src/programs/mdrun/pme_loadbal.c

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 2012,2013, by the GROMACS development team, led by
   5  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   6  * and including many others, as listed in the AUTHORS file in the
   7  * top-level source directory and at http://www.gromacs.org.
   8  *
   9  * GROMACS is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public License
  11  * as published by the Free Software Foundation; either version 2.1
  12  * of the License, or (at your option) any later version.
  13  *
  14  * GROMACS is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with GROMACS; if not, see
  21  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  22  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  23  *
  24  * If you want to redistribute modifications to GROMACS, please
  25  * consider that scientific software is very special. Version
  26  * control is crucial - bugs must be traceable. We will be happy to
  27  * consider code for inclusion in the official distribution, but
  28  * derived work must not be called official GROMACS. Details are found
  29  * in the README & COPYING files - if they are missing, get the
  30  * official version at http://www.gromacs.org.
  31  *
  32  * To help us fund GROMACS development, we humbly ask that you cite
  33  * the research papers on the package. Check out http://www.gromacs.org.
  34  */
  35 #ifdef HAVE_CONFIG_H
  36 #include <config.h>
  37 #endif
  38
  39 #include "smalloc.h"
  40 #include "network.h"
  41 #include "calcgrid.h"
  42 #include "pme.h"
  43 #include "vec.h"
  44 #include "domdec.h"
  45 #include "nbnxn_cuda_data_mgmt.h"
  46 #include "force.h"
  47 #include "macros.h"
  48 #include "md_logging.h"
  49 #include "pme_loadbal.h"
  50
  51 /* Parameters and setting for one PP-PME setup */
  52 typedef struct {
  53     real      rcut_coulomb;    /* Coulomb cut-off                              */
  54     real      rlist;           /* pair-list cut-off                            */
  55     real      rlistlong;       /* LR pair-list cut-off                         */
  56     int       nstcalclr;       /* frequency of evaluating long-range forces for group scheme */
  57     real      spacing;         /* (largest) PME grid spacing                   */
  58     ivec      grid;            /* the PME grid dimensions                      */
  59     real      grid_efficiency; /* ineffiency factor for non-uniform grids <= 1 */
  60     real      ewaldcoeff;      /* the Ewald coefficient                        */
  61     gmx_pme_t pmedata;         /* the data structure used in the PME code      */
  62
  63     int       count;           /* number of times this setup has been timed    */
  64     double    cycles;          /* the fastest time for this setup in cycles    */
  65 } pme_setup_t;
  66
  67 /* In the initial scan, step by grids that are at least a factor 0.8 coarser */
  68 #define PME_LB_GRID_SCALE_FAC  0.8
  69 /* In the initial scan, try to skip grids with uneven x/y/z spacing,
  70  * checking if the "efficiency" is more than 5% worse than the previous grid.
  71  */
  72 #define PME_LB_GRID_EFFICIENCY_REL_FAC  1.05
  73 /* Rerun up till 12% slower setups than the fastest up till now */
  74 #define PME_LB_SLOW_FAC  1.12
  75 /* If setups get more than 2% faster, do another round to avoid
  76  * choosing a slower setup due to acceleration or fluctuations.
