->cycles, cycles);
 502     }
 503
 504     if (set->cycles < pme_lb->setup[pme_lb->fastest].cycles)
 505     {
 506         pme_lb->fastest = pme_lb->cur;
 507
 508         if (DOMAINDECOMP(cr))
 509         {
 510             /* We found a new fastest setting, ensure that with subsequent
 511              * shorter cut-off's the dynamic load balancing does not make
 512              * the use of the current cut-off impossible. This solution is
 513              * a trade-off, as the PME load balancing and DD domain size
 514              * load balancing can interact in complex ways.
 515              * With the Verlet kernels, DD load imbalance will usually be
 516              * mainly due to bonded interaction imbalance, which will often
 517              * quickly push the domain boundaries beyond the limit for the
 518              * optimal, PME load balanced, cut-off. But it could be that
 519              * better overal performance can be obtained with a slightly
 520              * shorter cut-off and better DD load balancing.
 521              */
 522             change_dd_dlb_cutoff_limit(cr);
 523         }
 524     }
 525     cycles_fast = pme_lb->setup[pme_lb->fastest].cycles;
 526
 527     /* Check in stage 0 if we should stop scanning grids.
 528      * Stop when the time is more than SLOW_FAC longer than the fastest.
 529      */
 530     if (pme_lb->stage == 0 && pme_lb->cur > 0 &&
 531         cycles > pme_lb->setup[pme_lb->fastest].cycles*PME_LB_SLOW_FAC)
 532     {
 533         pme_lb->n = pme_lb->cur + 1;
 534         /* Done with scanning, go to stage 1 */
 535         switch_to_stage1(pme_lb);
 536     }
 537
 538     if (pme_lb->stage == 0)
 539     {
 540         int gridsize_start;
 541
 542         gridsize_start = set->grid[XX]*set->grid[YY]*set->grid[ZZ];
 543
 544         do
 545         {
 546             if (pme_lb->cur+1 < pme_lb->n)
 547             {
 548                 /* We had already generated the next setup */
 549                 OK = TRUE;
 550             }
 551             else
 552             {
 553                 /* Find the next setup */
 554                 OK = pme_loadbal_increase_cutoff(pme_lb, ir->pme_order, cr->dd);
 555
 556                 if (!OK)
 557                 {
 558                     pme_lb->elimited = epmelblimPMEGRID;
 559                 }
 560             }
 561
 562             if (OK && ir->ePBC != epbcNONE)
 563             {
 564                 OK = (sqr(pme_lb->setup[pme_lb->cur+1].rlistlong)
 565                       <= max_cutoff2(ir->ePBC, state->box));
 566                 if (!OK)
 567                 {
 568                     pme_lb->elimited = epmelblimBOX;
 569                 }
 570             }
 571
 572             if (OK)
 573             {
 574                 pme_lb->cur++;
 575
 576                 if (DOMAINDECOMP(cr))
 577                 {
 578                     OK = change_dd_cutoff(cr, state, ir,
 579                                           pme_lb->setup[pme_lb->cur].rlistlong);
 580                     if (!OK)
 581                     {
 582                         /* Failed: do not use this setup */
 583                         pme_lb->cur--;
 584                         pme_lb->elimited = epmelblimDD;
 585                     }
 586                 }
 587             }
 588             if (!OK)
 589             {
 590                 /* We hit the upper limit for the cut-off,
 591                  * the setup should not go further than cur.
 592                  */
 593                 pme_lb->n = pme_lb->cur + 1;
 594                 print_loadbal_limited(fp_err, fp_log, step, pme_lb);
 595                 /* Switch to the next stage */
 596                 switch_to_stage1(pme_lb);
 597             }
 598         }
 599         while (OK &&
 600                !(pme_lb->setup[pme_lb->cur].grid[XX]*
 601                  pme_lb->setup[pme_lb->cur].grid[YY]*
 602                  pme_lb->setup[pme_lb->cur].grid[ZZ] <
 603                  gridsize_start*PME_LB_GRID_SCALE_FAC
 604                  &&
 605                  pme_lb->setup[pme_lb->cur].grid_efficiency <
 606                  pme_lb->setup[pme_lb->cur-1].grid_efficiency*PME_LB_GRID_EFFICIENCY_REL_FAC));
 607     }
 608
 609     if (pme_lb->stage > 0 && pme_lb->end == 1)
 610     {
 611         pme_lb->cur   = 0;
 612         pme_lb->stage = pme_lb->nstage;
 613     }
 614     else if (pme_lb->stage > 0 && pme_lb->end > 1)
 615     {
 616         /* If stage = nstage-1:
 617          *   scan over all setups, rerunning only those setups
 618          *   which are not much slower than the fastest
 619          * else:
 620          *   use the next setup
 621          */
 622         do
 623         {
 624             pme_lb->cur++;
