2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2008,2009,2010,2011,2012,2013,2014,2015, 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.
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.
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.
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.
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.
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.
38 * \brief This file defines functions used by the domdec module
39 * in its initial setup phase.
41 * \author Berk Hess <hess@kth.se>
42 * \ingroup module_domdec
52 #include "gromacs/domdec/domdec.h"
53 #include "gromacs/domdec/domdec_struct.h"
54 #include "gromacs/gmxlib/network.h"
55 #include "gromacs/math/vec.h"
56 #include "gromacs/mdlib/perf_est.h"
57 #include "gromacs/mdtypes/commrec.h"
58 #include "gromacs/mdtypes/inputrec.h"
59 #include "gromacs/mdtypes/md_enums.h"
60 #include "gromacs/pbcutil/pbc.h"
61 #include "gromacs/utility/fatalerror.h"
62 #include "gromacs/utility/smalloc.h"
64 /*! \brief Margin for setting up the DD grid */
65 #define DD_GRID_MARGIN_PRES_SCALE 1.05
67 /*! \brief Factorize \p n.
69 * \param[in] n Value to factorize
70 * \param[out] fac Pointer to array of factors (to be allocated in this function)
71 * \param[out] mfac Pointer to array of the number of times each factor repeats in the factorization (to be allocated in this function)
72 * \return The number of unique factors
74 static int factorize(int n, int **fac, int **mfac)
80 gmx_fatal(FARGS, "Can only factorize positive integers.");
83 /* Decompose n in factors */
92 if (ndiv == 0 || (*fac)[ndiv-1] != d)
106 /*! \brief Find largest divisor of \p n smaller than \p n*/
107 static gmx_bool largest_divisor(int n)
109 int ndiv, *div, *mdiv, ldiv;
111 ndiv = factorize(n, &div, &mdiv);
119 /*! \brief Compute largest common divisor of \p n1 and \b n2 */
120 static int lcd(int n1, int n2)
125 for (i = 2; (i <= n1 && i <= n2); i++)
127 if (n1 % i == 0 && n2 % i == 0)
136 /*! \brief Returns TRUE when there are enough PME ranks for the ratio */
137 static gmx_bool fits_pme_ratio(int nrank_tot, int nrank_pme, float ratio)
139 return ((double)nrank_pme/(double)nrank_tot > 0.95*ratio);
142 /*! \brief Returns TRUE when npme out of ntot ranks doing PME is expected to give reasonable performance */
143 static gmx_bool fits_pp_pme_perf(int ntot, int npme, float ratio)
145 int ndiv, *div, *mdiv, ldiv;
146 int npp_root3, npme_root2;
148 ndiv = factorize(ntot - npme, &div, &mdiv);
153 npp_root3 = static_cast<int>(std::cbrt(ntot - npme) + 0.5);
154 npme_root2 = static_cast<int>(std::sqrt(static_cast<double>(npme)) + 0.5);
156 /* The check below gives a reasonable division:
157 * factor 5 allowed at 5 or more PP ranks,
158 * factor 7 allowed at 49 or more PP ranks.
160 if (ldiv > 3 + npp_root3)
165 /* Check if the number of PP and PME ranks have a reasonable sized
166 * denominator in common, such that we can use 2D PME decomposition
167 * when required (which requires nx_pp == nx_pme).
168 * The factor of 2 allows for a maximum ratio of 2^2=4
169 * between nx_pme and ny_pme.
171 if (lcd(ntot - npme, npme)*2 < npme_root2)
176 /* Does this division gives a reasonable PME load? */
177 return fits_pme_ratio(ntot, npme, ratio);
180 /*! \brief Make a guess for the number of PME ranks to use. */
181 static int guess_npme(FILE *fplog, gmx_mtop_t *mtop, t_inputrec *ir, matrix box,
187 ratio = pme_load_estimate(mtop, ir, box);
191 fprintf(fplog, "Guess for relative PME load: %.2f\n", ratio);
194 /* We assume the optimal rank ratio is close to the load ratio.
195 * The communication load is neglected,
196 * but (hopefully) this will balance out between PP and PME.