  77  */
  78 #define PME_LB_ACCEL_TOL 1.02
  79
  80 enum {
  81     epmelblimNO, epmelblimBOX, epmelblimDD, epmelblimPMEGRID, epmelblimNR
  82 };
  83
  84 const char *pmelblim_str[epmelblimNR] =
  85 { "no", "box size", "domain decompostion", "PME grid restriction" };
  86
  87 struct pme_load_balancing {
  88     int          nstage;             /* the current maximum number of stages */
  89
  90     real         cut_spacing;        /* the minimum cutoff / PME grid spacing ratio */
  91     real         rcut_vdw;           /* Vdw cutoff (does not change) */
  92     real         rcut_coulomb_start; /* Initial electrostatics cutoff */
  93     int          nstcalclr_start;    /* Initial electrostatics cutoff */
  94     real         rbuf_coulomb;       /* the pairlist buffer size */
  95     real         rbuf_vdw;           /* the pairlist buffer size */
  96     matrix       box_start;          /* the initial simulation box */
  97     int          n;                  /* the count of setup as well as the allocation size */
  98     pme_setup_t *setup;              /* the PME+cutoff setups */
  99     int          cur;                /* the current setup */
 100     int          fastest;            /* fastest setup up till now */
 101     int          start;              /* start of setup range to consider in stage>0 */
 102     int          end;                /* end   of setup range to consider in stage>0 */
 103     int          elimited;           /* was the balancing limited, uses enum above */
 104     int          cutoff_scheme;      /* Verlet or group cut-offs */
 105
 106     int          stage;              /* the current stage */
 107 };
 108
 109 void pme_loadbal_init(pme_load_balancing_t *pme_lb_p,
 110                       const t_inputrec *ir, matrix box,
 111                       const interaction_const_t *ic,
 112                       gmx_pme_t pmedata)
 113 {
 114     pme_load_balancing_t pme_lb;
 115     real                 spm, sp;
 116     int                  d;
 117
 118     snew(pme_lb, 1);
 119
 120     /* Any number of stages >= 2 is supported */
 121     pme_lb->nstage   = 2;
 122
 123     pme_lb->cutoff_scheme = ir->cutoff_scheme;
 124
 125     if (pme_lb->cutoff_scheme == ecutsVERLET)
 126     {
 127         pme_lb->rbuf_coulomb = ic->rlist - ic->rcoulomb;
 128         pme_lb->rbuf_vdw     = pme_lb->rbuf_coulomb;
 129     }
 130     else
 131     {
 132         if (ic->rcoulomb > ic->rlist)
 133         {
 134             pme_lb->rbuf_coulomb = ic->rlistlong - ic->rcoulomb;
 135         }
 136         else
 137         {
 138             pme_lb->rbuf_coulomb = ic->rlist - ic->rcoulomb;
 139         }
 140         if (ic->rvdw > ic->rlist)
 141         {
 142             pme_lb->rbuf_vdw = ic->rlistlong - ic->rvdw;
 143         }
 144         else
 145         {
 146             pme_lb->rbuf_vdw = ic->rlist - ic->rvdw;
 147         }
 148     }
 149
 150     copy_mat(box, pme_lb->box_start);
 151     if (ir->ePBC == epbcXY && ir->nwall == 2)
 152     {
 153         svmul(ir->wall_ewald_zfac, pme_lb->box_start[ZZ], pme_lb->box_start[ZZ]);
 154     }
 155
 156     pme_lb->n = 1;
 157     snew(pme_lb->setup, pme_lb->n);
 158
 159     pme_lb->rcut_vdw              = ic->rvdw;
 160     pme_lb->rcut_coulomb_start    = ir->rcoulomb;
 161     pme_lb->nstcalclr_start       = ir->nstcalclr;
 162
 163     pme_lb->cur                   = 0;
 164     pme_lb->setup[0].rcut_coulomb = ic->rcoulomb;
 165     pme_lb->setup[0].rlist        = ic->rlist;
 166     pme_lb->setup[0].rlistlong    = ic->rlistlong;
 167     pme_lb->setup[0].nstcalclr    = ir->nstcalclr;
 168     pme_lb->setup[0].grid[XX]     = ir->nkx;
 169     pme_lb->setup[0].grid[YY]     = ir->nky;
 170     pme_lb->setup[0].grid[ZZ]     = ir->nkz;
 171     pme_lb->setup[0].ewaldcoeff   = ic->ewaldcoeff;
 172
 173     pme_lb->setup[0].pmedata  = pmedata;
 174
 175     spm = 0;
 176     for (d = 0; d < DIM; d++)
 177     {
 178         sp = norm(pme_lb->box_start[d])/pme_lb->setup[0].grid[d];
 179         if (sp > spm)
 180         {
 181             spm = sp;
 182         }
 183     }
 184     pme_lb->setup[0].spacing = spm;
 185
 186     if (ir->fourier_spacing > 0)
 187     {
 188         pme_lb->cut_spacing = ir->rcoulomb/ir->fourier_spacing;
 189     }
 190     else
 191     {
 192         pme_lb->cut_spacing = ir->rcoulomb/pme_lb->setup[0].spacing;
 193     }
 194
 195     pme_lb->stage = 0;
 196
 197     pme_lb->fastest  = 0;
 198     pme_lb->start    = 0;
 199     pme_lb->end      = 0;
 200     pme_lb->elimited = epmelblimNO;
 201
 202     *pme_lb_p = pme_lb;
 203 }
 204
 205 static gmx_bool pme_loadbal_increase_cutoff(pme_load_balancing_t pme_lb,
 206                                             int                  pme_order,
 207                                             const gmx_domdec_t   *dd)
 208 {
 209     pme_setup_t *set;
 210     int          npmenodes_x, npmenodes_y;
 211     real         fac, sp;
 212     real         tmpr_coulomb, tmpr_vdw;
 213     int          d;
 214     gmx_bool     grid_ok;
 215
 216     /* Try to add a new setup with next larger cut-off to the list */
 217     pme_lb->n++;
 218     srenew(pme_lb->setup, pme_lb->n);
 219     set          = &pme_lb->setup[pme_lb->n-1];
 220     set->pmedata = NULL;
 221
 222     get_pme_nnodes(dd, &npmenodes_x, &npmenodes_y);
 223
 224     fac = 1;
 225     do
 226     {
 227         /* Avoid infinite while loop, which can occur at the minimum grid size.
