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 "types/commrec.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 #include "gromacs/utility/cstringutil.h"
  52 #include "gromacs/utility/smalloc.h"
  53
  54 /* Parameters and setting for one PP-PME setup */
  55 typedef struct {
  56     real      rcut_coulomb;    /* Coulomb cut-off                              */
  57     real      rlist;           /* pair-list cut-off                            */
  58     real      rlistlong;       /* LR pair-list cut-off                         */
  59     int       nstcalclr;       /* frequency of evaluating long-range forces for group scheme */
  60     real      spacing;         /* (largest) PME grid spacing                   */
  61     ivec      grid;            /* the PME grid dimensions                      */
  62     real      grid_efficiency; /* ineffiency factor for non-uniform grids <= 1 */
  63     real      ewaldcoeff_q;    /* Electrostatic Ewald coefficient            */
  64     real      ewaldcoeff_lj;   /* LJ Ewald coefficient, only for the call to send_switchgrid */
  65     gmx_pme_t pmedata;         /* the data structure used in the PME code      */
  66     int       count;           /* number of times this setup has been timed    */
  67     double    cycles;          /* the fastest time for this setup in cycles    */
  68 } pme_setup_t;
  69
  70 /* In the initial scan, step by grids that are at least a factor 0.8 coarser */
  71 #define PME_LB_GRID_SCALE_FAC  0.8
  72 /* In the initial scan, try to skip grids with uneven x/y/z spacing,
  73  * checking if the "efficiency" is more than 5% worse than the previous grid.
  74  */
  75 #define PME_LB_GRID_EFFICIENCY_REL_FAC  1.05
  76 /* Rerun up till 12% slower setups than the fastest up till now */
  77 #define PME_LB_SLOW_FAC  1.12
  78 /* If setups get more than 2% faster, do another round to avoid
  79  * choosing a slower setup due to acceleration or fluctuations.
  80  */
  81 #define PME_LB_ACCEL_TOL 1.02
  82
  83 enum {
  84     epmelblimNO, epmelblimBOX, epmelblimDD, epmelblimPMEGRID, epmelblimNR
  85 };
  86
  87 const char *pmelblim_str[epmelblimNR] =
  88 { "no", "box size", "domain decompostion", "PME grid restriction" };
  89
  90 struct pme_load_balancing {
  91     int          nstage;             /* the current maximum number of stages */
  92
  93     real         cut_spacing;        /* the minimum cutoff / PME grid spacing ratio */
  94     real         rcut_vdw;           /* Vdw cutoff (does not change) */
  95     real         rcut_coulomb_start; /* Initial electrostatics cutoff */
  96     int          nstcalclr_start;    /* Initial electrostatics cutoff */
  97     real         rbuf_coulomb;       /* the pairlist buffer size */
  98     real         rbuf_vdw;           /* the pairlist buffer size */
  99     matrix       box_start;          /* the initial simulation box */
 100     int          n;                  /* the count of setup as well as the allocation size */
 101     pme_setup_t *setup;              /* the PME+cutoff setups */
 102     int          cur;                /* the current setup */
 103     int          fastest;            /* fastest setup up till now */
 104     int          start;              /* start of setup range to consider in stage>0 */
 105     int          end;                /* end   of setup range to consider in stage>0 */
 106     int          elimited;           /* was the balancing limited, uses enum above */
 107     int          cutoff_scheme;      /* Verlet or group cut-offs */
 108
 109     int          stage;              /* the current stage */
 110 };
 111
 112 void pme_loadbal_init(pme_load_balancing_t *pme_lb_p,
 113                       const t_inputrec *ir, matrix box,
 114                       const interaction_const_t *ic,
 115                       gmx_pme_t pmedata)
 116 {
 117     pme_load_balancing_t pme_lb;