 625             if (pme_lb->cur == pme_lb->end)
 626             {
 627                 pme_lb->stage++;
 628                 pme_lb->cur = pme_lb->start;
 629             }
 630         }
 631         while (pme_lb->stage == pme_lb->nstage - 1 &&
 632                pme_lb->setup[pme_lb->cur].count > 0 &&
 633                pme_lb->setup[pme_lb->cur].cycles > cycles_fast*PME_LB_SLOW_FAC);
 634
 635         if (pme_lb->stage == pme_lb->nstage)
 636         {
 637             /* We are done optimizing, use the fastest setup we found */
 638             pme_lb->cur = pme_lb->fastest;
 639         }
 640     }
 641
 642     if (DOMAINDECOMP(cr) && pme_lb->stage > 0)
 643     {
 644         OK = change_dd_cutoff(cr, state, ir, pme_lb->setup[pme_lb->cur].rlistlong);
 645         if (!OK)
 646         {
 647             /* Failsafe solution */
 648             if (pme_lb->cur > 1 && pme_lb->stage == pme_lb->nstage)
 649             {
 650                 pme_lb->stage--;
 651             }
 652             pme_lb->fastest  = 0;
 653             pme_lb->start    = 0;
 654             pme_lb->end      = pme_lb->cur;
 655             pme_lb->cur      = pme_lb->start;
 656             pme_lb->elimited = epmelblimDD;
 657             print_loadbal_limited(fp_err, fp_log, step, pme_lb);
 658         }
 659     }
 660
 661     /* Change the Coulomb cut-off and the PME grid */
 662
 663     set = &pme_lb->setup[pme_lb->cur];
 664
 665     ic->rcoulomb     = set->rcut_coulomb;
 666     ic->rlist        = set->rlist;
 667     ic->rlistlong    = set->rlistlong;
 668     ir->nstcalclr    = set->nstcalclr;
 669     ic->ewaldcoeff_q = set->ewaldcoeff_q;
 670     /* TODO: centralize the code that sets the potentials shifts */
 671     if (ic->coulomb_modifier == eintmodPOTSHIFT)
 672     {
 673         ic->sh_ewald = gmx_erfc(ic->ewaldcoeff_q*ic->rcoulomb);
 674     }
 675     if (EVDW_PME(ic->vdwtype))
 676     {
 677         /* We have PME for both Coulomb and VdW, set rvdw equal to rcoulomb */
 678         ic->rvdw            = set->rcut_coulomb;
 679         ic->ewaldcoeff_lj   = set->ewaldcoeff_lj;
 680         if (ic->vdw_modifier == eintmodPOTSHIFT)
 681         {
 682             real crc2;
 683
 684             ic->dispersion_shift.cpot = -pow(ic->rvdw, -6.0);
 685             ic->repulsion_shift.cpot  = -pow(ic->rvdw, -12.0);
 686             ic->sh_invrc6             = -ic->dispersion_shift.cpot;
 687             crc2                      = sqr(ic->ewaldcoeff_lj*ic->rvdw);
 688             ic->sh_lj_ewald           = (exp(-crc2)*(1 + crc2 + 0.5*crc2*crc2) - 1)*pow(ic->rvdw, -6.0);
 689         }
 690     }
 691
 692     bUsesSimpleTables = uses_simple_tables(ir->cutoff_scheme, nbv, 0);
 693     if (pme_lb->cutoff_scheme == ecutsVERLET &&
 694         nbv->grp[0].kernel_type == nbnxnk8x8x8_CUDA)
 695     {
 696         nbnxn_cuda_pme_loadbal_update_param(nbv->cu_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))
 712         {
 713             gmx_barrier(cr);
 714         }
 715 #endif  /* GMX_THREAD_MPI */
 716     }
 717
 718     /* Usually we won't need the simple tables with GPUs.
 719      * But we do with hybrid acceleration and with free energy.
 720      * To avoid bugs, we always re-initialize the simple tables here.
 721      */
 722     init_interaction_const_tables(NULL, ic, bUsesSimpleTables, rtab);
 723
 724     if (cr->duty & DUTY_PME)
 725     {
 726         if (pme_lb->setup[pme_lb->cur].pmedata == NULL)
 727         {
 728             /* Generate a new PME data structure,
 729              * copying part of the old pointers.
 730              */
 731             gmx_pme_reinit(&set->pmedata,
 732                            cr, pme_lb->setup[0].pmedata, ir,
 733                            set->grid);
 734         }
 735         *pmedata = set->pmedata;
 736     }
 737     else
 738     {
 739         /* Tell our PME-only node to switch grid */
 740         gmx_pme_send_switchgrid(cr, set->grid, set->ewaldcoeff_q, set->ewaldcoeff_lj);
 741     }
 742
 743     if (debug)
 744     {
 745         print_grid(NULL, debug, "", "switched to", set, -1);
 746     }
 747
 748     if (pme_lb->stage == pme_lb->nstage)
 749     {
 750         print_grid(fp_err, fp_log, "", "optimal", set, -1);
 751     }
 752
 753     return TRUE;
 754 }
 755
 756 void restart_pme_loadbal(pme_load_balancing_t pme_lb, int n)
 757 {
 758     pme_lb->nstage += n;
 759 }
 760
 761 static int pme_grid_points(const pme_setup_t *setup)
 762 {
 763     return setup->grid[XX]*setup->grid[YY]*setup->grid[ZZ];
 764 }
 765
 766 static real pme_loadbal_rlist(const pme_setup_t *setup)
 767 {
 768     /* With the group cut-off scheme we can have twin-range either
 769      * for Coulomb or for VdW, so we need a check here.