199 if (!fits_pme_ratio(nrank_tot, nrank_tot/2, ratio))
201 /* We would need more than nrank_tot/2 PME only nodes,
202 * which is not possible. Since the PME load is very high,
203 * we will not loose much performance when all ranks do PME.
209 /* First try to find npme as a factor of nrank_tot up to nrank_tot/3.
210 * We start with a minimum PME node fraction of 1/16
211 * and avoid ratios which lead to large prime factors in nnodes-npme.
213 npme = (nrank_tot + 15)/16;
214 while (npme <= nrank_tot/3)
216 if (nrank_tot % npme == 0)
218 /* Note that fits_perf might change the PME grid,
219 * in the current implementation it does not.
221 if (fits_pp_pme_perf(nrank_tot, npme, ratio))
228 if (npme > nrank_tot/3)
230 /* Try any possible number for npme */
232 while (npme <= nrank_tot/2)
234 /* Note that fits_perf may change the PME grid */
235 if (fits_pp_pme_perf(nrank_tot, npme, ratio))
242 if (npme > nrank_tot/2)
244 gmx_fatal(FARGS, "Could not find an appropriate number of separate PME ranks. i.e. >= %5f*#ranks (%d) and <= #ranks/2 (%d) and reasonable performance wise (grid_x=%d, grid_y=%d).\n"
245 "Use the -npme option of mdrun or change the number of ranks or the PME grid dimensions, see the manual for details.",
246 ratio, (int)(0.95*ratio*nrank_tot + 0.5), nrank_tot/2, ir->nkx, ir->nky);
247 /* Keep the compiler happy */
255 "Will use %d particle-particle and %d PME only ranks\n"
256 "This is a guess, check the performance at the end of the log file\n",
257 nrank_tot - npme, npme);
260 "Will use %d particle-particle and %d PME only ranks\n"
261 "This is a guess, check the performance at the end of the log file\n",
262 nrank_tot - npme, npme);
268 /*! \brief Return \p n divided by \p f rounded up to the next integer. */
269 static int div_up(int n, int f)
271 return (n + f - 1)/f;
274 real comm_box_frac(ivec dd_nc, real cutoff, gmx_ddbox_t *ddbox)
280 for (i = 0; i < DIM; i++)
282 real bt = ddbox->box_size[i]*ddbox->skew_fac[i];
283 nw[i] = dd_nc[i]*cutoff/bt;
287 for (i = 0; i < DIM; i++)
292 for (j = i+1; j < DIM; j++)
296 comm_vol += nw[i]*nw[j]*M_PI/4;
297 for (k = j+1; k < DIM; k++)
301 comm_vol += nw[i]*nw[j]*nw[k]*M_PI/6;
312 /*! \brief Return whether the DD inhomogeneous in the z direction */
313 static gmx_bool inhomogeneous_z(const t_inputrec *ir)
315 return ((EEL_PME(ir->coulombtype) || ir->coulombtype == eelEWALD) &&
316 ir->ePBC == epbcXYZ && ir->ewald_geometry == eewg3DC);
319 /*! \brief Estimate cost of PME FFT communication
321 * This only takes the communication into account and not imbalance
322 * in the calculation. But the imbalance in communication and calculation
323 * are similar and therefore these formulas also prefer load balance
324 * in the FFT and pme_solve calculation.
326 static float comm_pme_cost_vol(int npme, int a, int b, int c)
328 /* We use a float here, since an integer might overflow */
333 comm_vol *= div_up(a, npme);
334 comm_vol *= div_up(b, npme);
340 /*! \brief Estimate cost of communication for a possible domain decomposition. */
341 static float comm_cost_est(real limit, real cutoff,
342 matrix box, gmx_ddbox_t *ddbox,
343 int natoms, t_inputrec *ir,
345 int npme_tot, ivec nc)
347 ivec npme = {1, 1, 1};
348 int i, j, nk, overlap;
350 float comm_vol, comm_vol_xf, comm_pme, cost_pbcdx;
351 /* This is the cost of a pbc_dx call relative to the cost
352 * of communicating the coordinate and force of an atom.
