2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 1991-2008
5 * Copyright (c) 2012,2013, by the GROMACS development team, led by
6 * David van der Spoel, Berk Hess, Erik Lindahl, and including many
7 * others, as listed in the AUTHORS file in the top-level source
8 * directory and at http://www.gromacs.org.
10 * GROMACS is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU Lesser General Public License
12 * as published by the Free Software Foundation; either version 2.1
13 * of the License, or (at your option) any later version.
15 * GROMACS is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * Lesser General Public License for more details.
20 * You should have received a copy of the GNU Lesser General Public
21 * License along with GROMACS; if not, see
22 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
23 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
25 * If you want to redistribute modifications to GROMACS, please
26 * consider that scientific software is very special. Version
27 * control is crucial - bugs must be traceable. We will be happy to
28 * consider code for inclusion in the official distribution, but
29 * derived work must not be called official GROMACS. Details are found
30 * in the README & COPYING files - if they are missing, get the
31 * official version at http://www.gromacs.org.
33 * To help us fund GROMACS development, we humbly ask that you cite
34 * the research papers on the package. Check out http://www.gromacs.org.
52 /* Margin for setting up the DD grid */
53 #define DD_GRID_MARGIN_PRES_SCALE 1.05
55 static int factorize(int n, int **fac, int **mfac)
61 gmx_fatal(FARGS, "Can only factorize positive integers.");
64 /* Decompose n in factors */
73 if (ndiv == 0 || (*fac)[ndiv-1] != d)
87 static gmx_bool largest_divisor(int n)
89 int ndiv, *div, *mdiv, ldiv;
91 ndiv = factorize(n, &div, &mdiv);
99 static int lcd(int n1, int n2)
104 for (i = 2; (i <= n1 && i <= n2); i++)
106 if (n1 % i == 0 && n2 % i == 0)
115 static gmx_bool fits_pme_ratio(int nnodes, int npme, float ratio)
117 return ((double)npme/(double)nnodes > 0.95*ratio);
120 static gmx_bool fits_pp_pme_perf(FILE *fplog,
121 t_inputrec *ir, matrix box, gmx_mtop_t *mtop,
122 int nnodes, int npme, float ratio)
124 int ndiv, *div, *mdiv, ldiv;
125 int npp_root3, npme_root2;
127 ndiv = factorize(nnodes-npme, &div, &mdiv);
132 npp_root3 = (int)(pow(nnodes-npme, 1.0/3.0) + 0.5);
133 npme_root2 = (int)(sqrt(npme) + 0.5);
135 /* The check below gives a reasonable division:
136 * factor 5 allowed at 5 or more PP nodes,
137 * factor 7 allowed at 49 or more PP nodes.
139 if (ldiv > 3 + npp_root3)
144 /* Check if the number of PP and PME nodes have a reasonable sized
145 * denominator in common, such that we can use 2D PME decomposition
146 * when required (which requires nx_pp == nx_pme).
147 * The factor of 2 allows for a maximum ratio of 2^2=4
148 * between nx_pme and ny_pme.
150 if (lcd(nnodes-npme, npme)*2 < npme_root2)
155 /* Does this division gives a reasonable PME load? */
156 return fits_pme_ratio(nnodes, npme, ratio);
159 static int guess_npme(FILE *fplog, gmx_mtop_t *mtop, t_inputrec *ir, matrix box,
166 ratio = pme_load_estimate(mtop, ir, box);
170 fprintf(fplog, "Guess for relative PME load: %.2f\n", ratio);
173 /* We assume the optimal node ratio is close to the load ratio.
174 * The communication load is neglected,
175 * but (hopefully) this will balance out between PP and PME.
178 if (!fits_pme_ratio(nnodes, nnodes/2, ratio))
180 /* We would need more than nnodes/2 PME only nodes,
181 * which is not possible. Since the PME load is very high,
182 * we will not loose much performance when all nodes do PME.
