src/mdlib/domdec_setup.c

   1 /* -*- mode: c; tab-width: 4; indent-tabs-mode: nil; c-basic-offset: 4; c-file-style: "stroustrup"; -*-
   2  *
   3  *
   4  * This file is part of Gromacs        Copyright (c) 1991-2008
   5  * David van der Spoel, Erik Lindahl, Berk Hess, University of Groningen.
   6  *
   7  * This program is free software; you can redistribute it and/or
   8  * modify it under the terms of the GNU General Public License
   9  * as published by the Free Software Foundation; either version 2
  10  * of the License, or (at your option) any later version.
  11  *
  12  * To help us fund GROMACS development, we humbly ask that you cite
  13  * the research papers on the package. Check out http://www.gromacs.org
  14  *
  15  * And Hey:
  16  * Gnomes, ROck Monsters And Chili Sauce
  17  */
  18
  19 #ifdef HAVE_CONFIG_H
  20 #include <config.h>
  21 #endif
  22
  23 #include <stdio.h>
  24 #include "domdec.h"
  25 #include "network.h"
  26 #include "perf_est.h"
  27 #include "physics.h"
  28 #include "smalloc.h"
  29 #include "typedefs.h"
  30 #include "vec.h"
  31 #include "names.h"
  32
  33 /* Margin for setting up the DD grid */
  34 #define DD_GRID_MARGIN_PRES_SCALE 1.05
  35
  36 static int factorize(int n,int **fac,int **mfac)
  37 {
  38         int d,ndiv;
  39
  40         /* Decompose n in factors */
  41         snew(*fac,n/2);
  42         snew(*mfac,n/2);
  43         d = 2;
  44         ndiv = 0;
  45         while (n > 1)
  46                 {
  47                 while (n % d == 0)
  48                         {
  49                         if (ndiv == 0 || (*fac)[ndiv-1] != d)
  50                                 {
  51                                 ndiv++;
  52                                 (*fac)[ndiv-1] = d;
  53                                 }
  54                         (*mfac)[ndiv-1]++;
  55                         n /= d;
  56                         }
  57                 d++;
  58                 }
  59
  60         return ndiv;
  61 }
  62
  63 static bool fits_pme_ratio(int nnodes,int npme,float ratio)
  64 {
  65     return ((double)npme/(double)nnodes > 0.95*ratio);
  66 }
  67
  68 static bool fits_pp_pme_perf(FILE *fplog,
  69                              t_inputrec *ir,matrix box,gmx_mtop_t *mtop,
  70                              int nnodes,int npme,float ratio)
  71 {
  72     int ndiv,*div,*mdiv,ldiv;
  73
  74     ndiv = factorize(nnodes-npme,&div,&mdiv);
  75     ldiv = div[ndiv-1];
  76     sfree(div);
  77     sfree(mdiv);
  78     /* The check below gives a reasonable division:
  79      * factor 5 allowed at 5 or more PP nodes,
  80      * factor 7 allowed at 49 or more PP nodes.
  81      */
  82     if (ldiv > 3 + (int)(pow(nnodes-npme,1.0/3.0) + 0.5))
  83     {
  84         return FALSE;
  85     }
  86
  87     /* Does this division gives a reasonable PME load? */
  88     return fits_pme_ratio(nnodes,npme,ratio);
  89 }
  90
  91 static int guess_npme(FILE *fplog,gmx_mtop_t *mtop,t_inputrec *ir,matrix box,
  92                                           int nnodes)
  93 {
  94         float ratio;
  95         int  npme,nkx,nky;
  96         t_inputrec ir_try;
  97
  98         ratio = pme_load_estimate(mtop,ir,box);
  99
 100         if (fplog)
 101     {
 102                 fprintf(fplog,"Guess for relative PME load: %.2f\n",ratio);
 103     }
 104
 105         /* We assume the optimal node ratio is close to the load ratio.
 106          * The communication load is neglected,
 107          * but (hopefully) this will balance out between PP and PME.