 228          * Note that in practice load balancing will stop before this point.
 229          * The factor 2.1 allows for the extreme case in which only grids
 230          * of powers of 2 are allowed (the current code supports more grids).
 231          */
 232         if (fac > 2.1)
 233         {
 234             pme_lb->n--;
 235
 236             return FALSE;
 237         }
 238
 239         fac *= 1.01;
 240         clear_ivec(set->grid);
 241         sp = calc_grid(NULL, pme_lb->box_start,
 242                        fac*pme_lb->setup[pme_lb->cur].spacing,
 243                        &set->grid[XX],
 244                        &set->grid[YY],
 245                        &set->grid[ZZ]);
 246
 247         /* As here we can't easily check if one of the PME nodes
 248          * uses threading, we do a conservative grid check.
 249          * This means we can't use pme_order or less grid lines
 250          * per PME node along x, which is not a strong restriction.
 251          */
 252         gmx_pme_check_restrictions(pme_order,
 253                                    set->grid[XX], set->grid[YY], set->grid[ZZ],
 254                                    npmenodes_x, npmenodes_y,
 255                                    TRUE,
 256                                    FALSE,
 257                                    &grid_ok);
 258     }
 259     while (sp <= 1.001*pme_lb->setup[pme_lb->cur].spacing || !grid_ok);
 260
 261     set->rcut_coulomb = pme_lb->cut_spacing*sp;
 262
 263     if (pme_lb->cutoff_scheme == ecutsVERLET)
 264     {
 265         set->rlist        = set->rcut_coulomb + pme_lb->rbuf_coulomb;
 266         /* We dont use LR lists with Verlet, but this avoids if-statements in further checks */
 267         set->rlistlong    = set->rlist;
 268     }
 269     else
 270     {
 271         tmpr_coulomb          = set->rcut_coulomb + pme_lb->rbuf_coulomb;
 272         tmpr_vdw              = pme_lb->rcut_vdw + pme_lb->rbuf_vdw;
 273         set->rlist            = min(tmpr_coulomb, tmpr_vdw);
 274         set->rlistlong        = max(tmpr_coulomb, tmpr_vdw);
 275
 276         /* Set the long-range update frequency */
 277         if (set->rlist == set->rlistlong)
 278         {
 279             /* No long-range interactions if the short-/long-range cutoffs are identical */
 280             set->nstcalclr = 0;
 281         }
 282         else if (pme_lb->nstcalclr_start == 0 || pme_lb->nstcalclr_start == 1)
 283         {
 284             /* We were not doing long-range before, but now we are since rlist!=rlistlong */
 285             set->nstcalclr = 1;
 286         }
 287         else
 288         {
 289             /* We were already doing long-range interactions from the start */
 290             if (pme_lb->rcut_vdw > pme_lb->rcut_coulomb_start)
 291             {
 292                 /* We were originally doing long-range VdW-only interactions.
 293                  * If rvdw is still longer than rcoulomb we keep the original nstcalclr,
 294                  * but if the coulomb cutoff has become longer we should update the long-range
 295                  * part every step.
 296                  */
 297                 set->nstcalclr = (tmpr_vdw > tmpr_coulomb) ? pme_lb->nstcalclr_start : 1;
 298             }
 299             else
 300             {
 301                 /* We were not doing any long-range interaction from the start,
 302                  * since it is not possible to do twin-range coulomb for the PME interaction.
