2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2005,2006,2007,2008,2009,2010,2011,2012,2013,2014,2015,2016,2017,2018,2019, 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.
53 #include "gromacs/domdec/collect.h"
54 #include "gromacs/domdec/dlb.h"
55 #include "gromacs/domdec/dlbtiming.h"
56 #include "gromacs/domdec/domdec_network.h"
57 #include "gromacs/domdec/ga2la.h"
58 #include "gromacs/domdec/gpuhaloexchange.h"
59 #include "gromacs/domdec/options.h"
60 #include "gromacs/domdec/partition.h"
61 #include "gromacs/gmxlib/network.h"
62 #include "gromacs/gmxlib/nrnb.h"
63 #include "gromacs/gpu_utils/gpu_utils.h"
64 #include "gromacs/hardware/hw_info.h"
65 #include "gromacs/listed_forces/manage_threading.h"
66 #include "gromacs/math/vec.h"
67 #include "gromacs/math/vectypes.h"
68 #include "gromacs/mdlib/calc_verletbuf.h"
69 #include "gromacs/mdlib/constr.h"
70 #include "gromacs/mdlib/constraintrange.h"
71 #include "gromacs/mdlib/updategroups.h"
72 #include "gromacs/mdlib/vsite.h"
73 #include "gromacs/mdtypes/commrec.h"
74 #include "gromacs/mdtypes/forceoutput.h"
75 #include "gromacs/mdtypes/inputrec.h"
76 #include "gromacs/mdtypes/mdrunoptions.h"
77 #include "gromacs/mdtypes/state.h"
78 #include "gromacs/pbcutil/ishift.h"
79 #include "gromacs/pbcutil/pbc.h"
80 #include "gromacs/pulling/pull.h"
81 #include "gromacs/timing/wallcycle.h"
82 #include "gromacs/topology/block.h"
83 #include "gromacs/topology/idef.h"
84 #include "gromacs/topology/ifunc.h"
85 #include "gromacs/topology/mtop_lookup.h"
86 #include "gromacs/topology/mtop_util.h"
87 #include "gromacs/topology/topology.h"
88 #include "gromacs/utility/basedefinitions.h"
89 #include "gromacs/utility/basenetwork.h"
90 #include "gromacs/utility/cstringutil.h"
91 #include "gromacs/utility/exceptions.h"
92 #include "gromacs/utility/fatalerror.h"
93 #include "gromacs/utility/gmxmpi.h"
94 #include "gromacs/utility/logger.h"
95 #include "gromacs/utility/real.h"
96 #include "gromacs/utility/smalloc.h"
97 #include "gromacs/utility/strconvert.h"
98 #include "gromacs/utility/stringstream.h"
99 #include "gromacs/utility/stringutil.h"
100 #include "gromacs/utility/textwriter.h"
102 #include "atomdistribution.h"
104 #include "cellsizes.h"
105 #include "distribute.h"
106 #include "domdec_constraints.h"
107 #include "domdec_internal.h"
108 #include "domdec_setup.h"
109 #include "domdec_vsite.h"
110 #include "redistribute.h"
113 // TODO remove this when moving domdec into gmx namespace
114 using gmx::DdRankOrder;
115 using gmx::DlbOption;
116 using gmx::DomdecOptions;
118 static const char *edlbs_names[int(DlbState::nr)] = { "off", "auto", "locked", "on", "on" };
120 /* The size per atom group of the cggl_flag buffer in gmx_domdec_comm_t */
123 /* The flags for the cggl_flag buffer in gmx_domdec_comm_t */
124 #define DD_FLAG_NRCG 65535
125 #define DD_FLAG_FW(d) (1<<(16+(d)*2))
126 #define DD_FLAG_BW(d) (1<<(16+(d)*2+1))
128 /* The DD zone order */
129 static const ivec dd_zo[DD_MAXZONE] =
130 {{0, 0, 0}, {1, 0, 0}, {1, 1, 0}, {0, 1, 0}, {0, 1, 1}, {0, 0, 1}, {1, 0, 1}, {1, 1, 1}};
132 /* The non-bonded zone-pair setup for domain decomposition
133 * The first number is the i-zone, the second number the first j-zone seen by
134 * this i-zone, the third number the last+1 j-zone seen by this i-zone.
135 * As is, this is for 3D decomposition, where there are 4 i-zones.
136 * With 2D decomposition use only the first 2 i-zones and a last+1 j-zone of 4.
137 * With 1D decomposition use only the first i-zone and a last+1 j-zone of 2.
140 ddNonbondedZonePairRanges[DD_MAXIZONE][3] = {{0, 0, 8},
147 #define dd_index(n,i) ((((i)[ZZ]*(n)[YY] + (i)[YY])*(n)[XX]) + (i)[XX])
149 static void index2xyz(ivec nc,int ind,ivec xyz)
151 xyz[XX] = ind % nc[XX];
152 xyz[YY] = (ind / nc[XX]) % nc[YY];
153 xyz[ZZ] = ind / (nc[YY]*nc[XX]);
157 static void ddindex2xyz(const ivec nc, int ind, ivec xyz)
159 xyz[XX] = ind / (nc[YY]*nc[ZZ]);
160 xyz[YY] = (ind / nc[ZZ]) % nc[YY];
161 xyz[ZZ] = ind % nc[ZZ];
164 static int ddcoord2ddnodeid(gmx_domdec_t *dd, ivec c)
168 const CartesianRankSetup &cartSetup = dd->comm->cartesianRankSetup;
169 const int ddindex = dd_index(dd->nc, c);
170 if (cartSetup.bCartesianPP_PME)
172 ddnodeid = cartSetup.ddindex2ddnodeid[ddindex];
174 else if (cartSetup.bCartesianPP)
177 MPI_Cart_rank(dd->mpi_comm_all, c, &ddnodeid);
188 int ddglatnr(const gmx_domdec_t *dd, int i)
198 if (i >= dd->comm->atomRanges.numAtomsTotal())
200 gmx_fatal(FARGS, "glatnr called with %d, which is larger than the local number of atoms (%d)", i, dd->comm->atomRanges.numAtomsTotal());
202 atnr = dd->globalAtomIndices[i] + 1;
208 t_block *dd_charge_groups_global(gmx_domdec_t *dd)
210 return &dd->comm->cgs_gl;
213 gmx::ArrayRef<const gmx::RangePartitioning> getUpdateGroupingPerMoleculetype(const gmx_domdec_t &dd)
215 GMX_RELEASE_ASSERT(dd.comm, "Need a valid dd.comm");
216 return dd.comm->systemInfo.updateGroupingPerMoleculetype;
219 void dd_store_state(gmx_domdec_t *dd, t_state *state)
223 if (state->ddp_count != dd->ddp_count)
225 gmx_incons("The MD state does not match the domain decomposition state");
228 state->cg_gl.resize(dd->ncg_home);
229 for (i = 0; i < dd->ncg_home; i++)
231 state->cg_gl[i] = dd->globalAtomGroupIndices[i];
234 state->ddp_count_cg_gl = dd->ddp_count;
237 gmx_domdec_zones_t *domdec_zones(gmx_domdec_t *dd)
239 return &dd->comm->zones;
242 void dd_get_ns_ranges(const gmx_domdec_t *dd, int icg,
243 int *jcg0, int *jcg1, ivec shift0, ivec shift1)
245 gmx_domdec_zones_t *zones;
248 zones = &dd->comm->zones;
251 while (icg >= zones->izone[izone].cg1)
260 else if (izone < zones->nizone)
262 *jcg0 = zones->izone[izone].jcg0;
266 gmx_fatal(FARGS, "DD icg %d out of range: izone (%d) >= nizone (%d)",
267 icg, izone, zones->nizone);
270 *jcg1 = zones->izone[izone].jcg1;
272 for (d = 0; d < dd->ndim; d++)
275 shift0[dim] = zones->izone[izone].shift0[dim];
276 shift1[dim] = zones->izone[izone].shift1[dim];
277 if (dd->comm->tric_dir[dim] || (isDlbOn(dd->comm) && d > 0))
279 /* A conservative approach, this can be optimized */
286 int dd_numHomeAtoms(const gmx_domdec_t &dd)
288 return dd.comm->atomRanges.numHomeAtoms();
291 int dd_natoms_mdatoms(const gmx_domdec_t *dd)
293 /* We currently set mdatoms entries for all atoms:
294 * local + non-local + communicated for vsite + constraints
297 return dd->comm->atomRanges.numAtomsTotal();
300 int dd_natoms_vsite(const gmx_domdec_t *dd)
302 return dd->comm->atomRanges.end(DDAtomRanges::Type::Vsites);
305 void dd_get_constraint_range(const gmx_domdec_t *dd, int *at_start, int *at_end)
307 *at_start = dd->comm->atomRanges.start(DDAtomRanges::Type::Constraints);
308 *at_end = dd->comm->atomRanges.end(DDAtomRanges::Type::Constraints);
311 void dd_move_x(gmx_domdec_t *dd,
313 gmx::ArrayRef<gmx::RVec> x,
314 gmx_wallcycle *wcycle)
316 wallcycle_start(wcycle, ewcMOVEX);
319 gmx_domdec_comm_t *comm;
320 gmx_domdec_comm_dim_t *cd;
321 rvec shift = {0, 0, 0};
322 gmx_bool bPBC, bScrew;
327 nat_tot = comm->atomRanges.numHomeAtoms();
328 for (int d = 0; d < dd->ndim; d++)
330 bPBC = (dd->ci[dd->dim[d]] == 0);
331 bScrew = (bPBC && dd->unitCellInfo.haveScrewPBC && dd->dim[d] == XX);
334 copy_rvec(box[dd->dim[d]], shift);
337 for (const gmx_domdec_ind_t &ind : cd->ind)
339 DDBufferAccess<gmx::RVec> sendBufferAccess(comm->rvecBuffer, ind.nsend[nzone + 1]);
340 gmx::ArrayRef<gmx::RVec> &sendBuffer = sendBufferAccess.buffer;
344 for (int j : ind.index)
346 sendBuffer[n] = x[j];
352 for (int j : ind.index)
354 /* We need to shift the coordinates */
355 for (int d = 0; d < DIM; d++)
357 sendBuffer[n][d] = x[j][d] + shift[d];
364 for (int j : ind.index)
367 sendBuffer[n][XX] = x[j][XX] + shift[XX];
369 * This operation requires a special shift force
370 * treatment, which is performed in calc_vir.