 770      * With the Verlet cut-off scheme rlist=rlistlong.
 771      */
 772     if (setup->rcut_coulomb > setup->rlist)
 773     {
 774         return setup->rlistlong;
 775     }
 776     else
 777     {
 778         return setup->rlist;
 779     }
 780 }
 781
 782 static void print_pme_loadbal_setting(FILE              *fplog,
 783                                       char              *name,
 784                                       const pme_setup_t *setup)
 785 {
 786     fprintf(fplog,
 787             "   %-7s %6.3f nm %6.3f nm     %3d %3d %3d   %5.3f nm  %5.3f nm\n",
 788             name,
 789             setup->rcut_coulomb, pme_loadbal_rlist(setup),
 790             setup->grid[XX], setup->grid[YY], setup->grid[ZZ],
 791             setup->spacing, 1/setup->ewaldcoeff_q);
 792 }
 793
 794 static void print_pme_loadbal_settings(pme_load_balancing_t pme_lb,
 795                                        t_commrec           *cr,
 796                                        FILE                *fplog,
 797                                        gmx_bool             bNonBondedOnGPU)
 798 {
 799     double pp_ratio, grid_ratio;
 800
 801     pp_ratio   = pow(pme_loadbal_rlist(&pme_lb->setup[pme_lb->cur])/pme_loadbal_rlist(&pme_lb->setup[0]), 3.0);
 802     grid_ratio = pme_grid_points(&pme_lb->setup[pme_lb->cur])/
 803         (double)pme_grid_points(&pme_lb->setup[0]);
 804
 805     fprintf(fplog, "\n");
 806     fprintf(fplog, "       P P   -   P M E   L O A D   B A L A N C I N G\n");
 807     fprintf(fplog, "\n");
 808     /* Here we only warn when the optimal setting is the last one */
 809     if (pme_lb->elimited != epmelblimNO &&
 810         pme_lb->cur == pme_loadbal_end(pme_lb)-1)
 811     {
 812         fprintf(fplog, " NOTE: The PP/PME load balancing was limited by the %s,\n",
 813                 pmelblim_str[pme_lb->elimited]);
 814         fprintf(fplog, "       you might not have reached a good load balance.\n");
 815         if (pme_lb->elimited == epmelblimDD)
 816         {
 817             fprintf(fplog, "       Try different mdrun -dd settings or lower the -dds value.\n");
 818         }
 819         fprintf(fplog, "\n");
 820     }
 821     fprintf(fplog, " PP/PME load balancing changed the cut-off and PME settings:\n");
 822     fprintf(fplog, "           particle-particle                    PME\n");
 823     fprintf(fplog, "            rcoulomb  rlist            grid      spacing   1/beta\n");
 824     print_pme_loadbal_setting(fplog, "initial", &pme_lb->setup[0]);
 825     print_pme_loadbal_setting(fplog, "final", &pme_lb->setup[pme_lb->cur]);
 826     fprintf(fplog, " cost-ratio           %4.2f             %4.2f\n",
 827             pp_ratio, grid_ratio);
 828     fprintf(fplog, " (note that these numbers concern only part of the total PP and PME load)\n");
 829
 830     if (pp_ratio > 1.5 && !bNonBondedOnGPU)
 831     {
 832         md_print_warn(cr, fplog,
 833                       "NOTE: PME load balancing increased the non-bonded workload by more than 50%%.\n"
 834                       "      For better performance, use (more) PME ranks (mdrun -npme),\n"
 835                       "      or if you are beyond the scaling limit, use fewer total ranks (or nodes).\n");
 836     }
 837     else
 838     {
 839         fprintf(fplog, "\n");
 840     }
 841 }
 842
 843 void pme_loadbal_done(pme_load_balancing_t pme_lb,
 844                       t_commrec *cr, FILE *fplog,
 845                       gmx_bool bNonBondedOnGPU)
 846 {
 847     if (fplog != NULL && (pme_lb->cur > 0 || pme_lb->elimited != epmelblimNO))
 848     {
 849         print_pme_loadbal_settings(pme_lb, cr, fplog, bNonBondedOnGPU);
 850     }
 851
 852     /* TODO: Here we should free all pointers in pme_lb,
 853      * but as it contains pme data structures,
 854      * we need to first make pme.c free all data.
 855      */
 856 }