353 * This will be machine dependent.
354 * These factors are for x86 with SMP or Infiniband.
356 float pbcdx_rect_fac = 0.1;
357 float pbcdx_tric_fac = 0.2;
360 /* Check the DD algorithm restrictions */
361 if ((ir->ePBC == epbcXY && ir->nwall < 2 && nc[ZZ] > 1) ||
362 (ir->ePBC == epbcSCREW && (nc[XX] == 1 || nc[YY] > 1 || nc[ZZ] > 1)))
367 if (inhomogeneous_z(ir) && nc[ZZ] > 1)
372 assert(ddbox->npbcdim <= DIM);
374 /* Check if the triclinic requirements are met */
375 for (i = 0; i < DIM; i++)
377 for (j = i+1; j < ddbox->npbcdim; j++)
379 if (box[j][i] != 0 || ir->deform[j][i] != 0 ||
380 (ir->epc != epcNO && ir->compress[j][i] != 0))
382 if (nc[j] > 1 && nc[i] == 1)
390 for (i = 0; i < DIM; i++)
392 bt[i] = ddbox->box_size[i]*ddbox->skew_fac[i];
394 /* Without PBC and with 2 cells, there are no lower limits on the cell size */
395 if (!(i >= ddbox->npbcdim && nc[i] <= 2) && bt[i] < nc[i]*limit)
399 /* With PBC, check if the cut-off fits in nc[i]-1 cells */
400 if (i < ddbox->npbcdim && nc[i] > 1 && (nc[i] - 1)*bt[i] < nc[i]*cutoff)
408 /* The following choices should match those
409 * in init_domain_decomposition in domdec.c.
411 if (nc[XX] == 1 && nc[YY] > 1)
416 else if (nc[YY] == 1)
423 /* Will we use 1D or 2D PME decomposition? */
424 npme[XX] = (npme_tot % nc[XX] == 0) ? nc[XX] : npme_tot;
425 npme[YY] = npme_tot/npme[XX];
429 /* When two dimensions are (nearly) equal, use more cells
430 * for the smallest index, so the decomposition does not
431 * depend sensitively on the rounding of the box elements.
433 for (i = 0; i < DIM; i++)
435 for (j = i+1; j < DIM; j++)
437 /* Check if the box size is nearly identical,
438 * in that case we prefer nx > ny and ny > nz.
440 if (fabs(bt[j] - bt[i]) < 0.01*bt[i] && nc[j] > nc[i])
442 /* The XX/YY check is a bit compact. If nc[YY]==npme[YY]
443 * this means the swapped nc has nc[XX]==npme[XX],
444 * and we can also swap X and Y for PME.
446 /* Check if dimension i and j are equivalent for PME.
447 * For x/y: if nc[YY]!=npme[YY], we can not swap x/y
448 * For y/z: we can not have PME decomposition in z
451 !((i == XX && j == YY && nc[YY] != npme[YY]) ||
452 (i == YY && j == ZZ && npme[YY] > 1)))
460 /* This function determines only half of the communication cost.
461 * All PP, PME and PP-PME communication is symmetric
462 * and the "back"-communication cost is identical to the forward cost.