372 sendBuffer[n][YY] = box[YY][YY] - x[j][YY];
373 sendBuffer[n][ZZ] = box[ZZ][ZZ] - x[j][ZZ];
378 DDBufferAccess<gmx::RVec> receiveBufferAccess(comm->rvecBuffer2, cd->receiveInPlace ? 0 : ind.nrecv[nzone + 1]);
380 gmx::ArrayRef<gmx::RVec> receiveBuffer;
381 if (cd->receiveInPlace)
383 receiveBuffer = gmx::arrayRefFromArray(x.data() + nat_tot, ind.nrecv[nzone + 1]);
387 receiveBuffer = receiveBufferAccess.buffer;
389 /* Send and receive the coordinates */
390 ddSendrecv(dd, d, dddirBackward,
391 sendBuffer, receiveBuffer);
393 if (!cd->receiveInPlace)
396 for (int zone = 0; zone < nzone; zone++)
398 for (int i = ind.cell2at0[zone]; i < ind.cell2at1[zone]; i++)
400 x[i] = receiveBuffer[j++];
404 nat_tot += ind.nrecv[nzone+1];
409 wallcycle_stop(wcycle, ewcMOVEX);
412 void dd_move_f(gmx_domdec_t *dd,
413 gmx::ForceWithShiftForces *forceWithShiftForces,
414 gmx_wallcycle *wcycle)
416 wallcycle_start(wcycle, ewcMOVEF);
418 gmx::ArrayRef<gmx::RVec> f = forceWithShiftForces->force();
419 gmx::ArrayRef<gmx::RVec> fshift = forceWithShiftForces->shiftForces();
421 gmx_domdec_comm_t &comm = *dd->comm;
422 int nzone = comm.zones.n/2;
423 int nat_tot = comm.atomRanges.end(DDAtomRanges::Type::Zones);
424 for (int d = dd->ndim-1; d >= 0; d--)
426 /* Only forces in domains near the PBC boundaries need to
427 consider PBC in the treatment of fshift */
428 const bool shiftForcesNeedPbc = (forceWithShiftForces->computeVirial() && dd->ci[dd->dim[d]] == 0);
429 const bool applyScrewPbc = (shiftForcesNeedPbc && dd->unitCellInfo.haveScrewPBC && dd->dim[d] == XX);
430 /* Determine which shift vector we need */
431 ivec vis = { 0, 0, 0 };
433 const int is = IVEC2IS(vis);
435 /* Loop over the pulses */
436 const gmx_domdec_comm_dim_t &cd = comm.cd[d];
437 for (int p = cd.numPulses() - 1; p >= 0; p--)
439 const gmx_domdec_ind_t &ind = cd.ind[p];
440 DDBufferAccess<gmx::RVec> receiveBufferAccess(comm.rvecBuffer, ind.nsend[nzone + 1]);
441 gmx::ArrayRef<gmx::RVec> &receiveBuffer = receiveBufferAccess.buffer;
443 nat_tot -= ind.nrecv[nzone+1];
445 DDBufferAccess<gmx::RVec> sendBufferAccess(comm.rvecBuffer2, cd.receiveInPlace ? 0 : ind.nrecv[nzone + 1]);
447 gmx::ArrayRef<gmx::RVec> sendBuffer;
448 if (cd.receiveInPlace)
450 sendBuffer = gmx::arrayRefFromArray(f.data() + nat_tot, ind.nrecv[nzone + 1]);
454 sendBuffer = sendBufferAccess.buffer;
456 for (int zone = 0; zone < nzone; zone++)
458 for (int i = ind.cell2at0[zone]; i < ind.cell2at1[zone]; i++)
460 sendBuffer[j++] = f[i];
464 /* Communicate the forces */
465 ddSendrecv(dd, d, dddirForward,
466 sendBuffer, receiveBuffer);
467 /* Add the received forces */
469 if (!shiftForcesNeedPbc)
471 for (int j : ind.index)
473 for (int d = 0; d < DIM; d++)
475 f[j][d] += receiveBuffer[n][d];
480 else if (!applyScrewPbc)
482 for (int j : ind.index)
484 for (int d = 0; d < DIM; d++)
486 f[j][d] += receiveBuffer[n][d];
488 /* Add this force to the shift force */
489 for (int d = 0; d < DIM; d++)
491 fshift[is][d] += receiveBuffer[n][d];
498 for (int j : ind.index)
500 /* Rotate the force */
501 f[j][XX] += receiveBuffer[n][XX];
502 f[j][YY] -= receiveBuffer[n][YY];
503 f[j][ZZ] -= receiveBuffer[n][ZZ];
504 if (shiftForcesNeedPbc)
506 /* Add this force to the shift force */
507 for (int d = 0; d < DIM; d++)
509 fshift[is][d] += receiveBuffer[n][d];
518 wallcycle_stop(wcycle, ewcMOVEF);
521 /* Convenience function for extracting a real buffer from an rvec buffer
523 * To reduce the number of temporary communication buffers and avoid
524 * cache polution, we reuse gmx::RVec buffers for storing reals.
525 * This functions return a real buffer reference with the same number
526 * of elements as the gmx::RVec buffer (so 1/3 of the size in bytes).
528 static gmx::ArrayRef<real>
529 realArrayRefFromRvecArrayRef(gmx::ArrayRef<gmx::RVec> arrayRef)
531 return gmx::arrayRefFromArray(as_rvec_array(arrayRef.data())[0],
535 void dd_atom_spread_real(gmx_domdec_t *dd, real v[])
538 gmx_domdec_comm_t *comm;
539 gmx_domdec_comm_dim_t *cd;
544 nat_tot = comm->atomRanges.numHomeAtoms();
545 for (int d = 0; d < dd->ndim; d++)
548 for (const gmx_domdec_ind_t &ind : cd->ind)
550 /* Note: We provision for RVec instead of real, so a factor of 3
551 * more than needed. The buffer actually already has this size
552 * and we pass a plain pointer below, so this does not matter.
554 DDBufferAccess<gmx::RVec> sendBufferAccess(comm->rvecBuffer, ind.nsend[nzone + 1]);
555 gmx::ArrayRef<real> sendBuffer = realArrayRefFromRvecArrayRef(sendBufferAccess.buffer);
557 for (int j : ind.index)
559 sendBuffer[n++] = v[j];
562 DDBufferAccess<gmx::RVec> receiveBufferAccess(comm->rvecBuffer2, cd->receiveInPlace ? 0 : ind.nrecv[nzone + 1]);
564 gmx::ArrayRef<real> receiveBuffer;
565 if (cd->receiveInPlace)
567 receiveBuffer = gmx::arrayRefFromArray(v + nat_tot, ind.nrecv[nzone + 1]);
571 receiveBuffer = realArrayRefFromRvecArrayRef(receiveBufferAccess.buffer);
573 /* Send and receive the data */
574 ddSendrecv(dd, d, dddirBackward,
575 sendBuffer, receiveBuffer);
576 if (!cd->receiveInPlace)
579 for (int zone = 0; zone < nzone; zone++)
581 for (int i = ind.cell2at0[zone]; i < ind.cell2at1[zone]; i++)
583 v[i] = receiveBuffer[j++];
587 nat_tot += ind.nrecv[nzone+1];
593 void dd_atom_sum_real(gmx_domdec_t *dd, real v[])
596 gmx_domdec_comm_t *comm;
597 gmx_domdec_comm_dim_t *cd;
601 nzone = comm->zones.n/2;
602 nat_tot = comm->atomRanges.end(DDAtomRanges::Type::Zones);
603 for (int d = dd->ndim-1; d >= 0; d--)
606 for (int p = cd->numPulses() - 1; p >= 0; p--)
608 const gmx_domdec_ind_t &ind = cd->ind[p];
610 /* Note: We provision for RVec instead of real, so a factor of 3
611 * more than needed. The buffer actually already has this size
612 * and we typecast, so this works as intended.
