2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2008,2009,2010,2011,2012,2013,2014,2015,2017,2018, 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
51 #include "gromacs/domdec/domdec.h"
52 #include "gromacs/domdec/domdec_struct.h"
53 #include "gromacs/ewald/pme.h"
54 #include "gromacs/gmxlib/network.h"
55 #include "gromacs/math/utilities.h"
56 #include "gromacs/math/vec.h"
57 #include "gromacs/mdlib/perf_est.h"
58 #include "gromacs/mdtypes/commrec.h"
59 #include "gromacs/mdtypes/inputrec.h"
60 #include "gromacs/mdtypes/md_enums.h"
61 #include "gromacs/pbcutil/pbc.h"
62 #include "gromacs/topology/topology.h"
63 #include "gromacs/utility/fatalerror.h"
64 #include "gromacs/utility/logger.h"
65 #include "gromacs/utility/stringutil.h"
67 /*! \brief Margin for setting up the DD grid */
68 #define DD_GRID_MARGIN_PRES_SCALE 1.05
70 /*! \brief Factorize \p n.
72 * \param[in] n Value to factorize
73 * \param[out] fac Vector of factors (to be allocated in this function)
74 * \param[out] mfac Vector with the number of times each factor repeats in the factorization (to be allocated in this function)
76 static void factorize(int n,
77 std::vector<int> *fac,
78 std::vector<int> *mfac)
82 gmx_fatal(FARGS, "Can only factorize positive integers.");
85 /* Decompose n in factors */
93 if (fac->empty() || fac->back() != d)
108 /*! \brief Find largest divisor of \p n smaller than \p n*/
109 static int largest_divisor(int n)
111 std::vector<int> div;
112 std::vector<int> mdiv;
113 factorize(n, &div, &mdiv);
118 /*! \brief Compute largest common divisor of \p n1 and \b n2 */
119 static int lcd(int n1, int n2)
124 for (i = 2; (i <= n1 && i <= n2); i++)
126 if (n1 % i == 0 && n2 % i == 0)
135 /*! \brief Returns TRUE when there are enough PME ranks for the ratio */
136 static gmx_bool fits_pme_ratio(int nrank_tot, int nrank_pme, float ratio)
138 return (static_cast<double>(nrank_pme)/static_cast<double>(nrank_tot) > 0.95*ratio);
141 /*! \brief Returns TRUE when npme out of ntot ranks doing PME is expected to give reasonable performance */
142 static gmx_bool fits_pp_pme_perf(int ntot, int npme, float ratio)
144 std::vector<int> div;
145 std::vector<int> mdiv;
146 factorize(ntot - npme, &div, &mdiv);
148 int npp_root3 = gmx::roundToInt(std::cbrt(ntot - npme));
149 int npme_root2 = gmx::roundToInt(std::sqrt(static_cast<double>(npme)));
151 /* The check below gives a reasonable division:
152 * factor 5 allowed at 5 or more PP ranks,
153 * factor 7 allowed at 49 or more PP ranks.
155 if (div.back() > 3 + npp_root3)
160 /* Check if the number of PP and PME ranks have a reasonable sized
161 * denominator in common, such that we can use 2D PME decomposition
162 * when required (which requires nx_pp == nx_pme).
163 * The factor of 2 allows for a maximum ratio of 2^2=4
164 * between nx_pme and ny_pme.
166 if (lcd(ntot - npme, npme)*2 < npme_root2)
171 /* Does this division gives a reasonable PME load? */
172 return fits_pme_ratio(ntot, npme, ratio);
175 /*! \brief Make a guess for the number of PME ranks to use. */
176 static int guess_npme(const gmx::MDLogger &mdlog,
177 const gmx_mtop_t *mtop, const t_inputrec *ir,
184 ratio = pme_load_estimate(mtop, ir, box);
186 GMX_LOG(mdlog.info).appendTextFormatted(
187 "Guess for relative PME load: %.2f", ratio);
189 /* We assume the optimal rank ratio is close to the load ratio.
190 * The communication load is neglected,
191 * but (hopefully) this will balance out between PP and PME.
194 if (!fits_pme_ratio(nrank_tot, nrank_tot/2, ratio))
196 /* We would need more than nrank_tot/2 PME only nodes,
197 * which is not possible. Since the PME load is very high,
198 * we will not loose much performance when all ranks do PME.