 303                  */
 304                 set->nstcalclr = 1;
 305             }
 306         }
 307     }
 308
 309     set->spacing      = sp;
 310     /* The grid efficiency is the size wrt a grid with uniform x/y/z spacing */
 311     set->grid_efficiency = 1;
 312     for (d = 0; d < DIM; d++)
 313     {
 314         set->grid_efficiency *= (set->grid[d]*sp)/norm(pme_lb->box_start[d]);
 315     }
 316     /* The Ewald coefficient is inversly proportional to the cut-off */
 317     set->ewaldcoeff =
 318         pme_lb->setup[0].ewaldcoeff*pme_lb->setup[0].rcut_coulomb/set->rcut_coulomb;
 319
 320     set->count   = 0;
 321     set->cycles  = 0;
 322
 323     if (debug)
 324     {
 325         fprintf(debug, "PME loadbal: grid %d %d %d, coulomb cutoff %f\n",
 326                 set->grid[XX], set->grid[YY], set->grid[ZZ], set->rcut_coulomb);
 327     }
 328     return TRUE;
 329 }
 330
 331 static void print_grid(FILE *fp_err, FILE *fp_log,
 332                        const char *pre,
 333                        const char *desc,
 334                        const pme_setup_t *set,
 335                        double cycles)
 336 {
 337     char buf[STRLEN], buft[STRLEN];
 338
 339     if (cycles >= 0)
 340     {
 341         sprintf(buft, ": %.1f M-cycles", cycles*1e-6);
 342     }
 343     else
 344     {
 345         buft[0] = '\0';
 346     }
 347     sprintf(buf, "%-11s%10s pme grid %d %d %d, coulomb cutoff %.3f%s",
 348             pre,
 349             desc, set->grid[XX], set->grid[YY], set->grid[ZZ], set->rcut_coulomb,
 350             buft);
 351     if (fp_err != NULL)
 352     {
 353         fprintf(fp_err, "\r%s\n", buf);
 354     }
 355     if (fp_log != NULL)
 356     {
 357         fprintf(fp_log, "%s\n", buf);
 358     }
 359 }
 360
 361 static int pme_loadbal_end(pme_load_balancing_t pme_lb)
 362 {
 363     /* In the initial stage only n is set; end is not set yet */
 364     if (pme_lb->end > 0)
 365     {
 366         return pme_lb->end;
 367     }
 368     else
 369     {
 370         return pme_lb->n;
 371     }
 372 }
 373
 374 static void print_loadbal_limited(FILE *fp_err, FILE *fp_log,
 375                                   gmx_large_int_t step,
 376                                   pme_load_balancing_t pme_lb)
 377 {
 378     char buf[STRLEN], sbuf[22];
 379
 380     sprintf(buf, "step %4s: the %s limits the PME load balancing to a coulomb cut-off of %.3f",
 381             gmx_step_str(step, sbuf),
 382             pmelblim_str[pme_lb->elimited],
 383             pme_lb->setup[pme_loadbal_end(pme_lb)-1].rcut_coulomb);
 384     if (fp_err != NULL)
 385     {
 386         fprintf(fp_err, "\r%s\n", buf);
 387     }
 388     if (fp_log != NULL)
 389     {
 390         fprintf(fp_log, "%s\n", buf);
 391     }
 392 }
 393
 394 static void switch_to_stage1(pme_load_balancing_t pme_lb)
 395 {
 396     pme_lb->start = 0;
 397     while (pme_lb->start+1 < pme_lb->n &&
 398            (pme_lb->setup[pme_lb->start].count == 0 ||
 399             pme_lb->setup[pme_lb->start].cycles >
 400             pme_lb->setup[pme_lb->fastest].cycles*PME_LB_SLOW_FAC))
 401     {
 402         pme_lb->start++;
 403     }
 404     while (pme_lb->start > 0 && pme_lb->setup[pme_lb->start-1].cycles == 0)
 405     {
 406         pme_lb->start--;
 407     }
 408
 409     pme_lb->end = pme_lb->n;
 410     if (pme_lb->setup[pme_lb->end-1].count > 0 &&
 411         pme_lb->setup[pme_lb->end-1].cycles >
 412         pme_lb->setup[pme_lb->fastest].cycles*PME_LB_SLOW_FAC)
 413     {
 414         pme_lb->end--;
 415     }
 416
 417     pme_lb->stage = 1;
 418
 419     /* Next we want to choose setup pme_lb->start, but as we will increase
 420      * pme_ln->cur by one right after returning, we subtract 1 here.