614 DDBufferAccess<gmx::RVec> receiveBufferAccess(comm->rvecBuffer, ind.nsend[nzone + 1]);
615 gmx::ArrayRef<real> receiveBuffer = realArrayRefFromRvecArrayRef(receiveBufferAccess.buffer);
616 nat_tot -= ind.nrecv[nzone + 1];
618 DDBufferAccess<gmx::RVec> sendBufferAccess(comm->rvecBuffer2, cd->receiveInPlace ? 0 : ind.nrecv[nzone + 1]);
620 gmx::ArrayRef<real> sendBuffer;
621 if (cd->receiveInPlace)
623 sendBuffer = gmx::arrayRefFromArray(v + nat_tot, ind.nrecv[nzone + 1]);
627 sendBuffer = realArrayRefFromRvecArrayRef(sendBufferAccess.buffer);
629 for (int zone = 0; zone < nzone; zone++)
631 for (int i = ind.cell2at0[zone]; i < ind.cell2at1[zone]; i++)
633 sendBuffer[j++] = v[i];
637 /* Communicate the forces */
638 ddSendrecv(dd, d, dddirForward,
639 sendBuffer, receiveBuffer);
640 /* Add the received forces */
642 for (int j : ind.index)
644 v[j] += receiveBuffer[n];
652 real dd_cutoff_multibody(const gmx_domdec_t *dd)
654 const gmx_domdec_comm_t &comm = *dd->comm;
655 const DDSystemInfo &systemInfo = comm.systemInfo;
658 if (systemInfo.haveInterDomainMultiBodyBondeds)
660 if (comm.cutoff_mbody > 0)
662 r = comm.cutoff_mbody;
666 /* cutoff_mbody=0 means we do not have DLB */
667 r = comm.cellsize_min[dd->dim[0]];
668 for (int di = 1; di < dd->ndim; di++)
670 r = std::min(r, comm.cellsize_min[dd->dim[di]]);
672 if (comm.systemInfo.filterBondedCommunication)
674 r = std::max(r, comm.cutoff_mbody);
678 r = std::min(r, systemInfo.cutoff);
686 real dd_cutoff_twobody(const gmx_domdec_t *dd)
690 r_mb = dd_cutoff_multibody(dd);
692 return std::max(dd->comm->systemInfo.cutoff, r_mb);
696 static void dd_cart_coord2pmecoord(const DDRankSetup &ddRankSetup,
697 const CartesianRankSetup &cartSetup,
701 const int nc = ddRankSetup.numPPCells[cartSetup.cartpmedim];
702 const int ntot = cartSetup.ntot[cartSetup.cartpmedim];
703 copy_ivec(coord, coord_pme);
704 coord_pme[cartSetup.cartpmedim] =
705 nc + (coord[cartSetup.cartpmedim]*(ntot - nc) + (ntot - nc)/2)/nc;
708 /* Returns the PME rank index in 0...npmenodes-1 for the PP cell with index ddCellIndex */
709 static int ddindex2pmeindex(const DDRankSetup &ddRankSetup,
710 const int ddCellIndex)
712 const int npp = ddRankSetup.numPPRanks;
713 const int npme = ddRankSetup.numRanksDoingPme;
715 /* Here we assign a PME node to communicate with this DD node
716 * by assuming that the major index of both is x.
717 * We add npme/2 to obtain an even distribution.
719 return (ddCellIndex*npme + npme/2)/npp;
722 static std::vector<int> dd_interleaved_pme_ranks(const DDRankSetup &ddRankSetup)
724 std::vector<int> pmeRanks(ddRankSetup.numRanksDoingPme);
727 for (int i = 0; i < ddRankSetup.numPPRanks; i++)
729 const int p0 = ddindex2pmeindex(ddRankSetup, i);
730 const int p1 = ddindex2pmeindex(ddRankSetup, i + 1);
731 if (i + 1 == ddRankSetup.numPPRanks || p1 > p0)
735 fprintf(debug, "pme_rank[%d] = %d\n", n, i+1+n);
737 pmeRanks[n] = i + 1 + n;
745 static int gmx_ddcoord2pmeindex(const t_commrec *cr, int x, int y, int z)
753 if (dd->comm->bCartesian) {
754 gmx_ddindex2xyz(dd->nc,ddindex,coords);
755 dd_coords2pmecoords(dd,coords,coords_pme);
756 copy_ivec(dd->ntot,nc);
757 nc[dd->cartpmedim] -= dd->nc[dd->cartpmedim];
758 coords_pme[dd->cartpmedim] -= dd->nc[dd->cartpmedim];
760 slab = (coords_pme[XX]*nc[YY] + coords_pme[YY])*nc[ZZ] + coords_pme[ZZ];
762 slab = (ddindex*cr->npmenodes + cr->npmenodes/2)/dd->nnodes;
768 slab = ddindex2pmeindex(dd->comm->ddRankSetup, dd_index(dd->nc, coords));
773 static int ddcoord2simnodeid(const t_commrec *cr, int x, int y, int z)
775 const CartesianRankSetup &cartSetup = cr->dd->comm->cartesianRankSetup;
776 ivec coords = { x, y, z };
779 if (cartSetup.bCartesianPP_PME)
782 MPI_Cart_rank(cr->mpi_comm_mysim, coords, &nodeid);
787 int ddindex = dd_index(cr->dd->nc, coords);
788 if (cartSetup.bCartesianPP)
790 nodeid = cartSetup.ddindex2simnodeid[ddindex];
794 if (cr->dd->comm->ddRankSetup.usePmeOnlyRanks)
796 nodeid = ddindex + gmx_ddcoord2pmeindex(cr, x, y, z);
808 static int dd_simnode2pmenode(const DDRankSetup &ddRankSetup,
809 const CartesianRankSetup &cartSetup,
810 gmx::ArrayRef<const int> pmeRanks,
811 const t_commrec gmx_unused *cr,
812 const int sim_nodeid)
816 /* This assumes a uniform x domain decomposition grid cell size */
817 if (cartSetup.bCartesianPP_PME)
820 ivec coord, coord_pme;
821 MPI_Cart_coords(cr->mpi_comm_mysim, sim_nodeid, DIM, coord);
822 if (coord[cartSetup.cartpmedim] < ddRankSetup.numPPCells[cartSetup.cartpmedim])
824 /* This is a PP rank */
825 dd_cart_coord2pmecoord(ddRankSetup, cartSetup, coord, coord_pme);
826 MPI_Cart_rank(cr->mpi_comm_mysim, coord_pme, &pmenode);
830 else if (cartSetup.bCartesianPP)
832 if (sim_nodeid < ddRankSetup.numPPRanks)
834 pmenode = ddRankSetup.numPPRanks + ddindex2pmeindex(ddRankSetup, sim_nodeid);
839 /* This assumes DD cells with identical x coordinates
840 * are numbered sequentially.