 118     real                 spm, sp;
 119     int                  d;
 120
 121     snew(pme_lb, 1);
 122
 123     /* Any number of stages >= 2 is supported */
 124     pme_lb->nstage   = 2;
 125
 126     pme_lb->cutoff_scheme = ir->cutoff_scheme;
 127
 128     if (pme_lb->cutoff_scheme == ecutsVERLET)
 129     {
 130         pme_lb->rbuf_coulomb = ic->rlist - ic->rcoulomb;
 131         pme_lb->rbuf_vdw     = pme_lb->rbuf_coulomb;
 132     }
 133     else
 134     {
 135         if (ic->rcoulomb > ic->rlist)
 136         {
 137             pme_lb->rbuf_coulomb = ic->rlistlong - ic->rcoulomb;
 138         }
 139         else
 140         {
 141             pme_lb->rbuf_coulomb = ic->rlist - ic->rcoulomb;
 142         }
 143         if (ic->rvdw > ic->rlist)
 144         {
 145             pme_lb->rbuf_vdw = ic->rlistlong - ic->rvdw;
 146         }
 147         else
 148         {
 149             pme_lb->rbuf_vdw = ic->rlist - ic->rvdw;
 150         }
 151     }
 152
 153     copy_mat(box, pme_lb->box_start);
 154     if (ir->ePBC == epbcXY && ir->nwall == 2)
 155     {
 156         svmul(ir->wall_ewald_zfac, pme_lb->box_start[ZZ], pme_lb->box_start[ZZ]);
 157     }
 158
 159     pme_lb->n = 1;
 160     snew(pme_lb->setup, pme_lb->n);
 161
 162     pme_lb->rcut_vdw                 = ic->rvdw;
 163     pme_lb->rcut_coulomb_start       = ir->rcoulomb;
 164     pme_lb->nstcalclr_start          = ir->nstcalclr;
 165
 166     pme_lb->cur                      = 0;
 167     pme_lb->setup[0].rcut_coulomb    = ic->rcoulomb;
 168     pme_lb->setup[0].rlist           = ic->rlist;
 169     pme_lb->setup[0].rlistlong       = ic->rlistlong;
 170     pme_lb->setup[0].nstcalclr       = ir->nstcalclr;
 171     pme_lb->setup[0].grid[XX]        = ir->nkx;
 172     pme_lb->setup[0].grid[YY]        = ir->nky;
 173     pme_lb->setup[0].grid[ZZ]        = ir->nkz;
 174     pme_lb->setup[0].ewaldcoeff_q    = ic->ewaldcoeff_q;
 175     pme_lb->setup[0].ewaldcoeff_lj   = ic->ewaldcoeff_lj;
 176
 177     pme_lb->setup[0].pmedata  = pmedata;
 178
 179     spm = 0;
 180     for (d = 0; d < DIM; d++)
 181     {
 182         sp = norm(pme_lb->box_start[d])/pme_lb->setup[0].grid[d];
 183         if (sp > spm)
 184         {
 185             spm = sp;
 186         }
 187     }
 188     pme_lb->setup[0].spacing = spm;
 189
 190     if (ir->fourier_spacing > 0)
 191     {
 192         pme_lb->cut_spacing = ir->rcoulomb/ir->fourier_spacing;
 193     }
 194     else
 195     {
 196         pme_lb->cut_spacing = ir->rcoulomb/pme_lb->setup[0].spacing;
 197     }
 198
 199     pme_lb->stage = 0;
 200
 201     pme_lb->fastest  = 0;
 202     pme_lb->start    = 0;
 203     pme_lb->end      = 0;
 204     pme_lb->elimited = epmelblimNO;
 205
 206     *pme_lb_p = pme_lb;
 207 }
 208
 209 static gmx_bool pme_loadbal_increase_cutoff(pme_load_balancing_t  pme_lb,
 210                                             int                   pme_order,
 211                                             const gmx_domdec_t   *dd)
 212 {
 213     pme_setup_t *set;
 214     int          npmenodes_x, npmenodes_y;
 215     real         fac, sp;
 216     real         tmpr_coulomb, tmpr_vdw;
 217     int          d;
 218     gmx_bool     grid_ok;
 219
 220     /* Try to add a new setup with next larger cut-off to the list */
 221     pme_lb->n++;
 222     srenew(pme_lb->setup, pme_lb->n);
 223     set          = &pme_lb->setup[pme_lb->n-1];
 224     set->pmedata = NULL;
 225
 226     get_pme_nnodes(dd, &npmenodes_x, &npmenodes_y);
 227
 228     fac = 1;
 229     do
 230     {
 231         /* Avoid infinite while loop, which can occur at the minimum grid size.
 232          * Note that in practice load balancing will stop before this point.
 233          * The factor 2.1 allows for the extreme case in which only grids
 234          * of powers of 2 are allowed (the current code supports more grids).