188 /* First try to find npme as a factor of nnodes up to nnodes/3.
189 * We start with a minimum PME node fraction of 1/16
190 * and avoid ratios which lead to large prime factors in nnodes-npme.
192 npme = (nnodes + 15)/16;
193 while (npme <= nnodes/3)
195 if (nnodes % npme == 0)
197 /* Note that fits_perf might change the PME grid,
198 * in the current implementation it does not.
200 if (fits_pp_pme_perf(fplog, ir, box, mtop, nnodes, npme, ratio))
209 /* Try any possible number for npme */
211 while (npme <= nnodes/2)
213 /* Note that fits_perf may change the PME grid */
214 if (fits_pp_pme_perf(fplog, ir, box, mtop, nnodes, npme, ratio))
223 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"
224 "Use the -npme option of mdrun or change the number of processors or the PME grid dimensions, see the manual for details.",
225 ratio, (int)(0.95*ratio*nnodes+0.5), nnodes/2, ir->nkx, ir->nky);
226 /* Keep the compiler happy */
234 "Will use %d particle-particle and %d PME only nodes\n"
235 "This is a guess, check the performance at the end of the log file\n",
239 "Will use %d particle-particle and %d PME only nodes\n"
240 "This is a guess, check the performance at the end of the log file\n",
247 static int div_up(int n, int f)
249 return (n + f - 1)/f;
252 real comm_box_frac(ivec dd_nc, real cutoff, gmx_ddbox_t *ddbox)
258 for (i = 0; i < DIM; i++)
260 bt[i] = ddbox->box_size[i]*ddbox->skew_fac[i];
261 nw[i] = dd_nc[i]*cutoff/bt[i];
266 for (i = 0; i < DIM; i++)
272 for (j = i+1; j < DIM; j++)
276 comm_vol += nw[i]*nw[j]*M_PI/4;
277 for (k = j+1; k < DIM; k++)
281 comm_vol += nw[i]*nw[j]*nw[k]*M_PI/6;
292 static gmx_bool inhomogeneous_z(const t_inputrec *ir)
294 return ((EEL_PME(ir->coulombtype) || ir->coulombtype == eelEWALD) &&
295 ir->ePBC == epbcXYZ && ir->ewald_geometry == eewg3DC);
298 /* Avoid integer overflows */
299 static float comm_pme_cost_vol(int npme, int a, int b, int c)
305 comm_vol *= div_up(a, npme);
306 comm_vol *= div_up(b, npme);
311 static float comm_cost_est(gmx_domdec_t *dd, real limit, real cutoff,
312 matrix box, gmx_ddbox_t *ddbox,
313 int natoms, t_inputrec *ir,
315 int npme_tot, ivec nc)
317 ivec npme = {1, 1, 1};
318 int i, j, k, nk, overlap;
320 float comm_vol, comm_vol_xf, comm_pme, cost_pbcdx;
321 /* This is the cost of a pbc_dx call relative to the cost
322 * of communicating the coordinate and force of an atom.
323 * This will be machine dependent.
324 * These factors are for x86 with SMP or Infiniband.
326 float pbcdx_rect_fac = 0.1;
327 float pbcdx_tric_fac = 0.2;
330 /* Check the DD algorithm restrictions */
331 if ((ir->ePBC == epbcXY && ir->nwall < 2 && nc[ZZ] > 1) ||
332 (ir->ePBC == epbcSCREW && (nc[XX] == 1 || nc[YY] > 1 || nc[ZZ] > 1)))
337 if (inhomogeneous_z(ir) && nc[ZZ] > 1)
342 assert(ddbox->npbcdim <= DIM);
344 /* Check if the triclinic requirements are met */
345 for (i = 0; i < DIM; i++)
347 for (j = i+1; j < ddbox->npbcdim; j++)
349 if (box[j][i] != 0 || ir->deform[j][i] != 0 ||
350 (ir->epc != epcNO && ir->compress[j][i] != 0))
352 if (nc[j] > 1 && nc[i] == 1)
360 for (i = 0; i < DIM; i++)
362 bt[i] = ddbox->box_size[i]*ddbox->skew_fac[i];
364 /* Without PBC and with 2 cells, there are no lower limits on the cell size */
365 if (!(i >= ddbox->npbcdim && nc[i] <= 2) && bt[i] < nc[i]*limit)
369 /* With PBC, check if the cut-off fits in nc[i]-1 cells */
370 if (i < ddbox->npbcdim && nc[i] > 1 && (nc[i] - 1)*bt[i] < nc[i]*cutoff)
378 /* The following choices should match those
379 * in init_domain_decomposition in domdec.c.