842 if (pmeRanks.empty())
844 if (sim_nodeid < ddRankSetup.numPPRanks)
846 /* The DD index equals the nodeid */
847 pmenode = ddRankSetup.numPPRanks + ddindex2pmeindex(ddRankSetup, sim_nodeid);
853 while (sim_nodeid > pmeRanks[i])
857 if (sim_nodeid < pmeRanks[i])
859 pmenode = pmeRanks[i];
867 NumPmeDomains getNumPmeDomains(const gmx_domdec_t *dd)
872 dd->comm->ddRankSetup.npmenodes_x,
873 dd->comm->ddRankSetup.npmenodes_y
884 std::vector<int> get_pme_ddranks(const t_commrec *cr,
887 const DDRankSetup &ddRankSetup = cr->dd->comm->ddRankSetup;
888 const CartesianRankSetup &cartSetup = cr->dd->comm->cartesianRankSetup;
889 GMX_RELEASE_ASSERT(ddRankSetup.usePmeOnlyRanks, "This function should only be called when PME-only ranks are in use");
890 std::vector<int> ddranks;
891 ddranks.reserve((ddRankSetup.numPPRanks + ddRankSetup.numRanksDoingPme - 1)/ddRankSetup.numRanksDoingPme);
893 for (int x = 0; x < ddRankSetup.numPPCells[XX]; x++)
895 for (int y = 0; y < ddRankSetup.numPPCells[YY]; y++)
897 for (int z = 0; z < ddRankSetup.numPPCells[ZZ]; z++)
899 if (cartSetup.bCartesianPP_PME)
901 ivec coord = { x, y, z };
903 dd_cart_coord2pmecoord(ddRankSetup, cartSetup, coord, coord_pme);
904 if (cr->dd->ci[XX] == coord_pme[XX] &&
905 cr->dd->ci[YY] == coord_pme[YY] &&
906 cr->dd->ci[ZZ] == coord_pme[ZZ])
908 ddranks.push_back(ddcoord2simnodeid(cr, x, y, z));
913 /* The slab corresponds to the nodeid in the PME group */
914 if (gmx_ddcoord2pmeindex(cr, x, y, z) == pmenodeid)
916 ddranks.push_back(ddcoord2simnodeid(cr, x, y, z));
925 static gmx_bool receive_vir_ener(const gmx_domdec_t *dd,
926 gmx::ArrayRef<const int> pmeRanks,
929 gmx_bool bReceive = TRUE;
931 const DDRankSetup &ddRankSetup = dd->comm->ddRankSetup;
932 if (ddRankSetup.usePmeOnlyRanks)
934 const CartesianRankSetup &cartSetup = dd->comm->cartesianRankSetup;
935 if (cartSetup.bCartesianPP_PME)
938 int pmenode = dd_simnode2pmenode(ddRankSetup, cartSetup, pmeRanks, cr, cr->sim_nodeid);
940 MPI_Cart_coords(cr->mpi_comm_mysim, cr->sim_nodeid, DIM, coords);
941 coords[cartSetup.cartpmedim]++;
942 if (coords[cartSetup.cartpmedim] < dd->nc[cartSetup.cartpmedim])
945 MPI_Cart_rank(cr->mpi_comm_mysim, coords, &rank);
946 if (dd_simnode2pmenode(ddRankSetup, cartSetup, pmeRanks, cr, rank) == pmenode)
948 /* This is not the last PP node for pmenode */
953 GMX_RELEASE_ASSERT(false, "Without MPI we should not have Cartesian PP-PME with #PMEnodes < #DDnodes");
958 int pmenode = dd_simnode2pmenode(ddRankSetup, cartSetup, pmeRanks, cr, cr->sim_nodeid);
959 if (cr->sim_nodeid+1 < cr->nnodes &&
960 dd_simnode2pmenode(ddRankSetup, cartSetup, pmeRanks, cr, cr->sim_nodeid + 1) == pmenode)
962 /* This is not the last PP node for pmenode */
971 static void set_slb_pme_dim_f(gmx_domdec_t *dd, int dim, real **dim_f)
973 gmx_domdec_comm_t *comm;
978 snew(*dim_f, dd->nc[dim]+1);
980 for (i = 1; i < dd->nc[dim]; i++)
982 if (comm->slb_frac[dim])
984 (*dim_f)[i] = (*dim_f)[i-1] + comm->slb_frac[dim][i-1];
988 (*dim_f)[i] = static_cast<real>(i)/static_cast<real>(dd->nc[dim]);
991 (*dim_f)[dd->nc[dim]] = 1;
994 static void init_ddpme(gmx_domdec_t *dd, gmx_ddpme_t *ddpme, int dimind)
996 const DDRankSetup &ddRankSetup = dd->comm->ddRankSetup;
998 if (dimind == 0 && dd->dim[0] == YY && ddRankSetup.npmenodes_x == 1)
1004 ddpme->dim = dimind;
1006 ddpme->dim_match = (ddpme->dim == dd->dim[dimind]);
1008 ddpme->nslab = (ddpme->dim == 0 ?
1009 ddRankSetup.npmenodes_x :
1010 ddRankSetup.npmenodes_y);
1012 if (ddpme->nslab <= 1)
1017 const int nso = ddRankSetup.numRanksDoingPme/ddpme->nslab;
1018 /* Determine for each PME slab the PP location range for dimension dim */
1019 snew(ddpme->pp_min, ddpme->nslab);
1020 snew(ddpme->pp_max, ddpme->nslab);
1021 for (int slab = 0; slab < ddpme->nslab; slab++)
1023 ddpme->pp_min[slab] = dd->nc[dd->dim[dimind]] - 1;
1024 ddpme->pp_max[slab] = 0;
1026 for (int i = 0; i < dd->nnodes; i++)
1029 ddindex2xyz(dd->nc, i, xyz);
1030 /* For y only use our y/z slab.
1031 * This assumes that the PME x grid size matches the DD grid size.
1033 if (dimind == 0 || xyz[XX] == dd->ci[XX])
1035 const int pmeindex = ddindex2pmeindex(ddRankSetup, i);
1039 slab = pmeindex/nso;
1043 slab = pmeindex % ddpme->nslab;
1045 ddpme->pp_min[slab] = std::min(ddpme->pp_min[slab], xyz[dimind]);
1046 ddpme->pp_max[slab] = std::max(ddpme->pp_max[slab], xyz[dimind]);
1050 set_slb_pme_dim_f(dd, ddpme->dim, &ddpme->slb_dim_f);
1053 int dd_pme_maxshift_x(const gmx_domdec_t *dd)
1055 const DDRankSetup &ddRankSetup = dd->comm->ddRankSetup;
1057 if (ddRankSetup.ddpme[0].dim == XX)
1059 return ddRankSetup.ddpme[0].maxshift;
1067 int dd_pme_maxshift_y(const gmx_domdec_t *dd)
1069 const DDRankSetup &ddRankSetup = dd->comm->ddRankSetup;
1071 if (ddRankSetup.ddpme[0].dim == YY)
1073 return ddRankSetup.ddpme[0].maxshift;
1075 else if (ddRankSetup.npmedecompdim >= 2 && ddRankSetup.ddpme[1].dim == YY)
1077 return ddRankSetup.ddpme[1].maxshift;
1085 bool ddHaveSplitConstraints(const gmx_domdec_t &dd)
1087 return dd.comm->systemInfo.haveSplitConstraints;
1090 void dd_cycles_add(const gmx_domdec_t *dd, float cycles, int ddCycl)
1092 /* Note that the cycles value can be incorrect, either 0 or some
1093 * extremely large value, when our thread migrated to another core
1094 * with an unsynchronized cycle counter. If this happens less often
1095 * that once per nstlist steps, this will not cause issues, since
1096 * we later subtract the maximum value from the sum over nstlist steps.
1097 * A zero count will slightly lower the total, but that's a small effect.
1098 * Note that the main purpose of the subtraction of the maximum value
1099 * is to avoid throwing off the load balancing when stalls occur due
1100 * e.g. system activity or network congestion.
1102 dd->comm->cycl[ddCycl] += cycles;
1103 dd->comm->cycl_n[ddCycl]++;
1104 if (cycles > dd->comm->cycl_max[ddCycl])
1106 dd->comm->cycl_max[ddCycl] = cycles;
1111 static void make_load_communicator(gmx_domdec_t *dd, int dim_ind, ivec loc)
1116 gmx_bool bPartOfGroup = FALSE;
1118 dim = dd->dim[dim_ind];
1119 copy_ivec(loc, loc_c);
1120 for (i = 0; i < dd->nc[dim]; i++)
1123 rank = dd_index(dd->nc, loc_c);
1124 if (rank == dd->rank)
1126 /* This process is part of the group */
1127 bPartOfGroup = TRUE;
1130 MPI_Comm_split(dd->mpi_comm_all, bPartOfGroup ? 0 : MPI_UNDEFINED, dd->rank,
1134 dd->comm->mpi_comm_load[dim_ind] = c_row;
1135 if (!isDlbDisabled(dd->comm))
1137 DDCellsizesWithDlb &cellsizes = dd->comm->cellsizesWithDlb[dim_ind];
1139 if (dd->ci[dim] == dd->master_ci[dim])
1141 /* This is the root process of this row */
1142 cellsizes.rowMaster = std::make_unique<RowMaster>();
1144 RowMaster &rowMaster = *cellsizes.rowMaster;
1145 rowMaster.cellFrac.resize(ddCellFractionBufferSize(dd, dim_ind));
1146 rowMaster.oldCellFrac.resize(dd->nc[dim] + 1);
1147 rowMaster.isCellMin.resize(dd->nc[dim]);
1150 rowMaster.bounds.resize(dd->nc[dim]);
1152 rowMaster.buf_ncd.resize(dd->nc[dim]);
1156 /* This is not a root process, we only need to receive cell_f */
1157 cellsizes.fracRow.resize(ddCellFractionBufferSize(dd, dim_ind));
1160 if (dd->ci[dim] == dd->master_ci[dim])
1162 snew(dd->comm->load[dim_ind].load, dd->nc[dim]*DD_NLOAD_MAX);
1168 void dd_setup_dlb_resource_sharing(t_commrec *cr,
1172 int physicalnode_id_hash;
1174 MPI_Comm mpi_comm_pp_physicalnode;
1176 if (!thisRankHasDuty(cr, DUTY_PP) || gpu_id < 0)
1178 /* Only ranks with short-ranged tasks (currently) use GPUs.