 235          */
 236         if (fac > 2.1)
 237         {
 238             pme_lb->n--;
 239
 240             return FALSE;
 241         }
 242
 243         fac *= 1.01;
 244         clear_ivec(set->grid);
 245         sp = calc_grid(NULL, pme_lb->box_start,
 246                        fac*pme_lb->setup[pme_lb->cur].spacing,
 247                        &set->grid[XX],
 248                        &set->grid[YY],
 249                        &set->grid[ZZ]);
 250
 251         /* As here we can't easily check if one of the PME nodes
 252          * uses threading, we do a conservative grid check.
 253          * This means we can't use pme_order or less grid lines
 254          * per PME node along x, which is not a strong restriction.
 255          */
 256         gmx_pme_check_restrictions(pme_order,
 257                                    set->grid[XX], set->grid[YY], set->grid[ZZ],
 258                                    npmenodes_x, npmenodes_y,
 259                                    TRUE,
 260                                    FALSE,
 261                                    &grid_ok);
 262     }
 263     while (sp <= 1.001*pme_lb->setup[pme_lb->cur].spacing || !grid_ok);
 264
 265     set->rcut_coulomb = pme_lb->cut_spacing*sp;
 266
 267     if (pme_lb->cutoff_scheme == ecutsVERLET)
 268     {
 269         set->rlist        = set->rcut_coulomb + pme_lb->rbuf_coulomb;
 270         /* We dont use LR lists with Verlet, but this avoids if-statements in further checks */
 271         set->rlistlong    = set->rlist;
 272     }
 273     else
 274     {
 275         tmpr_coulomb          = set->rcut_coulomb + pme_lb->rbuf_coulomb;
 276         tmpr_vdw              = pme_lb->rcut_vdw + pme_lb->rbuf_vdw;
 277         set->rlist            = min(tmpr_coulomb, tmpr_vdw);
 278         set->rlistlong        = max(tmpr_coulomb, tmpr_vdw);
 279
 280         /* Set the long-range update frequency */
 281         if (set->rlist == set->rlistlong)
 282         {
 283             /* No long-range interactions if the short-/long-range cutoffs are identical */
 284             set->nstcalclr = 0;
 285         }
 286         else if (pme_lb->nstcalclr_start == 0 || pme_lb->nstcalclr_start == 1)
 287         {
 288             /* We were not doing long-range before, but now we are since rlist!=rlistlong */
 289             set->nstcalclr = 1;
 290         }
 291         else
 292         {
 293             /* We were already doing long-range interactions from the start */
 294             if (pme_lb->rcut_vdw > pme_lb->rcut_coulomb_start)
 295             {
 296                 /* We were originally doing long-range VdW-only interactions.
 297                  * If rvdw is still longer than rcoulomb we keep the original nstcalclr,
 298                  * but if the coulomb cutoff has become longer we should update the long-range
 299                  * part every step.
 300                  */
 301                 set->nstcalclr = (tmpr_vdw > tmpr_coulomb) ? pme_lb->nstcalclr_start : 1;
 302             }
 303             else
 304             {
 305                 /* We were not doing any long-range interaction from the start,
 306                  * since it is not possible to do twin-range coulomb for the PME interaction.