381 if (nc[XX] == 1 && nc[YY] > 1)
386 else if (nc[YY] == 1)
393 /* Will we use 1D or 2D PME decomposition? */
394 npme[XX] = (npme_tot % nc[XX] == 0) ? nc[XX] : npme_tot;
395 npme[YY] = npme_tot/npme[XX];
399 /* When two dimensions are (nearly) equal, use more cells
400 * for the smallest index, so the decomposition does not
401 * depend sensitively on the rounding of the box elements.
403 for (i = 0; i < DIM; i++)
405 for (j = i+1; j < DIM; j++)
407 /* Check if the box size is nearly identical,
408 * in that case we prefer nx > ny and ny > nz.
410 if (fabs(bt[j] - bt[i]) < 0.01*bt[i] && nc[j] > nc[i])
412 /* The XX/YY check is a bit compact. If nc[YY]==npme[YY]
413 * this means the swapped nc has nc[XX]==npme[XX],
414 * and we can also swap X and Y for PME.
416 /* Check if dimension i and j are equivalent for PME.
417 * For x/y: if nc[YY]!=npme[YY], we can not swap x/y
418 * For y/z: we can not have PME decomposition in z
421 !((i == XX && j == YY && nc[YY] != npme[YY]) ||
422 (i == YY && j == ZZ && npme[YY] > 1)))
430 /* This function determines only half of the communication cost.
431 * All PP, PME and PP-PME communication is symmetric
432 * and the "back"-communication cost is identical to the forward cost.
435 comm_vol = comm_box_frac(nc, cutoff, ddbox);
438 for (i = 0; i < 2; i++)
440 /* Determine the largest volume for PME x/f redistribution */
441 if (nc[i] % npme[i] != 0)
445 comm_vol_xf = (npme[i] == 2 ? 1.0/3.0 : 0.5);
449 comm_vol_xf = 1.0 - lcd(nc[i], npme[i])/(double)npme[i];
451 comm_pme += 3*natoms*comm_vol_xf;
454 /* Grid overlap communication */
457 nk = (i == 0 ? ir->nkx : ir->nky);
458 overlap = (nk % npme[i] == 0 ? ir->pme_order-1 : ir->pme_order);
466 /* Old line comm_pme += npme[i]*overlap*ir->nkx*ir->nky*ir->nkz/nk; */
470 /* PME FFT communication volume.
471 * This only takes the communication into account and not imbalance
472 * in the calculation. But the imbalance in communication and calculation
473 * are similar and therefore these formulas also prefer load balance
474 * in the FFT and pme_solve calculation.