1179 * If we don't have GPUs assigned, there are no resources to share.
1184 physicalnode_id_hash = gmx_physicalnode_id_hash();
1190 fprintf(debug, "dd_setup_dd_dlb_gpu_sharing:\n");
1191 fprintf(debug, "DD PP rank %d physical node hash %d gpu_id %d\n",
1192 dd->rank, physicalnode_id_hash, gpu_id);
1194 /* Split the PP communicator over the physical nodes */
1195 /* TODO: See if we should store this (before), as it's also used for
1196 * for the nodecomm summation.
1198 // TODO PhysicalNodeCommunicator could be extended/used to handle
1199 // the need for per-node per-group communicators.
1200 MPI_Comm_split(dd->mpi_comm_all, physicalnode_id_hash, dd->rank,
1201 &mpi_comm_pp_physicalnode);
1202 MPI_Comm_split(mpi_comm_pp_physicalnode, gpu_id, dd->rank,
1203 &dd->comm->mpi_comm_gpu_shared);
1204 MPI_Comm_free(&mpi_comm_pp_physicalnode);
1205 MPI_Comm_size(dd->comm->mpi_comm_gpu_shared, &dd->comm->nrank_gpu_shared);
1209 fprintf(debug, "nrank_gpu_shared %d\n", dd->comm->nrank_gpu_shared);
1212 /* Note that some ranks could share a GPU, while others don't */
1214 if (dd->comm->nrank_gpu_shared == 1)
1216 MPI_Comm_free(&dd->comm->mpi_comm_gpu_shared);
1219 GMX_UNUSED_VALUE(cr);
1220 GMX_UNUSED_VALUE(gpu_id);
1224 static void make_load_communicators(gmx_domdec_t gmx_unused *dd)
1227 int dim0, dim1, i, j;
1232 fprintf(debug, "Making load communicators\n");
1235 dd->comm->load = new domdec_load_t[std::max(dd->ndim, 1)];
1236 snew(dd->comm->mpi_comm_load, std::max(dd->ndim, 1));
1244 make_load_communicator(dd, 0, loc);
1248 for (i = 0; i < dd->nc[dim0]; i++)
1251 make_load_communicator(dd, 1, loc);
1257 for (i = 0; i < dd->nc[dim0]; i++)
1261 for (j = 0; j < dd->nc[dim1]; j++)
1264 make_load_communicator(dd, 2, loc);
1271 fprintf(debug, "Finished making load communicators\n");
1276 /*! \brief Sets up the relation between neighboring domains and zones */
1277 static void setup_neighbor_relations(gmx_domdec_t *dd)
1279 int d, dim, i, j, m;
1281 gmx_domdec_zones_t *zones;
1282 gmx_domdec_ns_ranges_t *izone;
1283 GMX_ASSERT(dd->ndim >= 0, "Must have non-negative number of dimensions for DD");
1285 for (d = 0; d < dd->ndim; d++)
1288 copy_ivec(dd->ci, tmp);
1289 tmp[dim] = (tmp[dim] + 1) % dd->nc[dim];
1290 dd->neighbor[d][0] = ddcoord2ddnodeid(dd, tmp);
1291 copy_ivec(dd->ci, tmp);
1292 tmp[dim] = (tmp[dim] - 1 + dd->nc[dim]) % dd->nc[dim];
1293 dd->neighbor[d][1] = ddcoord2ddnodeid(dd, tmp);
1296 fprintf(debug, "DD rank %d neighbor ranks in dir %d are + %d - %d\n",
1299 dd->neighbor[d][1]);
1303 int nzone = (1 << dd->ndim);
1304 GMX_ASSERT(dd->ndim < DIM, "Invalid number of dimensions");
1305 int nizone = (1 << std::max(dd->ndim - 1, 0));
1306 assert(nizone >= 1 && nizone <= DD_MAXIZONE);
1308 zones = &dd->comm->zones;
1310 for (i = 0; i < nzone; i++)
1313 clear_ivec(zones->shift[i]);
1314 for (d = 0; d < dd->ndim; d++)
1316 zones->shift[i][dd->dim[d]] = dd_zo[i][m++];
1321 for (i = 0; i < nzone; i++)
1323 for (d = 0; d < DIM; d++)
1325 s[d] = dd->ci[d] - zones->shift[i][d];
1330 else if (s[d] >= dd->nc[d])
1336 zones->nizone = nizone;
1337 for (i = 0; i < zones->nizone; i++)
1339 assert(ddNonbondedZonePairRanges[i][0] == i);
1341 izone = &zones->izone[i];
1342 /* dd_zp3 is for 3D decomposition, for fewer dimensions use only
1343 * j-zones up to nzone.
1345 izone->j0 = std::min(ddNonbondedZonePairRanges[i][1], nzone);
1346 izone->j1 = std::min(ddNonbondedZonePairRanges[i][2], nzone);
1347 for (dim = 0; dim < DIM; dim++)
1349 if (dd->nc[dim] == 1)
1351 /* All shifts should be allowed */
1352 izone->shift0[dim] = -1;
1353 izone->shift1[dim] = 1;
1357 /* Determine the min/max j-zone shift wrt the i-zone */
1358 izone->shift0[dim] = 1;
1359 izone->shift1[dim] = -1;
1360 for (j = izone->j0; j < izone->j1; j++)
1362 int shift_diff = zones->shift[j][dim] - zones->shift[i][dim];
1363 if (shift_diff < izone->shift0[dim])
1365 izone->shift0[dim] = shift_diff;
1367 if (shift_diff > izone->shift1[dim])
1369 izone->shift1[dim] = shift_diff;
1376 if (!isDlbDisabled(dd->comm))
1378 dd->comm->cellsizesWithDlb.resize(dd->ndim);
1381 if (dd->comm->ddSettings.recordLoad)
1383 make_load_communicators(dd);
1387 static void make_pp_communicator(const gmx::MDLogger &mdlog,
1389 t_commrec gmx_unused *cr,
1390 bool gmx_unused reorder)
1393 gmx_domdec_comm_t *comm = dd->comm;
1394 CartesianRankSetup &cartSetup = comm->cartesianRankSetup;
1396 if (cartSetup.bCartesianPP)
1398 /* Set up cartesian communication for the particle-particle part */
1399 GMX_LOG(mdlog.info).appendTextFormatted(
1400 "Will use a Cartesian communicator: %d x %d x %d",
1401 dd->nc[XX], dd->nc[YY], dd->nc[ZZ]);
1404 for (int i = 0; i < DIM; i++)
1409 MPI_Cart_create(cr->mpi_comm_mygroup, DIM, dd->nc, periods, static_cast<int>(reorder),
1411 /* We overwrite the old communicator with the new cartesian one */
1412 cr->mpi_comm_mygroup = comm_cart;
1415 dd->mpi_comm_all = cr->mpi_comm_mygroup;
1416 MPI_Comm_rank(dd->mpi_comm_all, &dd->rank);
1418 if (cartSetup.bCartesianPP_PME)
1420 /* Since we want to use the original cartesian setup for sim,
1421 * and not the one after split, we need to make an index.
1423 cartSetup.ddindex2ddnodeid.resize(dd->nnodes);
1424 cartSetup.ddindex2ddnodeid[dd_index(dd->nc, dd->ci)] = dd->rank;
1425 gmx_sumi(dd->nnodes, cartSetup.ddindex2ddnodeid.data(), cr);
1426 /* Get the rank of the DD master,
1427 * above we made sure that the master node is a PP node.
1438 MPI_Allreduce(&rank, &dd->masterrank, 1, MPI_INT, MPI_SUM, dd->mpi_comm_all);
1440 else if (cartSetup.bCartesianPP)
1442 if (!comm->ddRankSetup.usePmeOnlyRanks)
1444 /* The PP communicator is also
1445 * the communicator for this simulation
1447 cr->mpi_comm_mysim = cr->mpi_comm_mygroup;
1449 cr->nodeid = dd->rank;
1451 MPI_Cart_coords(dd->mpi_comm_all, dd->rank, DIM, dd->ci);
1453 /* We need to make an index to go from the coordinates
1454 * to the nodeid of this simulation.