 307                  */
 308                 set->nstcalclr = 1;
 309             }
 310         }
 311     }
 312
 313     set->spacing      = sp;
 314     /* The grid efficiency is the size wrt a grid with uniform x/y/z spacing */
 315     set->grid_efficiency = 1;
 316     for (d = 0; d < DIM; d++)
 317     {
 318         set->grid_efficiency *= (set->grid[d]*sp)/norm(pme_lb->box_start[d]);
 319     }
 320     /* The Ewald coefficient is inversly proportional to the cut-off */
 321     set->ewaldcoeff_q =
 322         pme_lb->setup[0].ewaldcoeff_q*pme_lb->setup[0].rcut_coulomb/set->rcut_coulomb;
 323     /* We set ewaldcoeff_lj in set, even when LJ-PME is not used */
 324     set->ewaldcoeff_lj =
 325         pme_lb->setup[0].ewaldcoeff_lj*pme_lb->setup[0].rcut_coulomb/set->rcut_coulomb;
 326
 327     set->count   = 0;
 328     set->cycles  = 0;
 329
 330     if (debug)
 331     {
 332         fprintf(debug, "PME loadbal: grid %d %d %d, coulomb cutoff %f\n",
 333                 set->grid[XX], set->grid[YY], set->grid[ZZ], set->rcut_coulomb);
 334     }
 335     return TRUE;
 336 }
 337
 338 static void print_grid(FILE *fp_err, FILE *fp_log,
 339                        const char *pre,
 340                        const char *desc,
 341                        const pme_setup_t *set,
 342                        double cycles)
 343 {
 344     char buf[STRLEN], buft[STRLEN];
 345
 346     if (cycles >= 0)
 347     {
 348         sprintf(buft, ": %.1f M-cycles", cycles*1e-6);
 349     }
 350     else
 351     {
 352         buft[0] = '\0';
 353     }
 354     sprintf(buf, "%-11s%10s pme grid %d %d %d, coulomb cutoff %.3f%s",
 355             pre,
 356             desc, set->grid[XX], set->grid[YY], set->grid[ZZ], set->rcut_coulomb,
 357             buft);
 358     if (fp_err != NULL)
 359     {
 360         fprintf(fp_err, "\r%s\n", buf);
 361     }
 362     if (fp_log != NULL)
 363     {
 364         fprintf(fp_log, "%s\n", buf);
 365     }
 366 }
 367
 368 static int pme_loadbal_end(pme_load_balancing_t pme_lb)
 369 {
 370     /* In the initial stage only n is set; end is not set yet */
 371     if (pme_lb->end > 0)
 372     {
 373         return pme_lb->end;
 374     }
 375     else
 376     {
 377         return pme_lb->n;
 378     }
 379 }
 380
 381 static void print_loadbal_limited(FILE *fp_err, FILE *fp_log,
 382                                   gmx_int64_t step,
 383                                   pme_load_balancing_t pme_lb)
 384 {
 385     char buf[STRLEN], sbuf[22];
 386
 387     sprintf(buf, "step %4s: the %s limits the PME load balancing to a coulomb cut-off of %.3f",
 388             gmx_step_str(step, sbuf),
 389             pmelblim_str[pme_lb->elimited],
 390             pme_lb->setup[pme_loadbal_end(pme_lb)-1].rcut_coulomb);
 391     if (fp_err != NULL)
 392     {
 393         fprintf(fp_err, "\r%s\n", buf);
 394     }
 395     if (fp_log != NULL)
 396     {
 397         fprintf(fp_log, "%s\n", buf);
 398     }
 399 }
 400
 401 static void switch_to_stage1(pme_load_balancing_t pme_lb)
 402 {
 403     pme_lb->start = 0;
 404     while (pme_lb->start+1 < pme_lb->n &&
 405            (pme_lb->setup[pme_lb->start].count == 0 ||
 406             pme_lb->setup[pme_lb->start].cycles >
 407             pme_lb->setup[pme_lb->fastest].cycles*PME_LB_SLOW_FAC))
 408     {
 409         pme_lb->start++;
 410     }
 411     while (pme_lb->start > 0 && pme_lb->setup[pme_lb->start-1].cycles == 0)
 412     {
 413         pme_lb->start--;
 414     }
 415
 416     pme_lb->end = pme_lb->n;
 417     if (pme_lb->setup[pme_lb->end-1].count > 0 &&
 418         pme_lb->setup[pme_lb->end-1].cycles >
 419         pme_lb->setup[pme_lb->fastest].cycles*PME_LB_SLOW_FAC)
 420     {
 421         pme_lb->end--;
 422     }
 423
 424     pme_lb->stage = 1;
 425
 426     /* Next we want to choose setup pme_lb->start, but as we will increase
 427      * pme_ln->cur by one right after returning, we subtract 1 here.
 428      */
 429     pme_lb->cur = pme_lb->start - 1;
 430 }
 431
 432 gmx_bool pme_load_balance(pme_load_balancing_t pme_lb,
 433                           t_commrec           *cr,
 434                           FILE                *fp_err,
 435                           FILE                *fp_log,
 436                           t_inputrec          *ir,
 437                           t_state             *state,
 438                           double               cycles,
 439                           interaction_const_t *ic,
 440                           nonbonded_verlet_t  *nbv,
 441                           gmx_pme_t           *pmedata,
 442                           gmx_int64_t          step)
 443 {
 444     gmx_bool     OK;
 445     pme_setup_t *set;
 446     double       cycles_fast;
 447     char         buf[STRLEN], sbuf[22];
 448     real         rtab;
 449     gmx_bool     bUsesSimpleTables = TRUE;
 450
 451     if (pme_lb->stage == pme_lb->nstage)
 452     {
 453         return FALSE;
 454     }
 455
 456     if (PAR(cr))
 457     {
 458         gmx_sumd(1, &cycles, cr);
 459         cycles /= cr->nnodes;
 460     }
 461
 462     set = &pme_lb->setup[pme_lb->cur];
 463     set->count++;
 464
 465     rtab = ir->rlistlong + ir->tabext;
 466
 467     if (set->count % 2 == 1)
 468     {
 469         /* Skip the first cycle, because the first step after a switch
 470          * is much slower due to allocation and/or caching effects.
