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 "domdec.h"
  44 #include "nbnxn_cuda_data_mgmt.h"
  45 #include "force.h"
  46 #include "macros.h"
  47 #include "md_logging.h"
  48 #include "pme_loadbal.h"
  49
  50 #include "gromacs/math/vec.h"
  51 #include "gromacs/pbcutil/pbc.h"
  52 #include "gromacs/utility/cstringutil.h"
  53 #include "gromacs/utility/smalloc.h"
  54
  55 /* Parameters and setting for one PP-PME setup */
  56 typedef struct {
  57     real      rcut_coulomb;    /* Coulomb cut-off                              */
  58     real      rlist;           /* pair-list cut-off                            */
  59     real      rlistlong;       /* LR pair-list cut-off                         */
  60     int       nstcalclr;       /* frequency of evaluating long-range forces for group scheme */
  61     real      spacing;         /* (largest) PME grid spacing                   */
  62     ivec      grid;            /* the PME grid dimensions                      */
  63     real      grid_efficiency; /* ineffiency factor for non-uniform grids <= 1 */
  64     real      ewaldcoeff_q;    /* Electrostatic Ewald coefficient            */
  65     real      ewaldcoeff_lj;   /* LJ Ewald coefficient, only for the call to send_switchgrid */
  66     gmx_pme_t pmedata;         /* the data structure used in the PME code      */
  67     int       count;           /* number of times this setup has been timed    */
  68     double    cycles;          /* the fastest time for this setup in cycles    */
  69 } pme_setup_t;
  70
  71 /* In the initial scan, step by grids that are at least a factor 0.8 coarser */
  72 #define PME_LB_GRID_SCALE_FAC  0.8
  73 /* In the initial scan, try to skip grids with uneven x/y/z spacing,
  74  * checking if the "efficiency" is more than 5% worse than the previous grid.
  75  */
  76 #define PME_LB_GRID_EFFICIENCY_REL_FAC  1.05
  77 /* Rerun up till 12% slower setups than the fastest up till now */
  78 #define PME_LB_SLOW_FAC  1.12
  79 /* If setups get more than 2% faster, do another round to avoid
  80  * choosing a slower setup due to acceleration or fluctuations.
  81  */
  82 #define PME_LB_ACCEL_TOL 1.02
  83
  84 enum {
  85     epmelblimNO, epmelblimBOX, epmelblimDD, epmelblimPMEGRID, epmelblimNR
  86 };
  87
  88 const char *pmelblim_str[epmelblimNR] =
  89 { "no", "box size", "domain decompostion", "PME grid restriction" };
  90
  91 struct pme_load_balancing {
  92     int          nstage;             /* the current maximum number of stages */
  93
  94     real         cut_spacing;        /* the minimum cutoff / PME grid spacing ratio */
  95     real         rcut_vdw;           /* Vdw cutoff (does not change) */
  96     real         rcut_coulomb_start; /* Initial electrostatics cutoff */
  97     int          nstcalclr_start;    /* Initial electrostatics cutoff */
  98     real         rbuf_coulomb;       /* the pairlist buffer size */
  99     real         rbuf_vdw;           /* the pairlist buffer size */
 100     matrix       box_start;          /* the initial simulation box */
 101     int          n;                  /* the count of setup as well as the allocation size */
 102     pme_setup_t *setup;              /* the PME+cutoff setups */
 103     int          cur;                /* the current setup */
 104     int          fastest;            /* fastest setup up till now */
 105     int          start;              /* start of setup range to consider in stage>0 */
 106     int          end;                /* end   of setup range to consider in stage>0 */
 107     int          elimited;           /* was the balancing limited, uses enum above */
 108     int          cutoff_scheme;      /* Verlet or group cut-offs */
 109
 110     int          stage;              /* the current stage */
 111 };
 112
 113 void pme_loadbal_init(pme_load_balancing_t *pme_lb_p,