1456 cartSetup.ddindex2simnodeid.resize(dd->nnodes);
1457 std::vector<int> buf(dd->nnodes);
1458 if (thisRankHasDuty(cr, DUTY_PP))
1460 buf[dd_index(dd->nc, dd->ci)] = cr->sim_nodeid;
1462 /* Communicate the ddindex to simulation nodeid index */
1463 MPI_Allreduce(buf.data(), cartSetup.ddindex2simnodeid.data(), dd->nnodes, MPI_INT, MPI_SUM,
1464 cr->mpi_comm_mysim);
1466 /* Determine the master coordinates and rank.
1467 * The DD master should be the same node as the master of this sim.
1469 for (int i = 0; i < dd->nnodes; i++)
1471 if (cartSetup.ddindex2simnodeid[i] == 0)
1473 ddindex2xyz(dd->nc, i, dd->master_ci);
1474 MPI_Cart_rank(dd->mpi_comm_all, dd->master_ci, &dd->masterrank);
1479 fprintf(debug, "The master rank is %d\n", dd->masterrank);
1484 /* No Cartesian communicators */
1485 /* We use the rank in dd->comm->all as DD index */
1486 ddindex2xyz(dd->nc, dd->rank, dd->ci);
1487 /* The simulation master nodeid is 0, so the DD master rank is also 0 */
1489 clear_ivec(dd->master_ci);
1493 GMX_LOG(mdlog.info).appendTextFormatted(
1494 "Domain decomposition rank %d, coordinates %d %d %d\n",
1495 dd->rank, dd->ci[XX], dd->ci[YY], dd->ci[ZZ]);
1499 "Domain decomposition rank %d, coordinates %d %d %d\n\n",
1500 dd->rank, dd->ci[XX], dd->ci[YY], dd->ci[ZZ]);
1504 static void receive_ddindex2simnodeid(gmx_domdec_t *dd,
1508 CartesianRankSetup &cartSetup = dd->comm->cartesianRankSetup;
1510 if (!cartSetup.bCartesianPP_PME && cartSetup.bCartesianPP)
1512 cartSetup.ddindex2simnodeid.resize(dd->nnodes);
1513 std::vector<int> buf(dd->nnodes);
1514 if (thisRankHasDuty(cr, DUTY_PP))
1516 buf[dd_index(dd->nc, dd->ci)] = cr->sim_nodeid;
1518 /* Communicate the ddindex to simulation nodeid index */
1519 MPI_Allreduce(buf.data(), cartSetup.ddindex2simnodeid.data(), dd->nnodes, MPI_INT, MPI_SUM,
1520 cr->mpi_comm_mysim);
1523 GMX_UNUSED_VALUE(dd);
1524 GMX_UNUSED_VALUE(cr);
1528 static CartesianRankSetup
1529 split_communicator(const gmx::MDLogger &mdlog,
1531 const DdRankOrder ddRankOrder,
1532 bool gmx_unused reorder,
1533 const DDRankSetup &ddRankSetup,
1535 std::vector<int> *pmeRanks)
1537 CartesianRankSetup cartSetup;
1539 cartSetup.bCartesianPP = (ddRankOrder == DdRankOrder::cartesian);
1540 cartSetup.bCartesianPP_PME = false;
1542 const ivec &numDDCells = ddRankSetup.numPPCells;
1543 /* Initially we set ntot to the number of PP cells */
1544 copy_ivec(numDDCells, cartSetup.ntot);
1546 if (cartSetup.bCartesianPP)
1548 const int numDDCellsTot = ddRankSetup.numPPRanks;
1550 for (int i = 1; i < DIM; i++)
1552 bDiv[i] = ((ddRankSetup.numRanksDoingPme*numDDCells[i]) % numDDCellsTot == 0);
1554 if (bDiv[YY] || bDiv[ZZ])
1556 cartSetup.bCartesianPP_PME = TRUE;
1557 /* If we have 2D PME decomposition, which is always in x+y,
1558 * we stack the PME only nodes in z.
1559 * Otherwise we choose the direction that provides the thinnest slab
1560 * of PME only nodes as this will have the least effect
1561 * on the PP communication.
1562 * But for the PME communication the opposite might be better.
1564 if (bDiv[ZZ] && (ddRankSetup.npmenodes_y > 1 ||
1566 numDDCells[YY] > numDDCells[ZZ]))
1568 cartSetup.cartpmedim = ZZ;
1572 cartSetup.cartpmedim = YY;
1574 cartSetup.ntot[cartSetup.cartpmedim]
1575 += (ddRankSetup.numRanksDoingPme*numDDCells[cartSetup.cartpmedim])/numDDCellsTot;
1579 GMX_LOG(mdlog.info).appendTextFormatted(
1580 "Number of PME-only ranks (%d) is not a multiple of nx*ny (%d*%d) or nx*nz (%d*%d)",
1581 ddRankSetup.numRanksDoingPme,
1582 numDDCells[XX], numDDCells[YY],
1583 numDDCells[XX], numDDCells[ZZ]);
1584 GMX_LOG(mdlog.info).appendText("Will not use a Cartesian communicator for PP <-> PME\n");
1588 if (cartSetup.bCartesianPP_PME)
1594 GMX_LOG(mdlog.info).appendTextFormatted(
1595 "Will use a Cartesian communicator for PP <-> PME: %d x %d x %d",
1596 cartSetup.ntot[XX], cartSetup.ntot[YY], cartSetup.ntot[ZZ]);
1598 for (int i = 0; i < DIM; i++)
1603 MPI_Cart_create(cr->mpi_comm_mysim, DIM, cartSetup.ntot, periods, static_cast<int>(reorder),
1605 MPI_Comm_rank(comm_cart, &rank);
1606 if (MASTER(cr) && rank != 0)
1608 gmx_fatal(FARGS, "MPI rank 0 was renumbered by MPI_Cart_create, we do not allow this");
1611 /* With this assigment we loose the link to the original communicator
1612 * which will usually be MPI_COMM_WORLD, unless have multisim.
1614 cr->mpi_comm_mysim = comm_cart;
1615 cr->sim_nodeid = rank;
1617 MPI_Cart_coords(cr->mpi_comm_mysim, cr->sim_nodeid, DIM, ddCellIndex);
1619 GMX_LOG(mdlog.info).appendTextFormatted(
1620 "Cartesian rank %d, coordinates %d %d %d\n",
1621 cr->sim_nodeid, ddCellIndex[XX], ddCellIndex[YY], ddCellIndex[ZZ]);
1623 if (ddCellIndex[cartSetup.cartpmedim] < numDDCells[cartSetup.cartpmedim])
1627 if (!ddRankSetup.usePmeOnlyRanks ||
1628 ddCellIndex[cartSetup.cartpmedim] >= numDDCells[cartSetup.cartpmedim])
1630 cr->duty = DUTY_PME;
1633 /* Split the sim communicator into PP and PME only nodes */
1634 MPI_Comm_split(cr->mpi_comm_mysim,
1635 getThisRankDuties(cr),
1636 dd_index(cartSetup.ntot, ddCellIndex),
1637 &cr->mpi_comm_mygroup);
1639 GMX_UNUSED_VALUE(ddCellIndex);
1644 switch (ddRankOrder)
1646 case DdRankOrder::pp_pme:
1647 GMX_LOG(mdlog.info).appendText("Order of the ranks: PP first, PME last");
1649 case DdRankOrder::interleave:
1650 /* Interleave the PP-only and PME-only ranks */
1651 GMX_LOG(mdlog.info).appendText("Interleaving PP and PME ranks");
1652 *pmeRanks = dd_interleaved_pme_ranks(ddRankSetup);
1654 case DdRankOrder::cartesian:
1657 gmx_fatal(FARGS, "Invalid ddRankOrder=%d", static_cast<int>(ddRankOrder));
1660 if (dd_simnode2pmenode(ddRankSetup, cartSetup, *pmeRanks, cr, cr->sim_nodeid) == -1)
1662 cr->duty = DUTY_PME;
1669 /* Split the sim communicator into PP and PME only nodes */
1670 MPI_Comm_split(cr->mpi_comm_mysim,
1671 getThisRankDuties(cr),
1673 &cr->mpi_comm_mygroup);
1674 MPI_Comm_rank(cr->mpi_comm_mygroup, &cr->nodeid);
1678 GMX_LOG(mdlog.info).appendTextFormatted(
1679 "This rank does only %s work.\n",
1680 thisRankHasDuty(cr, DUTY_PP) ? "particle-particle" : "PME-mesh");
1685 /*! \brief Makes the PP communicator and the PME communicator, when needed
1687 * Returns the Cartesian rank setup.