 421      */
 422     pme_lb->cur = pme_lb->start - 1;
 423 }
 424
 425 gmx_bool pme_load_balance(pme_load_balancing_t pme_lb,
 426                           t_commrec           *cr,
 427                           FILE                *fp_err,
 428                           FILE                *fp_log,
 429                           t_inputrec          *ir,
 430                           t_state             *state,
 431                           double               cycles,
 432                           interaction_const_t *ic,
 433                           nonbonded_verlet_t  *nbv,
 434                           gmx_pme_t           *pmedata,
 435                           gmx_large_int_t      step)
 436 {
 437     gmx_bool     OK;
 438     pme_setup_t *set;
 439     double       cycles_fast;
 440     char         buf[STRLEN], sbuf[22];
 441     real         rtab;
 442     gmx_bool     bUsesSimpleTables = TRUE;
 443
 444     if (pme_lb->stage == pme_lb->nstage)
 445     {
 446         return FALSE;
 447     }
 448
 449     if (PAR(cr))
 450     {
 451         gmx_sumd(1, &cycles, cr);
 452         cycles /= cr->nnodes;
 453     }
 454
 455     set = &pme_lb->setup[pme_lb->cur];
 456     set->count++;
 457
 458     rtab = ir->rlistlong + ir->tabext;
 459
 460     if (set->count % 2 == 1)
 461     {
 462         /* Skip the first cycle, because the first step after a switch
 463          * is much slower due to allocation and/or caching effects.
 464          */
 465         return TRUE;
 466     }
 467
 468     sprintf(buf, "step %4s: ", gmx_step_str(step, sbuf));
 469     print_grid(fp_err, fp_log, buf, "timed with", set, cycles);
 470
 471     if (set->count <= 2)
 472     {
 473         set->cycles = cycles;
 474     }
 475     else
 476     {
 477         if (cycles*PME_LB_ACCEL_TOL < set->cycles &&
 478             pme_lb->stage == pme_lb->nstage - 1)
 479         {
 480             /* The performance went up a lot (due to e.g. DD load balancing).
 481              * Add a stage, keep the minima, but rescan all setups.
 482              */
 483             pme_lb->nstage++;
 484
 485             if (debug)
 486             {
 487                 fprintf(debug, "The performance for grid %d %d %d went from %.3f to %.1f M-cycles, this is more than %f\n"
 488                         "Increased the number stages to %d"
 489                         " and ignoring the previous performance\n",
 490                         set->grid[XX], set->grid[YY], set->grid[ZZ],
 491                         cycles*1e-6, set->cycles*1e-6, PME_LB_ACCEL_TOL,
 492                         pme_lb->nstage);
 493             }
 494         }
 495         set->cycles = min(set->cycles, cycles);
 496     }
 497
 498     if (set->cycles < pme_lb->setup[pme_lb->fastest].cycles)
 499     {
 500         pme_lb->fastest = pme_lb->cur;
 501
 502         if (DOMAINDECOMP(cr))
 503         {
 504             /* We found a new fastest setting, ensure that with subsequent
 505              * shorter cut-off's the dynamic load balancing does not make
 506              * the use of the current cut-off impossible. This solution is
 507              * a trade-off, as the PME load balancing and DD domain size
 508              * load balancing can interact in complex ways.
 509              * With the Verlet kernels, DD load imbalance will usually be
 510              * mainly due to bonded interaction imbalance, which will often
 511              * quickly push the domain boundaries beyond the limit for the
 512              * optimal, PME load balanced, cut-off. But it could be that
 513              * better overal performance can be obtained with a slightly
 514              * shorter cut-off and better DD load balancing.
 515              */
 516             change_dd_dlb_cutoff_limit(cr);
 517         }
 518     }
 519     cycles_fast = pme_lb->setup[pme_lb->fastest].cycles;
 520
 521     /* Check in stage 0 if we should stop scanning grids.
 522      * Stop when the time is more than SLOW_FAC longer than the fastest.