 108          */
 109
 110     if (!fits_pme_ratio(nnodes,nnodes/2,ratio))
 111     {
 112         /* We would need more than nnodes/2 PME only nodes,
 113          * which is not possible. Since the PME load is very high,
 114          * we will not loose much performance when all nodes do PME.
 115          */
 116
 117         return 0;
 118     }
 119
 120     /* First try to find npme as a factor of nnodes up to nnodes/3.
 121      * We start with a minimum PME node fraction of 1/16
 122      * and avoid ratios which lead to large prime factors in nnodes-npme.
 123      */
 124     npme = (nnodes + 15)/16;
 125     while (npme <= nnodes/3) {
 126         if (nnodes % npme == 0)
 127         {
 128             /* Note that fits_perf might change the PME grid,
 129              * in the current implementation it does not.
 130              */
 131             if (fits_pp_pme_perf(fplog,ir,box,mtop,nnodes,npme,ratio))
 132                         {
 133                                 break;
 134                         }
 135         }
 136         npme++;
 137     }
 138     if (npme > nnodes/3)
 139     {
 140         /* Try any possible number for npme */
 141         npme = 1;
 142         while (npme <= nnodes/2)
 143         {
 144             /* Note that fits_perf may change the PME grid */
 145             if (fits_pp_pme_perf(fplog,ir,box,mtop,nnodes,npme,ratio))
 146             {
 147                 break;
 148             }
 149             npme++;
 150         }
 151     }
 152     if (npme > nnodes/2)
 153     {
 154         gmx_fatal(FARGS,"Could not find an appropriate number of separate PME nodes. i.e. >= %5f*#nodes (%d) and <= #nodes/2 (%d) and reasonable performance wise (grid_x=%d, grid_y=%d).\n"
 155                   "Use the -npme option of mdrun or change the number of processors or the PME grid dimensions, see the manual for details.",
 156                   ratio,(int)(0.95*ratio*nnodes+0.5),nnodes/2,ir->nkx,ir->nky);
 157         /* Keep the compiler happy */
 158         npme = 0;
 159     }
 160     else
 161     {
 162         if (fplog)
 163         {
 164             fprintf(fplog,
 165                     "Will use %d particle-particle and %d PME only nodes\n"
 166                     "This is a guess, check the performance at the end of the log file\n",
 167                     nnodes-npme,npme);
 168         }
 169         fprintf(stderr,"\n"
 170                 "Will use %d particle-particle and %d PME only nodes\n"
 171                 "This is a guess, check the performance at the end of the log file\n",
 172                 nnodes-npme,npme);
 173     }
 174
 175     return npme;
 176 }
 177
 178 static int lcd(int n1,int n2)
 179 {
 180     int d,i;
 181
 182     d = 1;
 183     for(i=2; (i<=n1 && i<=n2); i++)
 184     {
 185         if (n1 % i == 0 && n2 % i == 0)
 186         {
 187             d = i;
 188         }
 189     }
 190
 191   return d;
 192 }
 193
 194 static int div_up(int n,int f)
 195 {
 196     return (n + f - 1)/f;
 197 }
 198
 199 real comm_box_frac(ivec dd_nc,real cutoff,gmx_ddbox_t *ddbox)
 200 {
 201     int  i,j,k,npp;
 202     rvec bt,nw;
 203     real comm_vol;
 204
 205     for(i=0; i<DIM; i++)
 206     {
 207         bt[i] = ddbox->box_size[i]*ddbox->skew_fac[i];
 208         nw[i] = dd_nc[i]*cutoff/bt[i];
 209     }
 210
 211     npp = 1;
 212     comm_vol = 0;
 213     for(i=0; i<DIM; i++)
 214     {
 215         if (dd_nc[i] > 1)
 216         {
 217             npp *= dd_nc[i];
 218             comm_vol += nw[i];
 219             for(j=i+1; j<DIM; j++)
 220             {
 221                 if (dd_nc[j] > 1)
 222                 {
 223                     comm_vol += nw[i]*nw[j]*M_PI/4;
 224                     for(k=j+1; k<DIM; k++)
 225                     {
 226                         if (dd_nc[k] > 1)
 227                         {
 228                             comm_vol += nw[i]*nw[j]*nw[k]*M_PI/6;
 229                         }
 230                     }
 231                 }
 232             }
 233         }
 234     }
 235
 236     return comm_vol;
 237 }
 238
 239 static bool inhomogeneous_z(const t_inputrec *ir)
 240 {
 241     return ((EEL_PME(ir->coulombtype) || ir->coulombtype==eelEWALD) &&
 242             ir->ePBC==epbcXYZ && ir->ewald_geometry==eewg3DC);
 243 }
 244
 245 static float comm_cost_est(gmx_domdec_t *dd,real limit,real cutoff,
 246                            matrix box,gmx_ddbox_t *ddbox,
 247                            int natoms,t_inputrec *ir,
 248                            float pbcdxr,
 249                            int npme_tot,ivec nc)
 250 {
 251     ivec npme={1,1,1};
 252     int  i,j,k,nk,overlap;
 253     rvec bt;
 254     float comm_vol,comm_vol_xf,comm_pme,cost_pbcdx;
 255     /* This is the cost of a pbc_dx call relative to the cost
 256      * of communicating the coordinate and force of an atom.