204 /* First try to find npme as a factor of nrank_tot up to nrank_tot/3.
205 * We start with a minimum PME node fraction of 1/16
206 * and avoid ratios which lead to large prime factors in nnodes-npme.
208 npme = (nrank_tot + 15)/16;
209 while (npme <= nrank_tot/3)
211 if (nrank_tot % npme == 0)
213 /* Note that fits_perf might change the PME grid,
214 * in the current implementation it does not.
216 if (fits_pp_pme_perf(nrank_tot, npme, ratio))
223 if (npme > nrank_tot/3)
225 /* Try any possible number for npme */
227 while (npme <= nrank_tot/2)
229 /* Note that fits_perf may change the PME grid */
230 if (fits_pp_pme_perf(nrank_tot, npme, ratio))
237 if (npme > nrank_tot/2)
239 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"
240 "Use the -npme option of mdrun or change the number of ranks or the PME grid dimensions, see the manual for details.",
241 ratio, gmx::roundToInt(0.95*ratio*nrank_tot), nrank_tot/2, ir->nkx, ir->nky);
245 GMX_LOG(mdlog.info).appendTextFormatted(
246 "Will use %d particle-particle and %d PME only ranks\n"
247 "This is a guess, check the performance at the end of the log file",
248 nrank_tot - npme, npme);
254 /*! \brief Return \p n divided by \p f rounded up to the next integer. */
255 static int div_up(int n, int f)
257 return (n + f - 1)/f;
260 real comm_box_frac(const ivec dd_nc, real cutoff, const gmx_ddbox_t *ddbox)
266 for (i = 0; i < DIM; i++)
268 real bt = ddbox->box_size[i]*ddbox->skew_fac[i];
269 nw[i] = dd_nc[i]*cutoff/bt;
273 for (i = 0; i < DIM; i++)
278 for (j = i+1; j < DIM; j++)
282 comm_vol += nw[i]*nw[j]*M_PI/4;
283 for (k = j+1; k < DIM; k++)
287 comm_vol += nw[i]*nw[j]*nw[k]*M_PI/6;
298 /*! \brief Return whether the DD inhomogeneous in the z direction */
299 static gmx_bool inhomogeneous_z(const t_inputrec *ir)
301 return ((EEL_PME(ir->coulombtype) || ir->coulombtype == eelEWALD) &&
302 ir->ePBC == epbcXYZ && ir->ewald_geometry == eewg3DC);
305 /*! \brief Estimate cost of PME FFT communication
307 * This only takes the communication into account and not imbalance
308 * in the calculation. But the imbalance in communication and calculation
309 * are similar and therefore these formulas also prefer load balance
310 * in the FFT and pme_solve calculation.
312 static float comm_pme_cost_vol(int npme, int a, int b, int c)
314 /* We use a float here, since an integer might overflow */
319 comm_vol *= div_up(a, npme);
320 comm_vol *= div_up(b, npme);
326 /*! \brief Estimate cost of communication for a possible domain decomposition. */
327 static float comm_cost_est(real limit, real cutoff,
328 const matrix box, const gmx_ddbox_t *ddbox,
329 int natoms, const t_inputrec *ir,
331 int npme_tot, ivec nc)
333 ivec npme = {1, 1, 1};
334 int i, j, nk, overlap;
336 float comm_vol, comm_vol_xf, comm_pme, cost_pbcdx;
337 /* This is the cost of a pbc_dx call relative to the cost
338 * of communicating the coordinate and force of an atom.
339 * This will be machine dependent.
340 * These factors are for x86 with SMP or Infiniband.