 523      */
 524     if (pme_lb->stage == 0 && pme_lb->cur > 0 &&
 525         cycles > pme_lb->setup[pme_lb->fastest].cycles*PME_LB_SLOW_FAC)
 526     {
 527         pme_lb->n = pme_lb->cur + 1;
 528         /* Done with scanning, go to stage 1 */
 529         switch_to_stage1(pme_lb);
 530     }
 531
 532     if (pme_lb->stage == 0)
 533     {
 534         int gridsize_start;
 535
 536         gridsize_start = set->grid[XX]*set->grid[YY]*set->grid[ZZ];
 537
 538         do
 539         {
 540             if (pme_lb->cur+1 < pme_lb->n)
 541             {
 542                 /* We had already generated the next setup */
 543                 OK = TRUE;
 544             }
 545             else
 546             {
 547                 /* Find the next setup */
 548                 OK = pme_loadbal_increase_cutoff(pme_lb, ir->pme_order, cr->dd);
 549
 550                 if (!OK)
 551                 {
 552                     pme_lb->elimited = epmelblimPMEGRID;
 553                 }
 554             }
 555
 556             if (OK && ir->ePBC != epbcNONE)
 557             {
 558                 OK = (sqr(pme_lb->setup[pme_lb->cur+1].rlistlong)
 559                       <= max_cutoff2(ir->ePBC, state->box));
 560                 if (!OK)
 561                 {
 562                     pme_lb->elimited = epmelblimBOX;
 563                 }
 564             }
 565
 566             if (OK)
 567             {
 568                 pme_lb->cur++;
 569
 570                 if (DOMAINDECOMP(cr))
 571                 {
 572                     OK = change_dd_cutoff(cr, state, ir,
 573                                           pme_lb->setup[pme_lb->cur].rlistlong);
 574                     if (!OK)
 575                     {
 576                         /* Failed: do not use this setup */
 577                         pme_lb->cur--;
 578                         pme_lb->elimited = epmelblimDD;
 579                     }
 580                 }
 581             }
 582             if (!OK)
 583             {
 584                 /* We hit the upper limit for the cut-off,
 585                  * the setup should not go further than cur.
 586                  */
 587                 pme_lb->n = pme_lb->cur + 1;
 588                 print_loadbal_limited(fp_err, fp_log, step, pme_lb);
 589                 /* Switch to the next stage */
 590                 switch_to_stage1(pme_lb);
 591             }
 592         }
 593         while (OK &&
 594                !(pme_lb->setup[pme_lb->cur].grid[XX]*
 595                  pme_lb->setup[pme_lb->cur].grid[YY]*
 596                  pme_lb->setup[pme_lb->cur].grid[ZZ] <
 597                  gridsize_start*PME_LB_GRID_SCALE_FAC
 598                  &&
 599                  pme_lb->setup[pme_lb->cur].grid_efficiency <
 600                  pme_lb->setup[pme_lb->cur-1].grid_efficiency*PME_LB_GRID_EFFICIENCY_REL_FAC));
 601     }
 602
 603     if (pme_lb->stage > 0 && pme_lb->end == 1)
 604     {
 605         pme_lb->cur   = 0;
 606         pme_lb->stage = pme_lb->nstage;
 607     }
 608     else if (pme_lb->stage > 0 && pme_lb->end > 1)
 609     {
 610         /* If stage = nstage-1:
 611          *   scan over all setups, rerunning only those setups
 612          *   which are not much slower than the fastest
 613          * else:
 614          *   use the next setup
 615          */
 616         do
 617         {
 618             pme_lb->cur++;
 619             if (pme_lb->cur == pme_lb->end)
 620             {
 621                 pme_lb->stage++;
 622                 pme_lb->cur = pme_lb->start;
 623             }
 624         }
 625         while (pme_lb->stage == pme_lb->nstage - 1 &&
 626                pme_lb->setup[pme_lb->cur].count > 0 &&
 627                pme_lb->setup[pme_lb->cur].cycles > cycles_fast*PME_LB_SLOW_FAC);
 628
 629         if (pme_lb->stage == pme_lb->nstage)
 630         {
 631             /* We are done optimizing, use the fastest setup we found */
 632             pme_lb->cur = pme_lb->fastest;
 633         }