 471          */
 472         return TRUE;
 473     }
 474
 475     sprintf(buf, "step %4s: ", gmx_step_str(step, sbuf));
 476     print_grid(fp_err, fp_log, buf, "timed with", set, cycles);
 477
 478     if (set->count <= 2)
 479     {
 480         set->cycles = cycles;
 481     }
 482     else
 483     {
 484         if (cycles*PME_LB_ACCEL_TOL < set->cycles &&
 485             pme_lb->stage == pme_lb->nstage - 1)
 486         {
 487             /* The performance went up a lot (due to e.g. DD load balancing).
 488              * Add a stage, keep the minima, but rescan all setups.
 489              */
 490             pme_lb->nstage++;
 491
 492             if (debug)
 493             {
 494                 fprintf(debug, "The performance for grid %d %d %d went from %.3f to %.1f M-cycles, this is more than %f\n"
 495                         "Increased the number stages to %d"
 496                         " and ignoring the previous performance\n",
 497                         set->grid[XX], set->grid[YY], set->grid[ZZ],
 498                         cycles*1e-6, set->cycles*1e-6, PME_LB_ACCEL_TOL,
 499                         pme_lb->nstage);
 500             }
 501         }
 502         set->cycles = min(set->cycles, cycles);
 503     }
 504
 505     if (set->cycles < pme_lb->setup[pme_lb->fastest].cycles)
 506     {
 507         pme_lb->fastest = pme_lb->cur;
 508
 509         if (DOMAINDECOMP(cr))
 510         {
 511             /* We found a new fastest setting, ensure that with subsequent
 512              * shorter cut-off's the dynamic load balancing does not make
 513              * the use of the current cut-off impossible. This solution is
 514              * a trade-off, as the PME load balancing and DD domain size
 515              * load balancing can interact in complex ways.
 516              * With the Verlet kernels, DD load imbalance will usually be
 517              * mainly due to bonded interaction imbalance, which will often
 518              * quickly push the domain boundaries beyond the limit for the
 519              * optimal, PME load balanced, cut-off. But it could be that
 520              * better overal performance can be obtained with a slightly
 521              * shorter cut-off and better DD load balancing.
 522              */
 523             change_dd_dlb_cutoff_limit(cr);
 524         }
 525     }
 526     cycles_fast = pme_lb->setup[pme_lb->fastest].cycles;
 527
 528     /* Check in stage 0 if we should stop scanning grids.
 529      * Stop when the time is more than SLOW_FAC longer than the fastest.
 530      */
 531     if (pme_lb->stage == 0 && pme_lb->cur > 0 &&
 532         cycles > pme_lb->setup[pme_lb->fastest].cycles*PME_LB_SLOW_FAC)
 533     {
 534         pme_lb->n = pme_lb->cur + 1;
 535         /* Done with scanning, go to stage 1 */
 536         switch_to_stage1(pme_lb);
 537     }
 538
 539     if (pme_lb->stage == 0)
 540     {
 541         int gridsize_start;
 542
 543         gridsize_start = set->grid[XX]*set->grid[YY]*set->grid[ZZ];
 544
 545         do
 546         {
 547             if (pme_lb->cur+1 < pme_lb->n)
 548             {
 549                 /* We had already generated the next setup */
 550                 OK = TRUE;
 551             }
 552             else
 553             {
 554                 /* Find the next setup */
 555                 OK = pme_loadbal_increase_cutoff(pme_lb, ir->pme_order, cr->dd);
 556
 557                 if (!OK)
 558                 {
 559                     pme_lb->elimited = epmelblimPMEGRID;
 560                 }
 561             }
 562
 563             if (OK && ir->ePBC != epbcNONE)
 564             {
 565                 OK = (sqr(pme_lb->setup[pme_lb->cur+1].rlistlong)
 566                       <= max_cutoff2(ir->ePBC, state->box));
 567                 if (!OK)
 568                 {
 569                     pme_lb->elimited = epmelblimBOX;
 570                 }
 571             }
 572
 573             if (OK)
 574             {
 575                 pme_lb->cur++;
 576
 577                 if (DOMAINDECOMP(cr))
 578                 {
 579                     OK = change_dd_cutoff(cr, state, ir,
 580                                           pme_lb->setup[pme_lb->cur].rlistlong);
 581                     if (!OK)
 582                     {
 583                         /* Failed: do not use this setup */
 584                         pme_lb->cur--;
 585                         pme_lb->elimited = epmelblimDD;
 586                     }
 587                 }
 588             }
 589             if (!OK)
 590             {
 591                 /* We hit the upper limit for the cut-off,
 592                  * the setup should not go further than cur.