476 comm_pme += comm_pme_cost_vol(npme[YY], ir->nky, ir->nkz, ir->nkx);
477 comm_pme += comm_pme_cost_vol(npme[XX], ir->nkx, ir->nky, ir->nkz);
479 /* Add cost of pbc_dx for bondeds */
481 if ((nc[XX] == 1 || nc[YY] == 1) || (nc[ZZ] == 1 && ir->ePBC != epbcXY))
483 if ((ddbox->tric_dir[XX] && nc[XX] == 1) ||
484 (ddbox->tric_dir[YY] && nc[YY] == 1))
486 cost_pbcdx = pbcdxr*pbcdx_tric_fac;
490 cost_pbcdx = pbcdxr*pbcdx_rect_fac;
497 "nc %2d %2d %2d %2d %2d vol pp %6.4f pbcdx %6.4f pme %9.3e tot %9.3e\n",
498 nc[XX], nc[YY], nc[ZZ], npme[XX], npme[YY],
499 comm_vol, cost_pbcdx, comm_pme,
500 3*natoms*(comm_vol + cost_pbcdx) + comm_pme);
503 return 3*natoms*(comm_vol + cost_pbcdx) + comm_pme;
506 static void assign_factors(gmx_domdec_t *dd,
507 real limit, real cutoff,
508 matrix box, gmx_ddbox_t *ddbox,
509 int natoms, t_inputrec *ir,
510 float pbcdxr, int npme,
511 int ndiv, int *div, int *mdiv, ivec ir_try, ivec opt)
518 ce = comm_cost_est(dd, limit, cutoff, box, ddbox,
519 natoms, ir, pbcdxr, npme, ir_try);
520 if (ce >= 0 && (opt[XX] == 0 ||
521 ce < comm_cost_est(dd, limit, cutoff, box, ddbox,
525 copy_ivec(ir_try, opt);
531 for (x = mdiv[0]; x >= 0; x--)
533 for (i = 0; i < x; i++)
535 ir_try[XX] *= div[0];
537 for (y = mdiv[0]-x; y >= 0; y--)
539 for (i = 0; i < y; i++)
541 ir_try[YY] *= div[0];
543 for (i = 0; i < mdiv[0]-x-y; i++)
545 ir_try[ZZ] *= div[0];
549 assign_factors(dd, limit, cutoff, box, ddbox, natoms, ir, pbcdxr, npme,
550 ndiv-1, div+1, mdiv+1, ir_try, opt);
552 for (i = 0; i < mdiv[0]-x-y; i++)
554 ir_try[ZZ] /= div[0];
556 for (i = 0; i < y; i++)
558 ir_try[YY] /= div[0];
561 for (i = 0; i < x; i++)
563 ir_try[XX] /= div[0];
568 static real optimize_ncells(FILE *fplog,
569 int nnodes_tot, int npme_only,
570 gmx_bool bDynLoadBal, real dlb_scale,
571 gmx_mtop_t *mtop, matrix box, gmx_ddbox_t *ddbox,
574 real cellsize_limit, real cutoff,
575 gmx_bool bInterCGBondeds, gmx_bool bInterCGMultiBody,
578 int npp, npme, ndiv, *div, *mdiv, d, nmax;
579 gmx_bool bExcl_pbcdx;
584 limit = cellsize_limit;
590 npp = nnodes_tot - npme_only;
591 if (EEL_PME(ir->coulombtype))
593 npme = (npme_only > 0 ? npme_only : npp);
602 /* For Ewald exclusions pbc_dx is not called */
604 (IR_EXCL_FORCES(*ir) && !EEL_FULL(ir->coulombtype));
605 pbcdxr = (double)n_bonded_dx(mtop, bExcl_pbcdx)/(double)mtop->natoms;
609 /* Every molecule is a single charge group: no pbc required */
612 /* Add a margin for DLB and/or pressure scaling */
615 if (dlb_scale >= 1.0)
617 gmx_fatal(FARGS, "The value for option -dds should be smaller than 1");
621 fprintf(fplog, "Scaling the initial minimum size with 1/%g (option -dds) = %g\n", dlb_scale, 1/dlb_scale);
625 else if (ir->epc != epcNO)
629 fprintf(fplog, "To account for pressure scaling, scaling the initial minimum size with %g\n", DD_GRID_MARGIN_PRES_SCALE);
630 limit *= DD_GRID_MARGIN_PRES_SCALE;
636 fprintf(fplog, "Optimizing the DD grid for %d cells with a minimum initial size of %.3f nm\n", npp, limit);