342 float pbcdx_rect_fac = 0.1;
343 float pbcdx_tric_fac = 0.2;
346 /* Check the DD algorithm restrictions */
347 if ((ir->ePBC == epbcXY && ir->nwall < 2 && nc[ZZ] > 1) ||
348 (ir->ePBC == epbcSCREW && (nc[XX] == 1 || nc[YY] > 1 || nc[ZZ] > 1)))
353 if (inhomogeneous_z(ir) && nc[ZZ] > 1)
358 assert(ddbox->npbcdim <= DIM);
360 /* Check if the triclinic requirements are met */
361 for (i = 0; i < DIM; i++)
363 for (j = i+1; j < ddbox->npbcdim; j++)
365 if (box[j][i] != 0 || ir->deform[j][i] != 0 ||
366 (ir->epc != epcNO && ir->compress[j][i] != 0))
368 if (nc[j] > 1 && nc[i] == 1)
376 for (i = 0; i < DIM; i++)
378 bt[i] = ddbox->box_size[i]*ddbox->skew_fac[i];
380 /* Without PBC and with 2 cells, there are no lower limits on the cell size */
381 if (!(i >= ddbox->npbcdim && nc[i] <= 2) && bt[i] < nc[i]*limit)
385 /* With PBC, check if the cut-off fits in nc[i]-1 cells */
386 if (i < ddbox->npbcdim && nc[i] > 1 && (nc[i] - 1)*bt[i] < nc[i]*cutoff)
394 /* The following choices should match those
395 * in init_domain_decomposition in domdec.c.
397 if (nc[XX] == 1 && nc[YY] > 1)
402 else if (nc[YY] == 1)
409 /* Will we use 1D or 2D PME decomposition? */
410 npme[XX] = (npme_tot % nc[XX] == 0) ? nc[XX] : npme_tot;
411 npme[YY] = npme_tot/npme[XX];
415 if (EEL_PME(ir->coulombtype) || EVDW_PME(ir->vdwtype))
417 /* Check the PME grid restrictions.
418 * Currently these can only be invalid here with too few grid lines
419 * along the x dimension per rank doing PME.
421 int npme_x = (npme_tot > 1 ? npme[XX] : nc[XX]);
423 /* Currently we don't have the OpenMP thread count available here.
424 * But with threads we have only tighter restrictions and it's
425 * probably better anyhow to avoid settings where we need to reduce
426 * grid lines over multiple ranks, as the thread check will do.
428 bool useThreads = true;
429 bool errorsAreFatal = false;
430 if (!gmx_pme_check_restrictions(ir->pme_order, ir->nkx, ir->nky, ir->nkz,
431 npme_x, useThreads, errorsAreFatal))
437 /* When two dimensions are (nearly) equal, use more cells
438 * for the smallest index, so the decomposition does not
439 * depend sensitively on the rounding of the box elements.
441 for (i = 0; i < DIM; i++)
443 for (j = i+1; j < DIM; j++)
445 /* Check if the box size is nearly identical,
446 * in that case we prefer nx > ny and ny > nz.
448 if (std::fabs(bt[j] - bt[i]) < 0.01*bt[i] && nc[j] > nc[i])
450 /* The XX/YY check is a bit compact. If nc[YY]==npme[YY]
451 * this means the swapped nc has nc[XX]==npme[XX],
452 * and we can also swap X and Y for PME.
454 /* Check if dimension i and j are equivalent for PME.
455 * For x/y: if nc[YY]!=npme[YY], we can not swap x/y
456 * For y/z: we can not have PME decomposition in z
459 !((i == XX && j == YY && nc[YY] != npme[YY]) ||
460 (i == YY && j == ZZ && npme[YY] > 1)))
468 /* This function determines only half of the communication cost.
469 * All PP, PME and PP-PME communication is symmetric
470 * and the "back"-communication cost is identical to the forward cost.