 593                  */
 594                 pme_lb->n = pme_lb->cur + 1;
 595                 print_loadbal_limited(fp_err, fp_log, step, pme_lb);
 596                 /* Switch to the next stage */
 597                 switch_to_stage1(pme_lb);
 598             }
 599         }
 600         while (OK &&
 601                !(pme_lb->setup[pme_lb->cur].grid[XX]*
 602                  pme_lb->setup[pme_lb->cur].grid[YY]*
 603                  pme_lb->setup[pme_lb->cur].grid[ZZ] <
 604                  gridsize_start*PME_LB_GRID_SCALE_FAC
 605                  &&
 606                  pme_lb->setup[pme_lb->cur].grid_efficiency <
 607                  pme_lb->setup[pme_lb->cur-1].grid_efficiency*PME_LB_GRID_EFFICIENCY_REL_FAC));
 608     }
 609
 610     if (pme_lb->stage > 0 && pme_lb->end == 1)
 611     {
 612         pme_lb->cur   = 0;
 613         pme_lb->stage = pme_lb->nstage;
 614     }
 615     else if (pme_lb->stage > 0 && pme_lb->end > 1)
 616     {
 617         /* If stage = nstage-1:
 618          *   scan over all setups, rerunning only those setups
 619          *   which are not much slower than the fastest
 620          * else:
 621          *   use the next setup
 622          */
 623         do
 624         {
 625             pme_lb->cur++;
 626             if (pme_lb->cur == pme_lb->end)
 627             {
 628                 pme_lb->stage++;
 629                 pme_lb->cur = pme_lb->start;
 630             }
 631         }
 632         while (pme_lb->stage == pme_lb->nstage - 1 &&
 633                pme_lb->setup[pme_lb->cur].count > 0 &&
 634                pme_lb->setup[pme_lb->cur].cycles > cycles_fast*PME_LB_SLOW_FAC);
 635
 636         if (pme_lb->stage == pme_lb->nstage)
 637         {
 638             /* We are done optimizing, use the fastest setup we found */
 639             pme_lb->cur = pme_lb->fastest;
 640         }
 641     }
 642
 643     if (DOMAINDECOMP(cr) && pme_lb->stage > 0)
 644     {
 645         OK = change_dd_cutoff(cr, state, ir, pme_lb->setup[pme_lb->cur].rlistlong);
 646         if (!OK)
 647         {
 648             /* Failsafe solution */
 649             if (pme_lb->cur > 1 && pme_lb->stage == pme_lb->nstage)
 650             {
 651                 pme_lb->stage--;
 652             }
 653             pme_lb->fastest  = 0;
 654             pme_lb->start    = 0;
 655             pme_lb->end      = pme_lb->cur;
 656             pme_lb->cur      = pme_lb->start;
 657             pme_lb->elimited = epmelblimDD;
 658             print_loadbal_limited(fp_err, fp_log, step, pme_lb);
 659         }
 660     }
 661
 662     /* Change the Coulomb cut-off and the PME grid */
 663
 664     set = &pme_lb->setup[pme_lb->cur];
 665
 666     ic->rcoulomb     = set->rcut_coulomb;
 667     ic->rlist        = set->rlist;
 668     ic->rlistlong    = set->rlistlong;
 669     ir->nstcalclr    = set->nstcalclr;
 670     ic->ewaldcoeff_q = set->ewaldcoeff_q;
 671     /* TODO: centralize the code that sets the potentials shifts */
 672     if (ic->coulomb_modifier == eintmodPOTSHIFT)
 673     {
 674         ic->sh_ewald = gmx_erfc(ic->ewaldcoeff_q*ic->rcoulomb);
 675     }
 676     if (EVDW_PME(ic->vdwtype))
 677     {
 678         /* We have PME for both Coulomb and VdW, set rvdw equal to rcoulomb */
 679         ic->rvdw            = set->rcut_coulomb;
 680         ic->ewaldcoeff_lj   = set->ewaldcoeff_lj;
 681         if (ic->vdw_modifier == eintmodPOTSHIFT)