 114                       const t_inputrec *ir, matrix box,
 115                       const interaction_const_t *ic,
 116                       gmx_pme_t pmedata)
 117 {
 118     pme_load_balancing_t pme_lb;
 119     real                 spm, sp;
 120     int                  d;
 121
 122     snew(pme_lb, 1);
 123
 124     /* Any number of stages >= 2 is supported */
 125     pme_lb->nstage   = 2;
 126
 127     pme_lb->cutoff_scheme = ir->cutoff_scheme;
 128
 129     if (pme_lb->cutoff_scheme == ecutsVERLET)
 130     {
 131         pme_lb->rbuf_coulomb = ic->rlist - ic->rcoulomb;
 132         pme_lb->rbuf_vdw     = pme_lb->rbuf_coulomb;
 133     }
 134     else
 135     {
 136         if (ic->rcoulomb > ic->rlist)
 137         {
 138             pme_lb->rbuf_coulomb = ic->rlistlong - ic->rcoulomb;
 139         }
 140         else
 141         {
 142             pme_lb->rbuf_coulomb = ic->rlist - ic->rcoulomb;
 143         }
 144         if (ic->rvdw > ic->rlist)
 145         {
 146             pme_lb->rbuf_vdw = ic->rlistlong - ic->rvdw;
 147         }
 148         else
 149         {
 150             pme_lb->rbuf_vdw = ic->rlist - ic->rvdw;
 151         }
 152     }
 153
 154     copy_mat(box, pme_lb->box_start);
 155     if (ir->ePBC == epbcXY && ir->nwall == 2)
 156     {
 157         svmul(ir->wall_ewald_zfac, pme_lb->box_start[ZZ], pme_lb->box_start[ZZ]);
 158     }
 159
 160     pme_lb->n = 1;
 161     snew(pme_lb->setup, pme_lb->n);
 162
 163     pme_lb->rcut_vdw                 = ic->rvdw;
 164     pme_lb->rcut_coulomb_start       = ir->rcoulomb;
 165     pme_lb->nstcalclr_start          = ir->nstcalclr;
 166
 167     pme_lb->cur                      = 0;
 168     pme_lb->setup[0].rcut_coulomb    = ic->rcoulomb;
 169     pme_lb->setup[0].rlist           = ic->rlist;
 170     pme_lb->setup[0].rlistlong       = ic->rlistlong;
 171     pme_lb->setup[0].nstcalclr       = ir->nstcalclr;
 172     pme_lb->setup[0].grid[XX]        = ir->nkx;
 173     pme_lb->setup[0].grid[YY]        = ir->nky;
 174     pme_lb->setup[0].grid[ZZ]        = ir->nkz;
 175     pme_lb->setup[0].ewaldcoeff_q    = ic->ewaldcoeff_q;
 176     pme_lb->setup[0].ewaldcoeff_lj   = ic->ewaldcoeff_lj;
 177
 178     pme_lb->setup[0].pmedata  = pmedata;
 179
 180     spm = 0;
 181     for (d = 0; d < DIM; d++)
 182     {
 183         sp = norm(pme_lb->box_start[d])/pme_lb->setup[0].grid[d];
 184         if (sp > spm)
 185         {
 186             spm = sp;
 187         }
 188     }
 189     pme_lb->setup[0].spacing = spm;
 190
 191     if (ir->fourier_spacing > 0)
 192     {
 193         pme_lb->cut_spacing = ir->rcoulomb/ir->fourier_spacing;
 194     }
 195     else
 196     {
 197         pme_lb->cut_spacing = ir->rcoulomb/pme_lb->setup[0].spacing;
 198     }
 199
 200     pme_lb->stage = 0;
 201
 202     pme_lb->fastest  = 0;
 203     pme_lb->start    = 0;
 204     pme_lb->end      = 0;
 205     pme_lb->elimited = epmelblimNO;
 206
 207     *pme_lb_p = pme_lb;
 208 }
 209
 210 static gmx_bool pme_loadbal_increase_cutoff(pme_load_balancing_t  pme_lb,
 211                                             int                   pme_order,
 212                                             const gmx_domdec_t   *dd)
 213 {
 214     pme_setup_t *set;
 215     int          npmenodes_x, npmenodes_y;
 216     real         fac, sp;
 217     real         tmpr_coulomb, tmpr_vdw;
 218     int          d;
 219     gmx_bool     grid_ok;
 220
 221     /* Try to add a new setup with next larger cut-off to the list */
 222     pme_lb->n++;
 223     srenew(pme_lb->setup, pme_lb->n);
 224     set          = &pme_lb->setup[pme_lb->n-1];
 225     set->pmedata = NULL;
 226
 227     get_pme_nnodes(dd, &npmenodes_x, &npmenodes_y);
 228
 229     fac = 1;
 230     do
 231     {
 232         /* Avoid infinite while loop, which can occur at the minimum grid size.