 634     }
 635
 636     if (DOMAINDECOMP(cr) && pme_lb->stage > 0)
 637     {
 638         OK = change_dd_cutoff(cr, state, ir, pme_lb->setup[pme_lb->cur].rlistlong);
 639         if (!OK)
 640         {
 641             /* Failsafe solution */
 642             if (pme_lb->cur > 1 && pme_lb->stage == pme_lb->nstage)
 643             {
 644                 pme_lb->stage--;
 645             }
 646             pme_lb->fastest  = 0;
 647             pme_lb->start    = 0;
 648             pme_lb->end      = pme_lb->cur;
 649             pme_lb->cur      = pme_lb->start;
 650             pme_lb->elimited = epmelblimDD;
 651             print_loadbal_limited(fp_err, fp_log, step, pme_lb);
 652         }
 653     }
 654
 655     /* Change the Coulomb cut-off and the PME grid */
 656
 657     set = &pme_lb->setup[pme_lb->cur];
 658
 659     ic->rcoulomb   = set->rcut_coulomb;
 660     ic->rlist      = set->rlist;
 661     ic->rlistlong  = set->rlistlong;
 662     ir->nstcalclr  = set->nstcalclr;
 663     ic->ewaldcoeff = set->ewaldcoeff;
 664
 665     bUsesSimpleTables = uses_simple_tables(ir->cutoff_scheme, nbv, 0);
 666     if (pme_lb->cutoff_scheme == ecutsVERLET &&
 667         nbv->grp[0].kernel_type == nbnxnk8x8x8_CUDA)
 668     {
 669         nbnxn_cuda_pme_loadbal_update_param(nbv->cu_nbv, ic);
 670
 671         /* With tMPI + GPUs some ranks may be sharing GPU(s) and therefore
 672          * also sharing texture references. To keep the code simple, we don't
 673          * treat texture references as shared resources, but this means that
 674          * the coulomb_tab texture ref will get updated by multiple threads.
 675          * Hence, to ensure that the non-bonded kernels don't start before all
 676          * texture binding operations are finished, we need to wait for all ranks
 677          * to arrive here before continuing.
 678          *
 679          * Note that we could omit this barrier if GPUs are not shared (or
 680          * texture objects are used), but as this is initialization code, there
 681          * is not point in complicating things.
 682          */
 683 #ifdef GMX_THREAD_MPI
 684         if (PAR(cr))
 685         {
 686             gmx_barrier(cr);
 687         }
 688 #endif /* GMX_THREAD_MPI */
 689     }
 690     else
 691     {
 692         init_interaction_const_tables(NULL, ic, bUsesSimpleTables,
 693                                       rtab);
 694     }
 695
 696     if (pme_lb->cutoff_scheme == ecutsVERLET && nbv->ngrp > 1)
 697     {
 698         init_interaction_const_tables(NULL, ic, bUsesSimpleTables,
 699                                       rtab);
 700     }
 701
 702     if (cr->duty & DUTY_PME)
 703     {
 704         if (pme_lb->setup[pme_lb->cur].pmedata == NULL)
 705         {
 706             /* Generate a new PME data structure,
 707              * copying part of the old pointers.
 708              */
 709             gmx_pme_reinit(&set->pmedata,
 710                            cr, pme_lb->setup[0].pmedata, ir,
 711                            set->grid);
 712         }
 713         *pmedata = set->pmedata;
 714     }
 715     else
 716     {
 717         /* Tell our PME-only node to switch grid */
 718         gmx_pme_send_switchgrid(cr, set->grid, set->ewaldcoeff);
 719     }
 720
 721     if (debug)
 722     {
 723         print_grid(NULL, debug, "", "switched to", set, -1);
 724     }
 725
 726     if (pme_lb->stage == pme_lb->nstage)
 727     {
 728         print_grid(fp_err, fp_log, "", "optimal", set, -1);
 729     }
 730
 731     return TRUE;
 732 }
 733
 734 void restart_pme_loadbal(pme_load_balancing_t pme_lb, int n)
 735 {
 736     pme_lb->nstage += n;
 737 }
 738
 739 static int pme_grid_points(const pme_setup_t *setup)
 740 {
 741     return setup->grid[XX]*setup->grid[YY]*setup->grid[ZZ];
 742 }
 743
 744 static real pme_loadbal_rlist(const pme_setup_t *setup)
 745 {
 746     /* With the group cut-off scheme we can have twin-range either
 747      * for Coulomb or for VdW, so we need a check here.