638 if (inhomogeneous_z(ir))
640 fprintf(fplog, "Ewald_geometry=%s: assuming inhomogeneous particle distribution in z, will not decompose in z.\n", eewg_names[ir->ewald_geometry]);
645 fprintf(fplog, "The maximum allowed number of cells is:");
646 for (d = 0; d < DIM; d++)
648 nmax = (int)(ddbox->box_size[d]*ddbox->skew_fac[d]/limit);
649 if (d >= ddbox->npbcdim && nmax < 2)
653 if (d == ZZ && inhomogeneous_z(ir))
657 fprintf(fplog, " %c %d", 'X' + d, nmax);
659 fprintf(fplog, "\n");
665 fprintf(debug, "Average nr of pbc_dx calls per atom %.2f\n", pbcdxr);
668 /* Decompose npp in factors */
669 ndiv = factorize(npp, &div, &mdiv);
675 assign_factors(dd, limit, cutoff, box, ddbox, mtop->natoms, ir, pbcdxr,
676 npme, ndiv, div, mdiv, itry, nc);
684 real dd_choose_grid(FILE *fplog,
685 t_commrec *cr, gmx_domdec_t *dd, t_inputrec *ir,
686 gmx_mtop_t *mtop, matrix box, gmx_ddbox_t *ddbox,
687 gmx_bool bDynLoadBal, real dlb_scale,
688 real cellsize_limit, real cutoff_dd,
689 gmx_bool bInterCGBondeds, gmx_bool bInterCGMultiBody)
691 gmx_large_int_t nnodes_div, ldiv;
696 nnodes_div = cr->nnodes;
697 if (EEL_PME(ir->coulombtype))
699 if (cr->npmenodes > 0)
704 "Can not have separate PME nodes with 2 or less nodes");
706 if (cr->npmenodes >= cr->nnodes)
709 "Can not have %d separate PME nodes with just %d total nodes",
710 cr->npmenodes, cr->nnodes);
713 /* If the user purposely selected the number of PME nodes,
714 * only check for large primes in the PP node count.
716 nnodes_div -= cr->npmenodes;
726 ldiv = largest_divisor(nnodes_div);
727 /* Check if the largest divisor is more than nnodes^2/3 */
728 if (ldiv*ldiv*ldiv > nnodes_div*nnodes_div)
730 gmx_fatal(FARGS, "The number of nodes you selected (%d) contains a large prime factor %d. In most cases this will lead to bad performance. Choose a number with smaller prime factors or set the decomposition (option -dd) manually.",
735 if (EEL_PME(ir->coulombtype))
737 if (cr->npmenodes < 0)
739 /* Use PME nodes when the number of nodes is more than 16 */
740 if (cr->nnodes <= 18)
745 fprintf(fplog, "Using %d separate PME nodes, as there are too few total\n nodes for efficient splitting\n", cr->npmenodes);
750 cr->npmenodes = guess_npme(fplog, mtop, ir, box, cr->nnodes);
753 fprintf(fplog, "Using %d separate PME nodes, as guessed by mdrun\n", cr->npmenodes);
761 fprintf(fplog, "Using %d separate PME nodes, per user request\n", cr->npmenodes);
766 limit = optimize_ncells(fplog, cr->nnodes, cr->npmenodes,
767 bDynLoadBal, dlb_scale,
768 mtop, box, ddbox, ir, dd,
769 cellsize_limit, cutoff_dd,
770 bInterCGBondeds, bInterCGMultiBody,
777 /* Communicate the information set by the master to all nodes */
778 gmx_bcast(sizeof(dd->nc), dd->nc, cr);
779 if (EEL_PME(ir->coulombtype))
781 gmx_bcast(sizeof(ir->nkx), &ir->nkx, cr);
782 gmx_bcast(sizeof(ir->nky), &ir->nky, cr);
783 gmx_bcast(sizeof(cr->npmenodes), &cr->npmenodes, cr);