 233          * Note that in practice load balancing will stop before this point.
 234          * The factor 2.1 allows for the extreme case in which only grids
 235          * of powers of 2 are allowed (the current code supports more grids).
 236          */
 237         if (fac > 2.1)
 238         {
 239             pme_lb->n--;
 240
 241             return FALSE;
 242         }
 243
 244         fac *= 1.01;
 245         clear_ivec(set->grid);
 246         sp = calc_grid(NULL, pme_lb->box_start,
 247                        fac*pme_lb->setup[pme_lb->cur].spacing,
 248                        &set->grid[XX],
 249                        &set->grid[YY],
 250                        &set->grid[ZZ]);
 251
 252         /* As here we can't easily check if one of the PME nodes
 253          * uses threading, we do a conservative grid check.
 254          * This means we can't use pme_order or less grid lines
 255          * per PME node along x, which is not a strong restriction.
 256          */
 257         gmx_pme_check_restrictions(pme_order,
 258                                    set->grid[XX], set->grid[YY], set->grid[ZZ],
 259                                    npmenodes_x, npmenodes_y,
 260                                    TRUE,
 261                                    FALSE,
 262                                    &grid_ok);
 263     }
 264     while (sp <= 1.001*pme_lb->setup[pme_lb->cur].spacing || !grid_ok);
 265
 266     set->rcut_coulomb = pme_lb->cut_spacing*sp;
 267
 268     if (pme_lb->cutoff_scheme == ecutsVERLET)
 269     {
 270         set->rlist        = set->rcut_coulomb + pme_lb->rbuf_coulomb;
 271         /* We dont use LR lists with Verlet, but this avoids if-statements in further checks */
 272         set->rlistlong    = set->rlist;
 273     }
 274     else
 275     {
 276         tmpr_coulomb          = set->rcut_coulomb + pme_lb->rbuf_coulomb;
 277         tmpr_vdw              = pme_lb->rcut_vdw + pme_lb->rbuf_vdw;
 278         set->rlist            = min(tmpr_coulomb, tmpr_vdw);
 279         set->rlistlong        = max(tmpr_coulomb, tmpr_vdw);
 280
 281         /* Set the long-range update frequency */
 282         if (set->rlist == set->rlistlong)
 283         {
 284             /* No long-range interactions if the short-/long-range cutoffs are identical */
 285             set->nstcalclr = 0;
 286         }
 287         else if (pme_lb->nstcalclr_start == 0 || pme_lb->nstcalclr_start == 1)
 288         {
 289             /* We were not doing long-range before, but now we are since rlist!=rlistlong */
 290             set->nstcalclr = 1;
 291         }
 292         else
 293         {
 294             /* We were already doing long-range interactions from the start */
 295             if (pme_lb->rcut_vdw > pme_lb->rcut_coulomb_start)
 296             {
 297                 /* We were originally doing long-range VdW-only interactions.
 298                  * If rvdw is still longer than rcoulomb we keep the original nstcalclr,
 299                  * but if the coulomb cutoff has become longer we should update the long-range
 300                  * part every step.
 301                  */
 302                 set->nstcalclr = (tmpr_vdw > tmpr_coulomb) ? pme_lb->nstcalclr_start : 1;
 303             }
 304             else
 305             {
 306                 /* We were not doing any long-range interaction from the start,
 307                  * since it is not possible to do twin-range coulomb for the PME interaction.