 682         {
 683             real crc2;
 684
 685             ic->dispersion_shift.cpot = -pow(ic->rvdw, -6.0);
 686             ic->repulsion_shift.cpot  = -pow(ic->rvdw, -12.0);
 687             ic->sh_invrc6             = -ic->dispersion_shift.cpot;
 688             crc2                      = sqr(ic->ewaldcoeff_lj*ic->rvdw);
 689             ic->sh_lj_ewald           = (exp(-crc2)*(1 + crc2 + 0.5*crc2*crc2) - 1)*pow(ic->rvdw, -6.0);
 690         }
 691     }
 692
 693     bUsesSimpleTables = uses_simple_tables(ir->cutoff_scheme, nbv, 0);
 694     if (pme_lb->cutoff_scheme == ecutsVERLET &&
 695         nbv->grp[0].kernel_type == nbnxnk8x8x8_CUDA)
 696     {
 697         nbnxn_cuda_pme_loadbal_update_param(nbv->cu_nbv, ic);
 698
 699         /* With tMPI + GPUs some ranks may be sharing GPU(s) and therefore
 700          * also sharing texture references. To keep the code simple, we don't
 701          * treat texture references as shared resources, but this means that
 702          * the coulomb_tab texture ref will get updated by multiple threads.
 703          * Hence, to ensure that the non-bonded kernels don't start before all
 704          * texture binding operations are finished, we need to wait for all ranks
 705          * to arrive here before continuing.
 706          *
 707          * Note that we could omit this barrier if GPUs are not shared (or
 708          * texture objects are used), but as this is initialization code, there
 709          * is not point in complicating things.
 710          */
 711 #ifdef GMX_THREAD_MPI
 712         if (PAR(cr))
 713         {
 714             gmx_barrier(cr);
 715         }
 716 #endif  /* GMX_THREAD_MPI */
 717     }
 718
 719     /* Usually we won't need the simple tables with GPUs.
 720      * But we do with hybrid acceleration and with free energy.
 721      * To avoid bugs, we always re-initialize the simple tables here.
 722      */
 723     init_interaction_const_tables(NULL, ic, bUsesSimpleTables, rtab);
 724
 725     if (cr->duty & DUTY_PME)
 726     {
 727         if (pme_lb->setup[pme_lb->cur].pmedata == NULL)
 728         {
 729             /* Generate a new PME data structure,
 730              * copying part of the old pointers.
 731              */
 732             gmx_pme_reinit(&set->pmedata,
 733                            cr, pme_lb->setup[0].pmedata, ir,
 734                            set->grid);
 735         }
 736         *pmedata = set->pmedata;
 737     }
 738     else
 739     {
 740         /* Tell our PME-only node to switch grid */
 741         gmx_pme_send_switchgrid(cr, set->grid, set->ewaldcoeff_q, set->ewaldcoeff_lj);
 742     }
 743
 744     if (debug)
 745     {
 746         print_grid(NULL, debug, "", "switched to", set, -1);
 747     }
 748
 749     if (pme_lb->stage == pme_lb->nstage)
 750     {
 751         print_grid(fp_err, fp_log, "", "optimal", set, -1);
 752     }
 753
 754     return TRUE;
 755 }
 756
 757 void restart_pme_loadbal(pme_load_balancing_t pme_lb, int n)
 758 {
 759     pme_lb->nstage += n;
 760 }
 761
 762 static int pme_grid_points(const pme_setup_t *setup)
 763 {
 764     return setup->grid[XX]*setup->grid[YY]*setup->grid[ZZ];
 765 }
 766
 767 static real pme_loadbal_rlist(const pme_setup_t *setup)
 768 {
 769     /* With the group cut-off scheme we can have twin-range either
 770      * for Coulomb or for VdW, so we need a check here.