473 comm_vol = comm_box_frac(nc, cutoff, ddbox);
476 for (i = 0; i < 2; i++)
478 /* Determine the largest volume for PME x/f redistribution */
479 if (nc[i] % npme[i] != 0)
483 comm_vol_xf = (npme[i] == 2 ? 1.0/3.0 : 0.5);
487 comm_vol_xf = 1.0 - lcd(nc[i], npme[i])/static_cast<double>(npme[i]);
489 comm_pme += 3*natoms*comm_vol_xf;
492 /* Grid overlap communication */
495 nk = (i == 0 ? ir->nkx : ir->nky);
496 overlap = (nk % npme[i] == 0 ? ir->pme_order-1 : ir->pme_order);
504 /* Old line comm_pme += npme[i]*overlap*ir->nkx*ir->nky*ir->nkz/nk; */
508 comm_pme += comm_pme_cost_vol(npme[YY], ir->nky, ir->nkz, ir->nkx);
509 comm_pme += comm_pme_cost_vol(npme[XX], ir->nkx, ir->nky, ir->nkz);
511 /* Add cost of pbc_dx for bondeds */
513 if ((nc[XX] == 1 || nc[YY] == 1) || (nc[ZZ] == 1 && ir->ePBC != epbcXY))
515 if ((ddbox->tric_dir[XX] && nc[XX] == 1) ||
516 (ddbox->tric_dir[YY] && nc[YY] == 1))
518 cost_pbcdx = pbcdxr*pbcdx_tric_fac;
522 cost_pbcdx = pbcdxr*pbcdx_rect_fac;
529 "nc %2d %2d %2d %2d %2d vol pp %6.4f pbcdx %6.4f pme %9.3e tot %9.3e\n",
530 nc[XX], nc[YY], nc[ZZ], npme[XX], npme[YY],
531 comm_vol, cost_pbcdx, comm_pme/(3*natoms),
532 comm_vol + cost_pbcdx + comm_pme/(3*natoms));
535 return 3*natoms*(comm_vol + cost_pbcdx) + comm_pme;
538 /*! \brief Assign penalty factors to possible domain decompositions, based on the estimated communication costs. */
539 static void assign_factors(const gmx_domdec_t *dd,
540 real limit, real cutoff,
541 const matrix box, const gmx_ddbox_t *ddbox,
542 int natoms, const t_inputrec *ir,
543 float pbcdxr, int npme,
544 int ndiv, const int *div, const int *mdiv,
545 ivec ir_try, ivec opt)
552 ce = comm_cost_est(limit, cutoff, box, ddbox,
553 natoms, ir, pbcdxr, npme, ir_try);
554 if (ce >= 0 && (opt[XX] == 0 ||
555 ce < comm_cost_est(limit, cutoff, box, ddbox,
559 copy_ivec(ir_try, opt);
565 for (x = mdiv[0]; x >= 0; x--)
567 for (i = 0; i < x; i++)
569 ir_try[XX] *= div[0];
571 for (y = mdiv[0]-x; y >= 0; y--)
573 for (i = 0; i < y; i++)
575 ir_try[YY] *= div[0];
577 for (i = 0; i < mdiv[0]-x-y; i++)
579 ir_try[ZZ] *= div[0];
583 assign_factors(dd, limit, cutoff, box, ddbox, natoms, ir, pbcdxr, npme,
584 ndiv-1, div+1, mdiv+1, ir_try, opt);
586 for (i = 0; i < mdiv[0]-x-y; i++)
588 ir_try[ZZ] /= div[0];
590 for (i = 0; i < y; i++)
592 ir_try[YY] /= div[0];
595 for (i = 0; i < x; i++)
597 ir_try[XX] /= div[0];
602 /*! \brief Determine the optimal distribution of DD cells for the simulation system and number of MPI ranks */
603 static real optimize_ncells(const gmx::MDLogger &mdlog,
604 int nnodes_tot, int npme_only,
605 gmx_bool bDynLoadBal, real dlb_scale,
606 const gmx_mtop_t *mtop,
607 const matrix box, const gmx_ddbox_t *ddbox,
608 const t_inputrec *ir,
610 real cellsize_limit, real cutoff,
611 gmx_bool bInterCGBondeds,
614 int npp, npme, d, nmax;
619 limit = cellsize_limit;
625 npp = nnodes_tot - npme_only;
626 if (EEL_PME(ir->coulombtype))
628 npme = (npme_only > 0 ? npme_only : npp);
637 /* If we can skip PBC for distance calculations in plain-C bondeds,
638 * we can save some time (e.g. 3D DD with pbc=xyz).
639 * Here we ignore SIMD bondeds as they always do (fast) PBC.