 308                  */
 309                 set->nstcalclr = 1;
 310             }
 311         }
 312     }
 313
 314     set->spacing      = sp;
 315     /* The grid efficiency is the size wrt a grid with uniform x/y/z spacing */
 316     set->grid_efficiency = 1;
 317     for (d = 0; d < DIM; d++)
 318     {
 319         set->grid_efficiency *= (set->grid[d]*sp)/norm(pme_lb->box_start[d]);
 320     }
 321     /* The Ewald coefficient is inversly proportional to the cut-off */
 322     set->ewaldcoeff_q =
 323         pme_lb->setup[0].ewaldcoeff_q*pme_lb->setup[0].rcut_coulomb/set->rcut_coulomb;
 324     /* We set ewaldcoeff_lj in set, even when LJ-PME is not used */
 325     set->ewaldcoeff_lj =
 326         pme_lb->setup[0].ewaldcoeff_lj*pme_lb->setup[0].rcut_coulomb/set->rcut_coulomb;
 327
 328     set->count   = 0;
 329     set->cycles  = 0;
 330
 331     if (debug)
 332     {
 333         fprintf(debug, "PME loadbal: grid %d %d %d, coulomb cutoff %f\n",
 334                 set->grid[XX], set->grid[YY], set->grid[ZZ], set->rcut_coulomb);
 335     }
 336     return TRUE;
 337 }
 338
 339 static void print_grid(FILE *fp_err, FILE *fp_log,
 340                        const char *pre,
 341                        const char *desc,
 342                        const pme_setup_t *set,
 343                        double cycles)
 344 {
 345     char buf[STRLEN], buft[STRLEN];
 346
 347     if (cycles >= 0)
 348     {
 349         sprintf(buft, ": %.1f M-cycles", cycles*1e-6);
 350     }
 351     else
 352     {
 353         buft[0] = '\0';
 354     }
 355     sprintf(buf, "%-11s%10s pme grid %d %d %d, coulomb cutoff %.3f%s",
 356             pre,
 357             desc, set->grid[XX], set->grid[YY], set->grid[ZZ], set->rcut_coulomb,
 358             buft);
 359     if (fp_err != NULL)
 360     {
 361         fprintf(fp_err, "\r%s\n", buf);
 362     }
 363     if (fp_log != NULL)
 364     {
 365         fprintf(fp_log, "%s\n", buf);
 366     }
 367 }
 368
 369 static int pme_loadbal_end(pme_load_balancing_t pme_lb)
 370 {
 371     /* In the initial stage only n is set; end is not set yet */
 372     if (pme_lb->end > 0)
 373     {
 374         return pme_lb->end;
 375     }
 376     else
 377     {
 378         return pme_lb->n;
 379     }
 380 }
 381
 382 static void print_loadbal_limited(FILE *fp_err, FILE *fp_log,
 383                                   gmx_int64_t step,
 384                                   pme_load_balancing_t pme_lb)
 385 {
 386     char buf[STRLEN], sbuf[22];
 387
 388     sprintf(buf, "step %4s: the %s limits the PME load balancing to a coulomb cut-off of %.3f",
 389             gmx_step_str(step, sbuf),
 390             pmelblim_str[pme_lb->elimited],
 391             pme_lb->setup[pme_loadbal_end(pme_lb)-1].rcut_coulomb);
 392     if (fp_err != NULL)
 393     {
 394         fprintf(fp_err, "\r%s\n", buf);
 395     }
 396     if (fp_log != NULL)
 397     {
 398         fprintf(fp_log, "%s\n", buf);
 399     }
 400 }
 401
 402 static void switch_to_stage1(pme_load_balancing_t pme_lb)
 403 {
 404     pme_lb->start = 0;
 405     while (pme_lb->start+1 < pme_lb->n &&
 406            (pme_lb->setup[pme_lb->start].count == 0 ||
 407             pme_lb->setup[pme_lb->start].cycles >
 408             pme_lb->setup[pme_lb->fastest].cycles*PME_LB_SLOW_FAC))
 409     {
 410         pme_lb->start++;
 411     }
 412     while (pme_lb->start > 0 && pme_lb->setup[pme_lb->start-1].cycles == 0)
 413     {
 414         pme_lb->start--;
 415     }
 416
 417     pme_lb->end = pme_lb->n;
 418     if (pme_lb->setup[pme_lb->end-1].count > 0 &&
 419         pme_lb->setup[pme_lb->end-1].cycles >
 420         pme_lb->setup[pme_lb->fastest].cycles*PME_LB_SLOW_FAC)
 421     {
 422         pme_lb->end--;
 423     }
 424
 425     pme_lb->stage = 1;
 426
 427     /* Next we want to choose setup pme_lb->start, but as we will increase
 428      * pme_ln->cur by one right after returning, we subtract 1 here.