1688 * Sets \p cr->mpi_comm_mygroup
1689 * For PP ranks, sets the DD PP cell index in \p ddCellIndex.
1690 * With separate PME ranks in interleaved order, set the PME ranks in \p pmeRanks.
1692 static CartesianRankSetup
1693 makeGroupCommunicators(const gmx::MDLogger &mdlog,
1694 const DDSettings &ddSettings,
1695 const DdRankOrder ddRankOrder,
1696 const DDRankSetup &ddRankSetup,
1699 std::vector<int> *pmeRanks)
1701 CartesianRankSetup cartSetup;
1703 if (ddRankSetup.usePmeOnlyRanks)
1705 /* Split the communicator into a PP and PME part */
1707 split_communicator(mdlog, cr, ddRankOrder, ddSettings.useCartesianReorder,
1708 ddRankSetup, ddCellIndex, pmeRanks);
1712 /* All nodes do PP and PME */
1713 /* We do not require separate communicators */
1714 cr->mpi_comm_mygroup = cr->mpi_comm_mysim;
1716 cartSetup.bCartesianPP = false;
1717 cartSetup.bCartesianPP_PME = false;
1723 /*! \brief For PP ranks, sets or makes the communicator
1725 * For PME ranks get the rank id.
1726 * For PP only ranks, sets the PME-only rank.
1728 static void setupGroupCommunication(const gmx::MDLogger &mdlog,
1729 const DDSettings &ddSettings,
1730 gmx::ArrayRef<const int> pmeRanks,
1734 const DDRankSetup &ddRankSetup = dd->comm->ddRankSetup;
1735 const CartesianRankSetup &cartSetup = dd->comm->cartesianRankSetup;
1737 if (thisRankHasDuty(cr, DUTY_PP))
1739 /* Copy or make a new PP communicator */
1741 /* We (possibly) reordered the nodes in split_communicator,
1742 * so it is no longer required in make_pp_communicator.
1744 const bool useCartesianReorder =
1745 (ddSettings.useCartesianReorder &&
1746 !cartSetup.bCartesianPP_PME);
1748 make_pp_communicator(mdlog, dd, cr, useCartesianReorder);
1752 receive_ddindex2simnodeid(dd, cr);
1755 if (!thisRankHasDuty(cr, DUTY_PME))
1757 /* Set up the commnuication to our PME node */
1758 dd->pme_nodeid = dd_simnode2pmenode(ddRankSetup, cartSetup, pmeRanks, cr, cr->sim_nodeid);
1759 dd->pme_receive_vir_ener = receive_vir_ener(dd, pmeRanks, cr);
1762 fprintf(debug, "My pme_nodeid %d receive ener %s\n",
1763 dd->pme_nodeid, gmx::boolToString(dd->pme_receive_vir_ener));
1768 dd->pme_nodeid = -1;
1771 /* We can not use DDMASTER(dd), because dd->masterrank is set later */
1774 dd->ma = std::make_unique<AtomDistribution>(dd->nc,
1775 dd->comm->cgs_gl.nr,
1776 dd->comm->cgs_gl.index[dd->comm->cgs_gl.nr]);
1780 static real *get_slb_frac(const gmx::MDLogger &mdlog,
1781 const char *dir, int nc, const char *size_string)
1783 real *slb_frac, tot;
1788 if (nc > 1 && size_string != nullptr)
1790 GMX_LOG(mdlog.info).appendTextFormatted(
1791 "Using static load balancing for the %s direction", dir);
1794 for (i = 0; i < nc; i++)
1797 sscanf(size_string, "%20lf%n", &dbl, &n);
1800 gmx_fatal(FARGS, "Incorrect or not enough DD cell size entries for direction %s: '%s'", dir, size_string);
1807 std::string relativeCellSizes = "Relative cell sizes:";
1808 for (i = 0; i < nc; i++)
1811 relativeCellSizes += gmx::formatString(" %5.3f", slb_frac[i]);
1813 GMX_LOG(mdlog.info).appendText(relativeCellSizes);
1819 static int multi_body_bondeds_count(const gmx_mtop_t *mtop)
1822 gmx_mtop_ilistloop_t iloop = gmx_mtop_ilistloop_init(mtop);
1824 while (const InteractionLists *ilists = gmx_mtop_ilistloop_next(iloop, &nmol))
1826 for (auto &ilist : extractILists(*ilists, IF_BOND))
1828 if (NRAL(ilist.functionType) > 2)
1830 n += nmol*(ilist.iatoms.size()/ilistStride(ilist));
1838 static int dd_getenv(const gmx::MDLogger &mdlog,
1839 const char *env_var, int def)
1845 val = getenv(env_var);
1848 if (sscanf(val, "%20d", &nst) <= 0)
1852 GMX_LOG(mdlog.info).appendTextFormatted(
1853 "Found env.var. %s = %s, using value %d",
1860 static void check_dd_restrictions(const gmx_domdec_t *dd,
1861 const t_inputrec *ir,
1862 const gmx::MDLogger &mdlog)
1864 if (ir->ePBC == epbcSCREW &&
1865 (dd->nc[XX] == 1 || dd->nc[YY] > 1 || dd->nc[ZZ] > 1))
1867 gmx_fatal(FARGS, "With pbc=%s can only do domain decomposition in the x-direction", epbc_names[ir->ePBC]);
1870 if (ir->ns_type == ensSIMPLE)
1872 gmx_fatal(FARGS, "Domain decomposition does not support simple neighbor searching, use grid searching or run with one MPI rank");
1875 if (ir->nstlist == 0)
1877 gmx_fatal(FARGS, "Domain decomposition does not work with nstlist=0");
1880 if (ir->comm_mode == ecmANGULAR && ir->ePBC != epbcNONE)
1882 GMX_LOG(mdlog.warning).appendText("comm-mode angular will give incorrect results when the comm group partially crosses a periodic boundary");
1886 static real average_cellsize_min(const gmx_ddbox_t &ddbox,
1887 const ivec numDomains)
1889 real r = ddbox.box_size[XX];
1890 for (int d = 0; d < DIM; d++)
1892 if (numDomains[d] > 1)
1894 /* Check using the initial average cell size */
1895 r = std::min(r, ddbox.box_size[d]*ddbox.skew_fac[d]/numDomains[d]);
1902 /*! \brief Depending on the DLB initial value return the DLB switched off state or issue an error.
1904 static DlbState forceDlbOffOrBail(DlbState cmdlineDlbState,
1905 const std::string &reasonStr,
1906 const gmx::MDLogger &mdlog)
1908 std::string dlbNotSupportedErr = "Dynamic load balancing requested, but ";
1909 std::string dlbDisableNote = "NOTE: disabling dynamic load balancing as ";
1911 if (cmdlineDlbState == DlbState::onUser)
1913 gmx_fatal(FARGS, "%s", (dlbNotSupportedErr + reasonStr).c_str());
1915 else if (cmdlineDlbState == DlbState::offCanTurnOn)
1917 GMX_LOG(mdlog.info).appendText(dlbDisableNote + reasonStr);
1919 return DlbState::offForever;
1922 /*! \brief Return the dynamic load balancer's initial state based on initial conditions and user inputs.
1924 * This function parses the parameters of "-dlb" command line option setting
1925 * corresponding state values. Then it checks the consistency of the determined
1926 * state with other run parameters and settings. As a result, the initial state
1927 * may be altered or an error may be thrown if incompatibility of options is detected.
1929 * \param [in] mdlog Logger.
1930 * \param [in] dlbOption Enum value for the DLB option.
1931 * \param [in] bRecordLoad True if the load balancer is recording load information.
1932 * \param [in] mdrunOptions Options for mdrun.
1933 * \param [in] ir Pointer mdrun to input parameters.
1934 * \returns DLB initial/startup state.