 257      * This will be machine dependent.
 258      * These factors are for x86 with SMP or Infiniband.
 259      */
 260     float pbcdx_rect_fac = 0.1;
 261     float pbcdx_tric_fac = 0.2;
 262
 263     /* Check the DD algorithm restrictions */
 264     if ((ir->ePBC == epbcXY && ir->nwall < 2 && nc[ZZ] > 1) ||
 265         (ir->ePBC == epbcSCREW && (nc[XX] == 1 || nc[YY] > 1 || nc[ZZ] > 1)))
 266     {
 267         return -1;
 268     }
 269
 270     if (inhomogeneous_z(ir) && nc[ZZ] > 1)
 271     {
 272         return -1;
 273     }
 274
 275     /* Check if the triclinic requirements are met */
 276     for(i=0; i<DIM; i++)
 277     {
 278         for(j=i+1; j<ddbox->npbcdim; j++)
 279         {
 280             if (box[j][i] != 0 || ir->deform[j][i] != 0 ||
 281                 (ir->epc != epcNO && ir->compress[j][i] != 0))
 282             {
 283                 if (nc[j] > 1 && nc[i] == 1)
 284                 {
 285                     return -1;
 286                 }
 287             }
 288         }
 289     }
 290
 291     for(i=0; i<DIM; i++)
 292     {
 293         bt[i] = ddbox->box_size[i]*ddbox->skew_fac[i];
 294
 295         /* Without PBC there are no cell size limits with 2 cells */
 296         if (!(i >= ddbox->npbcdim && nc[i] <= 2) && bt[i] < nc[i]*limit)
 297         {
 298             return -1;
 299         }
 300     }
 301
 302     if (npme_tot > 1)
 303     {
 304         /* Will we use 1D or 2D PME decomposition? */
 305         npme[XX] = (npme_tot % nc[XX] == 0) ? nc[XX] : npme_tot;
 306         npme[YY] = npme_tot/npme[XX];
 307     }
 308
 309     /* When two dimensions are (nearly) equal, use more cells
 310      * for the smallest index, so the decomposition does not
 311      * depend sensitively on the rounding of the box elements.
 312      */
 313     for(i=0; i<DIM; i++)
 314     {
 315         for(j=i+1; j<DIM; j++)
 316         {
 317             /* Check if dimension i and j are equivalent,
 318              * both for PME and for the box size.
 319              * The XX/YY check is a bit compact. If nc[YY]==npme[YY]
 320              * this means the swapped nc has nc[XX]=npme[XX],
 321              * and we can also swap X and Y for PME.
 322              */
 323             if (!((i == XX && j == YY && npme[YY] > 1 && nc[YY] != npme[YY]) ||
 324                   (i == YY && j == ZZ && npme[YY] > 1)) &&
 325                 fabs(bt[j] - bt[i]) < 0.01*bt[i] && nc[j] > nc[i])
 326             {
 327                 return -1;
 328             }
 329         }
 330     }
 331
 332     /* This function determines only half of the communication cost.