 429      */
 430     pme_lb->cur = pme_lb->start - 1;
 431 }
 432
 433 gmx_bool pme_load_balance(pme_load_balancing_t pme_lb,
 434                           t_commrec           *cr,
 435                           FILE                *fp_err,
 436                           FILE                *fp_log,
 437                           t_inputrec          *ir,
 438                           t_state             *state,
 439                           double               cycles,
 440                           interaction_const_t *ic,
 441                           nonbonded_verlet_t  *nbv,
 442                           gmx_pme_t           *pmedata,
 443                           gmx_int64_t          step)
 444 {
 445     gmx_bool     OK;
 446     pme_setup_t *set;
 447     double       cycles_fast;
 448     char         buf[STRLEN], sbuf[22];
 449     real         rtab;
 450     gmx_bool     bUsesSimpleTables = TRUE;
 451
 452     if (pme_lb->stage == pme_lb->nstage)
 453     {
 454         return FALSE;
 455     }
 456
 457     if (PAR(cr))
 458     {
 459         gmx_sumd(1, &cycles, cr);
 460         cycles /= cr->nnodes;
 461     }
 462
 463     set = &pme_lb->setup[pme_lb->cur];
 464     set->count++;
 465
 466     rtab = ir->rlistlong + ir->tabext;
 467
 468     if (set->count % 2 == 1)
 469     {
 470         /* Skip the first cycle, because the first step after a switch
 471          * is much slower due to allocation and/or caching effects.
 472          */
 473         return TRUE;
 474     }
 475
 476     sprintf(buf, "step %4s: ", gmx_step_str(step, sbuf));
 477     print_grid(fp_err, fp_log, buf, "timed with", set, cycles);
 478
 479     if (set->count <= 2)
 480     {
 481         set->cycles = cycles;
 482     }
 483     else
 484     {
 485         if (cycles*PME_LB_ACCEL_TOL < set->cycles &&
 486             pme_lb->stage == pme_lb->nstage - 1)
 487         {
 488             /* The performance went up a lot (due to e.g. DD load balancing).
 489              * Add a stage, keep the minima, but rescan all setups.
 490              */
 491             pme_lb->nstage++;
 492
 493             if (debug)
 494             {
 495                 fprintf(debug, "The performance for grid %d %d %d went from %.3f to %.1f M-cycles, this is more than %f\n"
 496                         "Increased the number stages to %d"
 497                         " and ignoring the previous performance\n",
 498                         set->grid[XX], set->grid[YY], set->grid[ZZ],
 499                         cycles*1e-6, set->cycles*1e-6, PME_LB_ACCEL_TOL,
 500                         pme_lb->nstage);
 501             }
 502         }
 503         set->cycles = min(set->cycles, cycles);
 504     }
 505
 506     if (set->cycles < pme_lb->setup[pme_lb->fastest].cycles)
 507     {
 508         pme_lb->fastest = pme_lb->cur;
 509
 510         if (DOMAINDECOMP(cr))
 511         {
 512             /* We found a new fastest setting, ensure that with subsequent
 513              * shorter cut-off's the dynamic load balancing does not make
 514              * the use of the current cut-off impossible. This solution is
 515              * a trade-off, as the PME load balancing and DD domain size
 516              * load balancing can interact in complex ways.
 517              * With the Verlet kernels, DD load imbalance will usually be
 518              * mainly due to bonded interaction imbalance, which will often
 519              * quickly push the domain boundaries beyond the limit for the
 520              * optimal, PME load balanced, cut-off. But it could be that
 521              * better overal performance can be obtained with a slightly
 522              * shorter cut-off and better DD load balancing.
 523              */
 524             change_dd_dlb_cutoff_limit(cr);
 525         }
 526     }
 527     cycles_fast = pme_lb->setup[pme_lb->fastest].cycles;
 528
 529     /* Check in stage 0 if we should stop scanning grids.
 530      * Stop when the time is more than SLOW_FAC longer than the fastest.