641 count_bonded_distances(mtop, ir, &pbcdxr, nullptr);
642 pbcdxr /= static_cast<double>(mtop->natoms);
646 /* Every molecule is a single charge group: no pbc required */
649 /* Add a margin for DLB and/or pressure scaling */
652 if (dlb_scale >= 1.0)
654 gmx_fatal(FARGS, "The value for option -dds should be smaller than 1");
656 GMX_LOG(mdlog.info).appendTextFormatted(
657 "Scaling the initial minimum size with 1/%g (option -dds) = %g",
658 dlb_scale, 1/dlb_scale);
661 else if (ir->epc != epcNO)
663 GMX_LOG(mdlog.info).appendTextFormatted(
664 "To account for pressure scaling, scaling the initial minimum size with %g",
665 DD_GRID_MARGIN_PRES_SCALE);
666 limit *= DD_GRID_MARGIN_PRES_SCALE;
669 GMX_LOG(mdlog.info).appendTextFormatted(
670 "Optimizing the DD grid for %d cells with a minimum initial size of %.3f nm",
673 if (inhomogeneous_z(ir))
675 GMX_LOG(mdlog.info).appendTextFormatted(
676 "Ewald_geometry=%s: assuming inhomogeneous particle distribution in z, will not decompose in z.",
677 eewg_names[ir->ewald_geometry]);
682 std::string maximumCells = "The maximum allowed number of cells is:";
683 for (d = 0; d < DIM; d++)
685 nmax = static_cast<int>(ddbox->box_size[d]*ddbox->skew_fac[d]/limit);
686 if (d >= ddbox->npbcdim && nmax < 2)
690 if (d == ZZ && inhomogeneous_z(ir))
694 maximumCells += gmx::formatString(" %c %d", 'X' + d, nmax);
696 GMX_LOG(mdlog.info).appendText(maximumCells);
701 fprintf(debug, "Average nr of pbc_dx calls per atom %.2f\n", pbcdxr);
704 /* Decompose npp in factors */
705 std::vector<int> div;
706 std::vector<int> mdiv;
707 factorize(npp, &div, &mdiv);
713 assign_factors(dd, limit, cutoff, box, ddbox, mtop->natoms, ir, pbcdxr,
714 npme, div.size(), div.data(), mdiv.data(), itry, nc);
719 real dd_choose_grid(const gmx::MDLogger &mdlog,
720 t_commrec *cr, gmx_domdec_t *dd,
721 const t_inputrec *ir,
722 const gmx_mtop_t *mtop,
723 const matrix box, const gmx_ddbox_t *ddbox,
725 gmx_bool bDynLoadBal, real dlb_scale,
726 real cellsize_limit, real cutoff_dd,
727 gmx_bool bInterCGBondeds)
729 int64_t nnodes_div, ldiv;
734 nnodes_div = cr->nnodes;
735 if (EEL_PME(ir->coulombtype))
739 if (nPmeRanks >= cr->nnodes)
742 "Cannot have %d separate PME ranks with just %d total ranks",
743 nPmeRanks, cr->nnodes);
746 /* If the user purposely selected the number of PME nodes,
747 * only check for large primes in the PP node count.
749 nnodes_div -= nPmeRanks;
759 ldiv = largest_divisor(nnodes_div);
760 /* Check if the largest divisor is more than nnodes^2/3 */
761 if (ldiv*ldiv*ldiv > nnodes_div*nnodes_div)
763 gmx_fatal(FARGS, "The number of ranks you selected (%ld) contains a large prime factor %ld. In most cases this will lead to bad performance. Choose a number with smaller prime factors or set the decomposition (option -dd) manually.",
768 if (EEL_PME(ir->coulombtype))
772 /* Use PME nodes when the number of nodes is more than 16 */
773 if (cr->nnodes <= 18)
776 GMX_LOG(mdlog.info).appendTextFormatted(
777 "Using %d separate PME ranks, as there are too few total\n"
778 " ranks for efficient splitting",
783 cr->npmenodes = guess_npme(mdlog, mtop, ir, box, cr->nnodes);
784 GMX_LOG(mdlog.info).appendTextFormatted(
785 "Using %d separate PME ranks, as guessed by mdrun", cr->npmenodes);
790 /* We checked above that nPmeRanks is a valid number */
791 cr->npmenodes = nPmeRanks;
792 GMX_LOG(mdlog.info).appendTextFormatted(
793 "Using %d separate PME ranks", cr->npmenodes);
794 // TODO: there was a ", per user request" note here, but it's not correct anymore,
795 // as with GPUs decision about nPmeRanks can be made in runner() as well.
796 // Consider a single spot for setting nPmeRanks.
800 limit = optimize_ncells(mdlog, cr->nnodes, cr->npmenodes,
801 bDynLoadBal, dlb_scale,
802 mtop, box, ddbox, ir, dd,
803 cellsize_limit, cutoff_dd,
811 /* Communicate the information set by the master to all nodes */
812 gmx_bcast(sizeof(dd->nc), dd->nc, cr);
813 if (EEL_PME(ir->coulombtype))
815 gmx_bcast(sizeof(cr->npmenodes), &cr->npmenodes, cr);