 333      * All PP, PME and PP-PME communication is symmetric
 334      * and the "back"-communication cost is identical to the forward cost.
 335      */
 336
 337     comm_vol = comm_box_frac(nc,cutoff,ddbox);
 338
 339     comm_pme = 0;
 340     for(i=0; i<2; i++)
 341     {
 342         /* Determine the largest volume for PME x/f redistribution */
 343         if (nc[i] % npme[i] != 0)
 344         {
 345             if (nc[i] > npme[i])
 346             {
 347                 comm_vol_xf = (npme[i]==2 ? 1.0/3.0 : 0.5);
 348             }
 349             else
 350             {
 351                 comm_vol_xf = 1.0 - lcd(nc[i],npme[i])/(double)npme[i];
 352             }
 353             comm_pme += 3*natoms*comm_vol_xf;
 354         }
 355
 356         /* Grid overlap communication */
 357         if (npme[i] > 1)
 358         {
 359             nk = (i==0 ? ir->nkx : ir->nky);
 360             overlap = (nk % npme[i] == 0 ? ir->pme_order-1 : ir->pme_order);
 361             comm_pme += npme[i]*overlap*ir->nkx*ir->nky*ir->nkz/nk;
 362         }
 363     }
 364
 365     /* PME FFT communication volume.
 366      * This only takes the communication into account and not imbalance
 367      * in the calculation. But the imbalance in communication and calculation
 368      * are similar and therefore these formulas also prefer load balance
 369      * in the FFT and pme_solve calculation.
 370      */
 371     comm_pme += (npme[YY] - 1)*npme[YY]*div_up(ir->nky,npme[YY])*div_up(ir->nkz,npme[YY])*ir->nkx;
 372     comm_pme += (npme[XX] - 1)*npme[XX]*div_up(ir->nkx,npme[XX])*div_up(ir->nky,npme[XX])*ir->nkz;
 373
 374     /* Add cost of pbc_dx for bondeds */
 375     cost_pbcdx = 0;
 376     if ((nc[XX] == 1 || nc[YY] == 1) || (nc[ZZ] == 1 && ir->ePBC != epbcXY))
 377     {
 378         if ((ddbox->tric_dir[XX] && nc[XX] == 1) ||
 379             (ddbox->tric_dir[YY] && nc[YY] == 1))
 380         {
 381             cost_pbcdx = pbcdxr*pbcdx_tric_fac;
 382         }
 383         else
 384         {
 385             cost_pbcdx = pbcdxr*pbcdx_rect_fac;
 386         }
 387     }
 388
 389     if (debug)
 390     {
 391         fprintf(debug,
 392                 "nc %2d %2d %2d %2d %2d vol pp %6.4f pbcdx %6.4f pme %9.3e tot %9.3e\n",
 393                 nc[XX],nc[YY],nc[ZZ],npme[XX],npme[YY],
 394                 comm_vol,cost_pbcdx,comm_pme,
 395                 3*natoms*(comm_vol + cost_pbcdx) + comm_pme);
 396     }
 397
 398     return 3*natoms*(comm_vol + cost_pbcdx) + comm_pme;
 399 }
 400
 401 static void assign_factors(gmx_domdec_t *dd,
 402                            real limit,real cutoff,
 403                            matrix box,gmx_ddbox_t *ddbox,
 404                            int natoms,t_inputrec *ir,
 405                            float pbcdxr,int npme,
 406                            int ndiv,int *div,int *mdiv,ivec ir_try,ivec opt)
 407 {
 408     int x,y,z,i;
 409     float ce;
 410
 411     if (ndiv == 0)
 412     {
 413         ce = comm_cost_est(dd,limit,cutoff,box,ddbox,
 414                            natoms,ir,pbcdxr,npme,ir_try);
 415         if (ce >= 0 && (opt[XX] == 0 ||
 416                         ce < comm_cost_est(dd,limit,cutoff,box,ddbox,
 417                                            natoms,ir,pbcdxr,
 418                                            npme,opt)))
 419         {
 420             copy_ivec(ir_try,opt);
 421         }
 422
 423         return;
 424     }
 425
 426     for(x=mdiv[0]; x>=0; x--)
 427     {
 428         for(i=0; i<x; i++)
 429         {