 531      */
 532     if (pme_lb->stage == 0 && pme_lb->cur > 0 &&
 533         cycles > pme_lb->setup[pme_lb->fastest].cycles*PME_LB_SLOW_FAC)
 534     {
 535         pme_lb->n = pme_lb->cur + 1;
 536         /* Done with scanning, go to stage 1 */
 537         switch_to_stage1(pme_lb);
 538     }
 539
 540     if (pme_lb->stage == 0)
 541     {
 542         int gridsize_start;
 543
 544         gridsize_start = set->grid[XX]*set->grid[YY]*set->grid[ZZ];
 545
 546         do
 547         {
 548             if (pme_lb->cur+1 < pme_lb->n)
 549             {
 550                 /* We had already generated the next setup */
 551                 OK = TRUE;
 552             }
 553             else
 554             {
 555                 /* Find the next setup */
 556                 OK = pme_loadbal_increase_cutoff(pme_lb, ir->pme_order, cr->dd);
 557
 558                 if (!OK)
 559                 {
 560                     pme_lb->elimited = epmelblimPMEGRID;
 561                 }
 562             }
 563
 564             if (OK && ir->ePBC != epbcNONE)
 565             {
 566                 OK = (sqr(pme_lb->setup[pme_lb->cur+1].rlistlong)
 567                       <= max_cutoff2(ir->ePBC, state->box));
 568                 if (!OK)
 569                 {
 570                     pme_lb->elimited = epmelblimBOX;
 571                 }
 572             }
 573
 574             if (OK)
 575             {
 576                 pme_lb->cur++;
 577
 578                 if (DOMAINDECOMP(cr))
 579                 {
 580                     OK = change_dd_cutoff(cr, state, ir,
 581                                           pme_lb->setup[pme_lb->cur].rlistlong);
 582                     if (!OK)
 583                     {
 584                         /* Failed: do not use this setup */
 585                         pme_lb->cur--;
 586                         pme_lb->elimited = epmelblimDD;
 587                     }
 588                 }
 589             }
 590             if (!OK)
 591             {
 592                 /* We hit the upper limit for the cut-off,
 593                  * the setup should not go further than cur.
 594                  */
 595                 pme_lb->n = pme_lb->cur + 1;
 596                 print_loadbal_limited(fp_err, fp_log, step, pme_lb);
 597                 /* Switch to the next stage */
 598                 switch_to_stage1(pme_lb);
 599             }
 600         }
 601         while (OK &&
 602                !(pme_lb->setup[pme_lb->cur].grid[XX]*
 603                  pme_lb->setup[pme_lb->cur].grid[YY]*
 604                  pme_lb->setup[pme_lb->cur].grid[ZZ] <
 605                  gridsize_start*PME_LB_GRID_SCALE_FAC
 606                  &&
 607                  pme_lb->setup[pme_lb->cur].grid_efficiency <
 608                  pme_lb->setup[pme_lb->cur-1].grid_efficiency*PME_LB_GRID_EFFICIENCY_REL_FAC));
 609     }
 610
 611     if (pme_lb->stage > 0 && pme_lb->end == 1)
 612     {
 613         pme_lb->cur   = 0;
 614         pme_lb->stage = pme_lb->nstage;
 615     }
 616     else if (pme_lb->stage > 0 && pme_lb->end > 1)
 617     {
 618         /* If stage = nstage-1:
 619          *   scan over all setups, rerunning only those setups
 620          *   which are not much slower than the fastest
 621          * else:
 622          *   use the next setup
 623          */
 624         do
 625         {
 626             pme_lb->cur++;
 627             if (pme_lb->cur == pme_lb->end)
 628             {
 629                 pme_lb->stage++;
 630                 pme_lb->cur = pme_lb->start;
 631             }
 632         }
 633         while (pme_lb->stage == pme_lb->nstage - 1 &&
 634                pme_lb->setup[pme_lb->cur].count > 0 &&
 635                pme_lb->setup[pme_lb->cur].cycles > cycles_fast*PME_LB_SLOW_FAC);
 636
 637         if (pme_lb->stage == pme_lb->nstage)
 638         {
 639             /* We are done optimizing, use the fastest setup we found */
 640             pme_lb->cur = pme_lb->fastest;
 641         }
 642     }
 643
 644     if (DOMAINDECOMP(cr) && pme_lb->stage > 0)
 645     {
 646         OK = change_dd_cutoff(cr, state, ir, pme_lb->setup[pme_lb->cur].rlistlong);
 647         if (!OK)
 648         {
 649             /* Failsafe solution */
 650             if (pme_lb->cur > 1 && pme_lb->stage == pme_lb->nstage)
 651             {
 652                 pme_lb->stage--;
 653             }
 654             pme_lb->fastest  = 0;
 655             pme_lb->start    = 0;