 771      * With the Verlet cut-off scheme rlist=rlistlong.
 772      */
 773     if (setup->rcut_coulomb > setup->rlist)
 774     {
 775         return setup->rlistlong;
 776     }
 777     else
 778     {
 779         return setup->rlist;
 780     }
 781 }
 782
 783 static void print_pme_loadbal_setting(FILE              *fplog,
 784                                       char              *name,
 785                                       const pme_setup_t *setup)
 786 {
 787     fprintf(fplog,
 788             "   %-7s %6.3f nm %6.3f nm     %3d %3d %3d   %5.3f nm  %5.3f nm\n",
 789             name,
 790             setup->rcut_coulomb, pme_loadbal_rlist(setup),
 791             setup->grid[XX], setup->grid[YY], setup->grid[ZZ],
 792             setup->spacing, 1/setup->ewaldcoeff_q);
 793 }
 794
 795 static void print_pme_loadbal_settings(pme_load_balancing_t pme_lb,
 796                                        t_commrec           *cr,
 797                                        FILE                *fplog,
 798                                        gmx_bool             bNonBondedOnGPU)
 799 {
 800     double pp_ratio, grid_ratio;
 801
 802     pp_ratio   = pow(pme_loadbal_rlist(&pme_lb->setup[pme_lb->cur])/pme_loadbal_rlist(&pme_lb->setup[0]), 3.0);
 803     grid_ratio = pme_grid_points(&pme_lb->setup[pme_lb->cur])/
 804         (double)pme_grid_points(&pme_lb->setup[0]);
 805
 806     fprintf(fplog, "\n");
 807     fprintf(fplog, "       P P   -   P M E   L O A D   B A L A N C I N G\n");
 808     fprintf(fplog, "\n");
 809     /* Here we only warn when the optimal setting is the last one */
 810     if (pme_lb->elimited != epmelblimNO &&
 811         pme_lb->cur == pme_loadbal_end(pme_lb)-1)
 812     {
 813         fprintf(fplog, " NOTE: The PP/PME load balancing was limited by the %s,\n",
 814                 pmelblim_str[pme_lb->elimited]);
 815         fprintf(fplog, "       you might not have reached a good load balance.\n");
 816         if (pme_lb->elimited == epmelblimDD)
 817         {
 818             fprintf(fplog, "       Try different mdrun -dd settings or lower the -dds value.\n");
 819         }
 820         fprintf(fplog, "\n");
 821     }
 822     fprintf(fplog, " PP/PME load balancing changed the cut-off and PME settings:\n");
 823     fprintf(fplog, "           particle-particle                    PME\n");
 824     fprintf(fplog, "            rcoulomb  rlist            grid      spacing   1/beta\n");
 825     print_pme_loadbal_setting(fplog, "initial", &pme_lb->setup[0]);
 826     print_pme_loadbal_setting(fplog, "final", &pme_lb->setup[pme_lb->cur]);
 827     fprintf(fplog, " cost-ratio           %4.2f             %4.2f\n",
 828             pp_ratio, grid_ratio);
 829     fprintf(fplog, " (note that these numbers concern only part of the total PP and PME load)\n");
 830
 831     if (pp_ratio > 1.5 && !bNonBondedOnGPU)
 832     {
 833         md_print_warn(cr, fplog,
 834                       "NOTE: PME load balancing increased the non-bonded workload by more than 50%%.\n"
 835                       "      For better performance use (more) PME nodes (mdrun -npme),\n"
 836                       "      or in case you are beyond the scaling limit, use less nodes in total.\n");
 837     }
 838     else
 839     {
 840         fprintf(fplog, "\n");
 841     }
 842 }
 843
 844 void pme_loadbal_done(pme_load_balancing_t pme_lb,
 845                       t_commrec *cr, FILE *fplog,
 846                       gmx_bool bNonBondedOnGPU)
 847 {
 848     if (fplog != NULL && (pme_lb->cur > 0 || pme_lb->elimited != epmelblimNO))
 849     {
 850         print_pme_loadbal_settings(pme_lb, cr, fplog, bNonBondedOnGPU);
 851     }
 852
 853     /* TODO: Here we should free all pointers in pme_lb,
 854      * but as it contains pme data structures,
 855      * we need to first make pme.c free all data.
 856      */
 857 }