 430             ir_try[XX] *= div[0];
 431         }
 432         for(y=mdiv[0]-x; y>=0; y--)
 433         {
 434             for(i=0; i<y; i++)
 435             {
 436                 ir_try[YY] *= div[0];
 437             }
 438             for(i=0; i<mdiv[0]-x-y; i++)
 439             {
 440                 ir_try[ZZ] *= div[0];
 441             }
 442
 443             /* recurse */
 444             assign_factors(dd,limit,cutoff,box,ddbox,natoms,ir,pbcdxr,npme,
 445                            ndiv-1,div+1,mdiv+1,ir_try,opt);
 446
 447             for(i=0; i<mdiv[0]-x-y; i++)
 448             {
 449                 ir_try[ZZ] /= div[0];
 450             }
 451             for(i=0; i<y; i++)
 452             {
 453                 ir_try[YY] /= div[0];
 454             }
 455         }
 456         for(i=0; i<x; i++)
 457         {
 458             ir_try[XX] /= div[0];
 459         }
 460     }
 461 }
 462
 463 static real optimize_ncells(FILE *fplog,
 464                             int nnodes_tot,int npme_only,
 465                             bool bDynLoadBal,real dlb_scale,
 466                             gmx_mtop_t *mtop,matrix box,gmx_ddbox_t *ddbox,
 467                             t_inputrec *ir,
 468                             gmx_domdec_t *dd,
 469                             real cellsize_limit,real cutoff,
 470                             bool bInterCGBondeds,bool bInterCGMultiBody,
 471                             ivec nc)
 472 {
 473     int npp,npme,ndiv,*div,*mdiv,d,nmax;
 474     bool bExcl_pbcdx;
 475     float pbcdxr;
 476     real limit;
 477     ivec itry;
 478
 479     limit  = cellsize_limit;
 480
 481     dd->nc[XX] = 1;
 482     dd->nc[YY] = 1;
 483     dd->nc[ZZ] = 1;
 484
 485     npp = nnodes_tot - npme_only;
 486     if (EEL_PME(ir->coulombtype))
 487     {
 488         npme = (npme_only > 0 ? npme_only : npp);
 489     }
 490     else
 491     {
 492         npme = 0;
 493     }
 494
 495     if (bInterCGBondeds)
 496     {
 497         /* For Ewald exclusions pbc_dx is not called */
 498         bExcl_pbcdx =
 499             (IR_EXCL_FORCES(*ir) && !EEL_FULL(ir->coulombtype));
 500         pbcdxr = (double)n_bonded_dx(mtop,bExcl_pbcdx)/(double)mtop->natoms;
 501     }
 502     else
 503     {
 504         /* Every molecule is a single charge group: no pbc required */
 505         pbcdxr = 0;
 506     }
 507     /* Add a margin for DLB and/or pressure scaling */
 508     if (bDynLoadBal)
 509     {
 510         if (dlb_scale >= 1.0)
 511         {
 512             gmx_fatal(FARGS,"The value for option -dds should be smaller than 1");
 513         }
 514         if (fplog)
 515         {
 516             fprintf(fplog,"Scaling the initial minimum size with 1/%g (option -dds) = %g\n",dlb_scale,1/dlb_scale);
 517         }
 518         limit /= dlb_scale;
 519     }
 520     else if (ir->epc != epcNO)
 521     {
 522         if (fplog)
 523         {
 524             fprintf(fplog,"To account for pressure scaling, scaling the initial minimum size with %g\n",DD_GRID_MARGIN_PRES_SCALE);
 525             limit *= DD_GRID_MARGIN_PRES_SCALE;
 526         }
 527     }
 528
 529     if (fplog)
 530     {
 531         fprintf(fplog,"Optimizing the DD grid for %d cells with a minimum initial size of %.3f nm\n",npp,limit);
 532
 533         if (inhomogeneous_z(ir))
 534         {
 535             fprintf(fplog,"Ewald_geometry=%s: assuming inhomogeneous particle distribution in z, will not decompose in z.\n",eewg_names[ir->ewald_geometry]);
 536         }
 537
 538         if (limit > 0)
 539         {