 656             pme_lb->end      = pme_lb->cur;
 657             pme_lb->cur      = pme_lb->start;
 658             pme_lb->elimited = epmelblimDD;
 659             print_loadbal_limited(fp_err, fp_log, step, pme_lb);
 660         }
 661     }
 662
 663     /* Change the Coulomb cut-off and the PME grid */
 664
 665     set = &pme_lb->setup[pme_lb->cur];
 666
 667     ic->rcoulomb     = set->rcut_coulomb;
 668     ic->rlist        = set->rlist;
 669     ic->rlistlong    = set->rlistlong;
 670     ir->nstcalclr    = set->nstcalclr;
 671     ic->ewaldcoeff_q = set->ewaldcoeff_q;
 672     /* TODO: centralize the code that sets the potentials shifts */
 673     if (ic->coulomb_modifier == eintmodPOTSHIFT)
 674     {
 675         ic->sh_ewald = gmx_erfc(ic->ewaldcoeff_q*ic->rcoulomb);
 676     }
 677     if (EVDW_PME(ic->vdwtype))
 678     {
 679         /* We have PME for both Coulomb and VdW, set rvdw equal to rcoulomb */
 680         ic->rvdw            = set->rcut_coulomb;
 681         ic->ewaldcoeff_lj   = set->ewaldcoeff_lj;
 682         if (ic->vdw_modifier == eintmodPOTSHIFT)
 683         {
 684             real crc2;
 685
 686             ic->dispersion_shift.cpot = -pow(ic->rvdw, -6.0);
 687             ic->repulsion_shift.cpot  = -pow(ic->rvdw, -12.0);
 688             ic->sh_invrc6             = -ic->dispersion_shift.cpot;
 689             crc2                      = sqr(ic->ewaldcoeff_lj*ic->rvdw);
 690             ic->sh_lj_ewald           = (exp(-crc2)*(1 + crc2 + 0.5*crc2*crc2) - 1)*pow(ic->rvdw, -6.0);
 691         }
 692     }
 693
 694     bUsesSimpleTables = uses_simple_tables(ir->cutoff_scheme, nbv, 0);
 695     if (pme_lb->cutoff_scheme == ecutsVERLET &&
 696         nbv->grp[0].kernel_type == nbnxnk8x8x8_CUDA)
 697     {
 698         nbnxn_cuda_pme_loadbal_update_param(nbv->cu_nbv, ic);
 699
 700         /* With tMPI + GPUs some ranks may be sharing GPU(s) and therefore
 701          * also sharing texture references. To keep the code simple, we don't
 702          * treat texture references as shared resources, but this means that
 703          * the coulomb_tab texture ref will get updated by multiple threads.
 704          * Hence, to ensure that the non-bonded kernels don't start before all
 705          * texture binding operations are finished, we need to wait for all ranks
 706          * to arrive here before continuing.
 707          *
 708          * Note that we could omit this barrier if GPUs are not shared (or
 709          * texture objects are used), but as this is initialization code, there
 710          * is not point in complicating things.
 711          */
 712 #ifdef GMX_THREAD_MPI
 713         if (PAR(cr))
 714         {
 715             gmx_barrier(cr);
 716         }
 717 #endif  /* GMX_THREAD_MPI */
 718     }
 719
 720     /* Usually we won't need the simple tables with GPUs.
 721      * But we do with hybrid acceleration and with free energy.
 722      * To avoid bugs, we always re-initialize the simple tables here.
 723      */
 724     init_interaction_const_tables(NULL, ic, bUsesSimpleTables, rtab);
 725
 726     if (cr->duty & DUTY_PME)
 727     {
 728         if (pme_lb->setup[pme_lb->cur].pmedata == NULL)
 729         {
 730             /* Generate a new PME data structure,
 731              * copying part of the old pointers.
 732              */
 733             gmx_pme_reinit(&set->pmedata,
 734                            cr, pme_lb->setup[0].pmedata, ir,
 735                            set->grid);
 736         }
 737         *pmedata = set->pmedata;
 738     }
 739     else
 740     {
 741         /* Tell our PME-only node to switch grid */
 742         gmx_pme_send_switchgrid(cr, set->grid, set->ewaldcoeff_q, set->ewaldcoeff_lj);
 743     }
 744
 745     if (debug)
 746     {
 747         print_grid(NULL, debug, "", "switched to", set, -1);
 748     }
 749
 750     if (pme_lb->stage == pme_lb->nstage)
 751     {
 752         print_grid(fp_err, fp_log, "", "optimal", set, -1);
 753     }
 754
 755     return TRUE;
 756 }
 757
 758 void restart_pme_loadbal(pme_load_balancing_t pme_lb, int n)
 759 {
 760     pme_lb->nstage += n;
 761 }
 762
 763 static int pme_grid_points(const pme_setup_t *setup)
 764 {
 765     return setup->grid[XX]*setup->grid[YY]*setup->grid[ZZ];
 766 }
 767
 768 static real pme_loadbal_rlist(const pme_setup_t *setup)
 769 {
 770     /* With the group cut-off scheme we can have twin-range either
 771      * for Coulomb or for VdW, so we need a check here.