465 comm_vol = comm_box_frac(nc, cutoff, ddbox);
468 for (i = 0; i < 2; i++)
470 /* Determine the largest volume for PME x/f redistribution */
471 if (nc[i] % npme[i] != 0)
475 comm_vol_xf = (npme[i] == 2 ? 1.0/3.0 : 0.5);
479 comm_vol_xf = 1.0 - lcd(nc[i], npme[i])/(double)npme[i];
481 comm_pme += 3*natoms*comm_vol_xf;
484 /* Grid overlap communication */
487 nk = (i == 0 ? ir->nkx : ir->nky);
488 overlap = (nk % npme[i] == 0 ? ir->pme_order-1 : ir->pme_order);
496 /* Old line comm_pme += npme[i]*overlap*ir->nkx*ir->nky*ir->nkz/nk; */
500 comm_pme += comm_pme_cost_vol(npme[YY], ir->nky, ir->nkz, ir->nkx);
501 comm_pme += comm_pme_cost_vol(npme[XX], ir->nkx, ir->nky, ir->nkz);
503 /* Add cost of pbc_dx for bondeds */
505 if ((nc[XX] == 1 || nc[YY] == 1) || (nc[ZZ] == 1 && ir->ePBC != epbcXY))
507 if ((ddbox->tric_dir[XX] && nc[XX] == 1) ||
508 (ddbox->tric_dir[YY] && nc[YY] == 1))
510 cost_pbcdx = pbcdxr*pbcdx_tric_fac;
514 cost_pbcdx = pbcdxr*pbcdx_rect_fac;
521 "nc %2d %2d %2d %2d %2d vol pp %6.4f pbcdx %6.4f pme %9.3e tot %9.3e\n",
522 nc[XX], nc[YY], nc[ZZ], npme[XX], npme[YY],
523 comm_vol, cost_pbcdx, comm_pme/(3*natoms),
524 comm_vol + cost_pbcdx + comm_pme/(3*natoms));
527 return 3*natoms*(comm_vol + cost_pbcdx) + comm_pme;
530 /*! \brief Assign penalty factors to possible domain decompositions, based on the estimated communication costs. */
531 static void assign_factors(gmx_domdec_t *dd,
532 real limit, real cutoff,
533 matrix box, gmx_ddbox_t *ddbox,
534 int natoms, t_inputrec *ir,
535 float pbcdxr, int npme,
536 int ndiv, int *div, int *mdiv, ivec ir_try, ivec opt)
543 ce = comm_cost_est(limit, cutoff, box, ddbox,
544 natoms, ir, pbcdxr, npme, ir_try);
545 if (ce >= 0 && (opt[XX] == 0 ||
546 ce < comm_cost_est(limit, cutoff, box, ddbox,
550 copy_ivec(ir_try, opt);
556 for (x = mdiv[0]; x >= 0; x--)
558 for (i = 0; i < x; i++)
560 ir_try[XX] *= div[0];
562 for (y = mdiv[0]-x; y >= 0; y--)
564 for (i = 0; i < y; i++)
566 ir_try[YY] *= div[0];
568 for (i = 0; i < mdiv[0]-x-y; i++)
570 ir_try[ZZ] *= div[0];
574 assign_factors(dd, limit, cutoff, box, ddbox, natoms, ir, pbcdxr, npme,
575 ndiv-1, div+1, mdiv+1, ir_try, opt);
577 for (i = 0; i < mdiv[0]-x-y; i++)
579 ir_try[ZZ] /= div[0];
581 for (i = 0; i < y; i++)
583 ir_try[YY] /= div[0];
586 for (i = 0; i < x; i++)
588 ir_try[XX] /= div[0];
593 /*! \brief Determine the optimal distribution of DD cells for the simulation system and number of MPI ranks */
594 static real optimize_ncells(FILE *fplog,
595 int nnodes_tot, int npme_only,
596 gmx_bool bDynLoadBal, real dlb_scale,
597 gmx_mtop_t *mtop, matrix box, gmx_ddbox_t *ddbox,
600 real cellsize_limit, real cutoff,
601 gmx_bool bInterCGBondeds,
604 int npp, npme, ndiv, *div, *mdiv, d, nmax;
609 limit = cellsize_limit;
615 npp = nnodes_tot - npme_only;
616 if (EEL_PME(ir->coulombtype))
618 npme = (npme_only > 0 ? npme_only : npp);
627 /* If we can skip PBC for distance calculations in plain-C bondeds,
628 * we can save some time (e.g. 3D DD with pbc=xyz).
629 * Here we ignore SIMD bondeds as they always do (fast) PBC.