 540             fprintf(fplog,"The maximum allowed number of cells is:");
 541             for(d=0; d<DIM; d++)
 542             {
 543                 nmax = (int)(ddbox->box_size[d]*ddbox->skew_fac[d]/limit);
 544                 if (d >= ddbox->npbcdim && nmax < 2)
 545                 {
 546                     nmax = 2;
 547                 }
 548                 if (d == ZZ && inhomogeneous_z(ir))
 549                 {
 550                     nmax = 1;
 551                 }
 552                 fprintf(fplog," %c %d",'X' + d,nmax);
 553             }
 554             fprintf(fplog,"\n");
 555         }
 556     }
 557
 558     if (debug)
 559     {
 560         fprintf(debug,"Average nr of pbc_dx calls per atom %.2f\n",pbcdxr);
 561     }
 562
 563     /* Decompose npp in factors */
 564     ndiv = factorize(npp,&div,&mdiv);
 565
 566     itry[XX] = 1;
 567     itry[YY] = 1;
 568     itry[ZZ] = 1;
 569     clear_ivec(nc);
 570     assign_factors(dd,limit,cutoff,box,ddbox,mtop->natoms,ir,pbcdxr,
 571                    npme,ndiv,div,mdiv,itry,nc);
 572
 573     sfree(div);
 574     sfree(mdiv);
 575
 576     return limit;
 577 }
 578
 579 real dd_choose_grid(FILE *fplog,
 580                     t_commrec *cr,gmx_domdec_t *dd,t_inputrec *ir,
 581                     gmx_mtop_t *mtop,matrix box,gmx_ddbox_t *ddbox,
 582                     bool bDynLoadBal,real dlb_scale,
 583                     real cellsize_limit,real cutoff_dd,
 584                     bool bInterCGBondeds,bool bInterCGMultiBody)
 585 {
 586     int  npme,nkx,nky;
 587     real limit;
 588
 589     if (MASTER(cr))
 590     {
 591         if (EEL_PME(ir->coulombtype))
 592         {
 593             if (cr->npmenodes >= 0)
 594             {
 595                 if (cr->nnodes <= 2 && cr->npmenodes > 0)
 596                 {
 597                     gmx_fatal(FARGS,
 598                               "Can not have separate PME nodes with 2 or less nodes");
 599                 }
 600             }
 601             else
 602             {
 603                 if (cr->nnodes <= 10)
 604                 {
 605                     cr->npmenodes = 0;
 606                 }
 607                 else
 608                 {
 609                     cr->npmenodes = guess_npme(fplog,mtop,ir,box,cr->nnodes);
 610                 }
 611             }
 612             if (fplog)
 613             {
 614                 fprintf(fplog,"Using %d separate PME nodes\n",cr->npmenodes);
 615             }
 616         }
 617         else
 618         {
 619             if (cr->npmenodes < 0)
 620             {
 621                 cr->npmenodes = 0;
 622             }
 623         }
 624
 625         limit = optimize_ncells(fplog,cr->nnodes,cr->npmenodes,
 626                                 bDynLoadBal,dlb_scale,
 627                                 mtop,box,ddbox,ir,dd,
 628                                 cellsize_limit,cutoff_dd,
 629                                 bInterCGBondeds,bInterCGMultiBody,
 630                                 dd->nc);
 631     }
 632     else
 633     {
 634         limit = 0;
 635     }
 636     /* Communicate the information set by the master to all nodes */
 637     gmx_bcast(sizeof(dd->nc),dd->nc,cr);
 638     if (EEL_PME(ir->coulombtype))
 639     {
 640         gmx_bcast(sizeof(ir->nkx),&ir->nkx,cr);
 641         gmx_bcast(sizeof(ir->nky),&ir->nky,cr);
 642         gmx_bcast(sizeof(cr->npmenodes),&cr->npmenodes,cr);
 643     }
 644     else
 645     {
 646         cr->npmenodes = 0;
 647     }
 648
 649     return limit;
 650 }