 772      * With the Verlet cut-off scheme rlist=rlistlong.
 773      */
 774     if (setup->rcut_coulomb > setup->rlist)
 775     {
 776         return setup->rlistlong;
 777     }
 778     else
 779     {
 780         return setup->rlist;
 781     }
 782 }
 783
 784 static void print_pme_loadbal_setting(FILE              *fplog,
 785                                       char              *name,
 786                                       const pme_setup_t *setup)
 787 {
 788     fprintf(fplog,
 789             "   %-7s %6.3f nm %6.3f nm     %3d %3d %3d   %5.3f nm  %5.3f nm\n",
 790             name,
 791             setup->rcut_coulomb, pme_loadbal_rlist(setup),
 792             setup->grid[XX], setup->grid[YY], setup->grid[ZZ],
 793             setup->spacing, 1/setup->ewaldcoeff_q);
 794 }
 795
 796 static void print_pme_loadbal_settings(pme_load_balancing_t pme_lb,
 797                                        t_commrec           *cr,
 798                                        FILE                *fplog,
 799                                        gmx_bool             bNonBondedOnGPU)
 800 {
 801     double pp_ratio, grid_ratio;
 802
 803     pp_ratio   = pow(pme_loadbal_rlist(&pme_lb->setup[pme_lb->cur])/pme_loadbal_rlist(&pme_lb->setup[0]), 3.0);
 804     grid_ratio = pme_grid_points(&pme_lb->setup[pme_lb->cur])/
 805         (double)pme_grid_points(&pme_lb->setup[0]);
 806
 807     fprintf(fplog, "\n");
 808     fprintf(fplog, "       P P   -   P M E   L O A D   B A L A N C I N G\n");
 809     fprintf(fplog, "\n");
 810     /* Here we only warn when the optimal setting is the last one */
 811     if (pme_lb->elimited != epmelblimNO &&
 812         pme_lb->cur == pme_loadbal_end(pme_lb)-1)
 813     {
 814         fprintf(fplog, " NOTE: The PP/PME load balancing was limited by the %s,\n",
 815                 pmelblim_str[pme_lb->elimited]);
 816         fprintf(fplog, "       you might not have reached a good load balance.\n");
 817         if (pme_lb->elimited == epmelblimDD)
 818         {
 819             fprintf(fplog, "       Try different mdrun -dd settings or lower the -dds value.\n");
 820         }
 821         fprintf(fplog, "\n");
 822     }
 823     fprintf(fplog, " PP/PME load balancing changed the cut-off and PME settings:\n");
 824     fprintf(fplog, "           particle-particle                    PME\n");
 825     fprintf(fplog, "            rcoulomb  rlist            grid      spacing   1/beta\n");
 826     print_pme_loadbal_setting(fplog, "initial", &pme_lb->setup[0]);
 827     print_pme_loadbal_setting(fplog, "final", &pme_lb->setup[pme_lb->cur]);
 828     fprintf(fplog, " cost-ratio           %4.2f             %4.2f\n",
 829             pp_ratio, grid_ratio);
 830     fprintf(fplog, " (note that these numbers concern only part of the total PP and PME load)\n");
 831
 832     if (pp_ratio > 1.5 && !bNonBondedOnGPU)
 833     {
 834         md_print_warn(cr, fplog,
 835                       "NOTE: PME load balancing increased the non-bonded workload by more than 50%%.\n"
 836                       "      For better performance, use (more) PME ranks (mdrun -npme),\n"
 837                       "      or if you are beyond the scaling limit, use fewer total ranks (or nodes).\n");
 838     }
 839     else
 840     {
 841         fprintf(fplog, "\n");
 842     }
 843 }
 844
 845 void pme_loadbal_done(pme_load_balancing_t pme_lb,
 846                       t_commrec *cr, FILE *fplog,
 847                       gmx_bool bNonBondedOnGPU)
 848 {
 849     if (fplog != NULL && (pme_lb->cur > 0 || pme_lb->elimited != epmelblimNO))
 850     {
 851         print_pme_loadbal_settings(pme_lb, cr, fplog, bNonBondedOnGPU);
 852     }
 853
 854     /* TODO: Here we should free all pointers in pme_lb,
 855      * but as it contains pme data structures,
 856      * we need to first make pme.c free all data.
 857      */
 858 }