631 count_bonded_distances(mtop, ir, &pbcdxr, NULL);
632 pbcdxr /= (double)mtop->natoms;
636 /* Every molecule is a single charge group: no pbc required */
639 /* Add a margin for DLB and/or pressure scaling */
642 if (dlb_scale >= 1.0)
644 gmx_fatal(FARGS, "The value for option -dds should be smaller than 1");
648 fprintf(fplog, "Scaling the initial minimum size with 1/%g (option -dds) = %g\n", dlb_scale, 1/dlb_scale);
652 else if (ir->epc != epcNO)
656 fprintf(fplog, "To account for pressure scaling, scaling the initial minimum size with %g\n", DD_GRID_MARGIN_PRES_SCALE);
657 limit *= DD_GRID_MARGIN_PRES_SCALE;
663 fprintf(fplog, "Optimizing the DD grid for %d cells with a minimum initial size of %.3f nm\n", npp, limit);
665 if (inhomogeneous_z(ir))
667 fprintf(fplog, "Ewald_geometry=%s: assuming inhomogeneous particle distribution in z, will not decompose in z.\n", eewg_names[ir->ewald_geometry]);
672 fprintf(fplog, "The maximum allowed number of cells is:");
673 for (d = 0; d < DIM; d++)
675 nmax = (int)(ddbox->box_size[d]*ddbox->skew_fac[d]/limit);
676 if (d >= ddbox->npbcdim && nmax < 2)
680 if (d == ZZ && inhomogeneous_z(ir))
684 fprintf(fplog, " %c %d", 'X' + d, nmax);
686 fprintf(fplog, "\n");
692 fprintf(debug, "Average nr of pbc_dx calls per atom %.2f\n", pbcdxr);
695 /* Decompose npp in factors */
696 ndiv = factorize(npp, &div, &mdiv);
702 assign_factors(dd, limit, cutoff, box, ddbox, mtop->natoms, ir, pbcdxr,
703 npme, ndiv, div, mdiv, itry, nc);
711 real dd_choose_grid(FILE *fplog,
712 t_commrec *cr, gmx_domdec_t *dd, t_inputrec *ir,
713 gmx_mtop_t *mtop, matrix box, gmx_ddbox_t *ddbox,
714 gmx_bool bDynLoadBal, real dlb_scale,
715 real cellsize_limit, real cutoff_dd,
716 gmx_bool bInterCGBondeds)
718 gmx_int64_t nnodes_div, ldiv;
723 nnodes_div = cr->nnodes;
724 if (EEL_PME(ir->coulombtype))
726 if (cr->npmenodes > 0)
728 if (cr->npmenodes >= cr->nnodes)
731 "Cannot have %d separate PME ranks with just %d total ranks",
732 cr->npmenodes, cr->nnodes);
735 /* If the user purposely selected the number of PME nodes,
736 * only check for large primes in the PP node count.
738 nnodes_div -= cr->npmenodes;
748 ldiv = largest_divisor(nnodes_div);
749 /* Check if the largest divisor is more than nnodes^2/3 */
750 if (ldiv*ldiv*ldiv > nnodes_div*nnodes_div)
752 gmx_fatal(FARGS, "The number of ranks 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.",
757 if (EEL_PME(ir->coulombtype))
759 if (cr->npmenodes < 0)
761 /* Use PME nodes when the number of nodes is more than 16 */
762 if (cr->nnodes <= 18)
767 fprintf(fplog, "Using %d separate PME ranks, as there are too few total\n ranks for efficient splitting\n", cr->npmenodes);
772 cr->npmenodes = guess_npme(fplog, mtop, ir, box, cr->nnodes);
775 fprintf(fplog, "Using %d separate PME ranks, as guessed by mdrun\n", cr->npmenodes);
783 fprintf(fplog, "Using %d separate PME ranks, per user request\n", cr->npmenodes);
788 limit = optimize_ncells(fplog, cr->nnodes, cr->npmenodes,
789 bDynLoadBal, dlb_scale,
790 mtop, box, ddbox, ir, dd,
791 cellsize_limit, cutoff_dd,
799 /* Communicate the information set by the master to all nodes */
800 gmx_bcast(sizeof(dd->nc), dd->nc, cr);
801 if (EEL_PME(ir->coulombtype))
803 gmx_bcast(sizeof(ir->nkx), &ir->nkx, cr);
804 gmx_bcast(sizeof(ir->nky), &ir->nky, cr);
805 gmx_bcast(sizeof(cr->npmenodes), &cr->npmenodes, cr);