 748      * With the Verlet cut-off scheme rlist=rlistlong.
 749      */
 750     if (setup->rcut_coulomb > setup->rlist)
 751     {
 752         return setup->rlistlong;
 753     }
 754     else
 755     {
 756         return setup->rlist;
 757     }
 758 }
 759
 760 static void print_pme_loadbal_setting(FILE              *fplog,
 761                                       char              *name,
 762                                       const pme_setup_t *setup)
 763 {
 764     fprintf(fplog,
 765             "   %-7s %6.3f nm %6.3f nm     %3d %3d %3d   %5.3f nm  %5.3f nm\n",
 766             name,
 767             setup->rcut_coulomb, pme_loadbal_rlist(setup),
 768             setup->grid[XX], setup->grid[YY], setup->grid[ZZ],
 769             setup->spacing, 1/setup->ewaldcoeff);
 770 }
 771
 772 static void print_pme_loadbal_settings(pme_load_balancing_t pme_lb,
 773                                        t_commrec           *cr,
 774                                        FILE                *fplog,
 775                                        gmx_bool             bNonBondedOnGPU)
 776 {
 777     double pp_ratio, grid_ratio;
 778
 779     pp_ratio   = pow(pme_loadbal_rlist(&pme_lb->setup[pme_lb->cur])/pme_loadbal_rlist(&pme_lb->setup[0]), 3.0);
 780     grid_ratio = pme_grid_points(&pme_lb->setup[pme_lb->cur])/
 781         (double)pme_grid_points(&pme_lb->setup[0]);
 782
 783     fprintf(fplog, "\n");
 784     fprintf(fplog, "       P P   -   P M E   L O A D   B A L A N C I N G\n");
 785     fprintf(fplog, "\n");
 786     /* Here we only warn when the optimal setting is the last one */
 787     if (pme_lb->elimited != epmelblimNO &&
 788         pme_lb->cur == pme_loadbal_end(pme_lb)-1)
 789     {
 790         fprintf(fplog, " NOTE: The PP/PME load balancing was limited by the %s,\n",
 791                 pmelblim_str[pme_lb->elimited]);
 792         fprintf(fplog, "       you might not have reached a good load balance.\n");
 793         if (pme_lb->elimited == epmelblimDD)
 794         {
 795             fprintf(fplog, "       Try different mdrun -dd settings or lower the -dds value.\n");
 796         }
 797         fprintf(fplog, "\n");
 798     }
 799     fprintf(fplog, " PP/PME load balancing changed the cut-off and PME settings:\n");
 800     fprintf(fplog, "           particle-particle                    PME\n");
 801     fprintf(fplog, "            rcoulomb  rlist            grid      spacing   1/beta\n");
 802     print_pme_loadbal_setting(fplog, "initial", &pme_lb->setup[0]);
 803     print_pme_loadbal_setting(fplog, "final", &pme_lb->setup[pme_lb->cur]);
 804     fprintf(fplog, " cost-ratio           %4.2f             %4.2f\n",
 805             pp_ratio, grid_ratio);
 806     fprintf(fplog, " (note that these numbers concern only part of the total PP and PME load)\n");
 807
 808     if (pp_ratio > 1.5 && !bNonBondedOnGPU)
 809     {
 810         md_print_warn(cr, fplog,
 811                       "NOTE: PME load balancing increased the non-bonded workload by more than 50%%.\n"
 812                       "      For better performance use (more) PME nodes (mdrun -npme),\n"
 813                       "      or in case you are beyond the scaling limit, use less nodes in total.\n");
 814     }
 815     else
 816     {
 817         fprintf(fplog, "\n");
 818     }
 819 }
 820
 821 void pme_loadbal_done(pme_load_balancing_t pme_lb,
 822                       t_commrec *cr, FILE *fplog,
 823                       gmx_bool bNonBondedOnGPU)
 824 {
 825     if (fplog != NULL && (pme_lb->cur > 0 || pme_lb->elimited != epmelblimNO))
 826     {
 827         print_pme_loadbal_settings(pme_lb, cr, fplog, bNonBondedOnGPU);
 828     }
 829
 830     /* TODO: Here we should free all pointers in pme_lb,
 831      * but as it contains pme data structures,
 832      * we need to first make pme.c free all data.
 833      */
 834 }