1936 static DlbState determineInitialDlbState(const gmx::MDLogger &mdlog,
1937 DlbOption dlbOption, gmx_bool bRecordLoad,
1938 const gmx::MdrunOptions &mdrunOptions,
1939 const t_inputrec *ir)
1941 DlbState dlbState = DlbState::offCanTurnOn;
1945 case DlbOption::turnOnWhenUseful: dlbState = DlbState::offCanTurnOn; break;
1946 case DlbOption::no: dlbState = DlbState::offUser; break;
1947 case DlbOption::yes: dlbState = DlbState::onUser; break;
1948 default: gmx_incons("Invalid dlbOption enum value");
1951 /* Reruns don't support DLB: bail or override auto mode */
1952 if (mdrunOptions.rerun)
1954 std::string reasonStr = "it is not supported in reruns.";
1955 return forceDlbOffOrBail(dlbState, reasonStr, mdlog);
1958 /* Unsupported integrators */
1959 if (!EI_DYNAMICS(ir->eI))
1961 auto reasonStr = gmx::formatString("it is only supported with dynamics, not with integrator '%s'.", EI(ir->eI));
1962 return forceDlbOffOrBail(dlbState, reasonStr, mdlog);
1965 /* Without cycle counters we can't time work to balance on */
1968 std::string reasonStr = "cycle counters unsupported or not enabled in the operating system kernel.";
1969 return forceDlbOffOrBail(dlbState, reasonStr, mdlog);
1972 if (mdrunOptions.reproducible)
1974 std::string reasonStr = "you started a reproducible run.";
1977 case DlbState::offUser:
1979 case DlbState::offForever:
1980 GMX_RELEASE_ASSERT(false, "DlbState::offForever is not a valid initial state");
1982 case DlbState::offCanTurnOn:
1983 return forceDlbOffOrBail(dlbState, reasonStr, mdlog);
1984 case DlbState::onCanTurnOff:
1985 GMX_RELEASE_ASSERT(false, "DlbState::offCanTurnOff is not a valid initial state");
1987 case DlbState::onUser:
1988 return forceDlbOffOrBail(dlbState, reasonStr + " In load balanced runs binary reproducibility cannot be ensured.", mdlog);
1990 gmx_fatal(FARGS, "Death horror: undefined case (%d) for load balancing choice", static_cast<int>(dlbState));
1997 /* Sets the order of the DD dimensions, returns the number of DD dimensions */
1998 static int set_dd_dim(const ivec numDDCells,
1999 const DDSettings &ddSettings,
2003 if (ddSettings.useDDOrderZYX)
2005 /* Decomposition order z,y,x */
2006 for (int dim = DIM - 1; dim >= 0; dim--)
2008 if (numDDCells[dim] > 1)
2016 /* Decomposition order x,y,z */
2017 for (int dim = 0; dim < DIM; dim++)
2019 if (numDDCells[dim] > 1)
2028 /* Set dim[0] to avoid extra checks on ndim in several places */
2035 static gmx_domdec_comm_t *init_dd_comm()
2037 gmx_domdec_comm_t *comm = new gmx_domdec_comm_t;
2039 comm->n_load_have = 0;
2040 comm->n_load_collect = 0;
2042 comm->haveTurnedOffDlb = false;
2044 for (int i = 0; i < static_cast<int>(DDAtomRanges::Type::Number); i++)
2046 comm->sum_nat[i] = 0;
2050 comm->load_step = 0;
2053 clear_ivec(comm->load_lim);
2057 /* This should be replaced by a unique pointer */
2058 comm->balanceRegion = ddBalanceRegionAllocate();
2063 /* Returns whether mtop contains constraints and/or vsites */
2064 static bool systemHasConstraintsOrVsites(const gmx_mtop_t &mtop)
2066 auto ilistLoop = gmx_mtop_ilistloop_init(mtop);
2068 while (const InteractionLists *ilists = gmx_mtop_ilistloop_next(ilistLoop, &nmol))
2070 if (!extractILists(*ilists, IF_CONSTRAINT | IF_VSITE).empty())
2079 static void setupUpdateGroups(const gmx::MDLogger &mdlog,
2080 const gmx_mtop_t &mtop,
2081 const t_inputrec &inputrec,
2082 const real cutoffMargin,
2083 DDSystemInfo *systemInfo)
2085 /* When we have constraints and/or vsites, it is beneficial to use
2086 * update groups (when possible) to allow independent update of groups.
2088 if (!systemHasConstraintsOrVsites(mtop))
2090 /* No constraints or vsites, atoms can be updated independently */
2094 systemInfo->updateGroupingPerMoleculetype = gmx::makeUpdateGroups(mtop);
2095 systemInfo->useUpdateGroups =
2096 (!systemInfo->updateGroupingPerMoleculetype.empty() &&
2097 getenv("GMX_NO_UPDATEGROUPS") == nullptr);
2099 if (systemInfo->useUpdateGroups)
2101 int numUpdateGroups = 0;
2102 for (const auto &molblock : mtop.molblock)
2104 numUpdateGroups += molblock.nmol*systemInfo->updateGroupingPerMoleculetype[molblock.type].numBlocks();
2107 systemInfo->maxUpdateGroupRadius =
2108 computeMaxUpdateGroupRadius(mtop,
2109 systemInfo->updateGroupingPerMoleculetype,
2110 maxReferenceTemperature(inputrec));
2112 /* To use update groups, the large domain-to-domain cutoff distance
2113 * should be compatible with the box size.
2115 systemInfo->useUpdateGroups = (atomToAtomIntoDomainToDomainCutoff(*systemInfo, 0) < cutoffMargin);
2117 if (systemInfo->useUpdateGroups)
2119 GMX_LOG(mdlog.info).appendTextFormatted(
2120 "Using update groups, nr %d, average size %.1f atoms, max. radius %.3f nm\n",
2122 mtop.natoms/static_cast<double>(numUpdateGroups),
2123 systemInfo->maxUpdateGroupRadius);
2127 GMX_LOG(mdlog.info).appendTextFormatted("The combination of rlist and box size prohibits the use of update groups\n");
2128 systemInfo->updateGroupingPerMoleculetype.clear();
2133 UnitCellInfo::UnitCellInfo(const t_inputrec &ir) :
2134 npbcdim(ePBC2npbcdim(ir.ePBC)),
2135 numBoundedDimensions(inputrec2nboundeddim(&ir)),
2136 ddBoxIsDynamic(numBoundedDimensions < DIM || inputrecDynamicBox(&ir)),
2137 haveScrewPBC(ir.ePBC == epbcSCREW)
2141 /*! \brief Generate the simulation system information */
2143 getSystemInfo(const gmx::MDLogger &mdlog,
2145 const DomdecOptions &options,
2146 const gmx_mtop_t *mtop,
2147 const t_inputrec *ir,
2149 gmx::ArrayRef<const gmx::RVec> xGlobal)
2151 const real tenPercentMargin = 1.1;
2153 DDSystemInfo systemInfo;
2155 /* We need to decide on update groups early, as this affects communication distances */
2156 systemInfo.useUpdateGroups = false;
2157 if (ir->cutoff_scheme == ecutsVERLET)
2159 real cutoffMargin = std::sqrt(max_cutoff2(ir->ePBC, box)) - ir->rlist;
2160 setupUpdateGroups(mdlog, *mtop, *ir, cutoffMargin, &systemInfo);
2163 // TODO: Check whether all bondeds are within update groups
2164 systemInfo.haveInterDomainBondeds = (mtop->natoms > gmx_mtop_num_molecules(*mtop) ||
2165 mtop->bIntermolecularInteractions);
2166 systemInfo.haveInterDomainMultiBodyBondeds = (multi_body_bondeds_count(mtop) > 0);
2168 if (systemInfo.useUpdateGroups)
2170 systemInfo.haveSplitConstraints = false;
2171 systemInfo.haveSplitSettles = false;
2175 systemInfo.haveSplitConstraints = gmx::inter_charge_group_constraints(*mtop);
2176 systemInfo.haveSplitSettles = gmx::inter_charge_group_settles(*mtop);
2181 /* Set the cut-off to some very large value,
2182 * so we don't need if statements everywhere in the code.
2183 * We use sqrt, since the cut-off is squared in some places.
2185 systemInfo.cutoff = GMX_CUTOFF_INF;
2189 systemInfo.cutoff = atomToAtomIntoDomainToDomainCutoff(systemInfo, ir->rlist);
2191 systemInfo.minCutoffForMultiBody = 0;
2193 /* Determine the minimum cell size limit, affected by many factors */
2194 systemInfo.cellsizeLimit = 0;
2195 systemInfo.filterBondedCommunication = false;
2197 /* We do not allow home atoms to move beyond the neighboring domain
2198 * between domain decomposition steps, which limits the cell size.
2199 * Get an estimate of cell size limit due to atom displacement.
2200 * In most cases this is a large overestimate, because it assumes
2201 * non-interaction atoms.
2202 * We set the chance to 1 in a trillion steps.
2204 constexpr real c_chanceThatAtomMovesBeyondDomain = 1e-12;
2205 const real limitForAtomDisplacement =
2206 minCellSizeForAtomDisplacement(*mtop, *ir,
2207 systemInfo.updateGroupingPerMoleculetype,
2208 c_chanceThatAtomMovesBeyondDomain);
2209 GMX_LOG(mdlog.info).appendTextFormatted(
2210 "Minimum cell size due to atom displacement: %.3f nm",
2211 limitForAtomDisplacement);
2213 systemInfo.cellsizeLimit = std::max(systemInfo.cellsizeLimit,
2214 limitForAtomDisplacement);
2216 /* TODO: PME decomposition currently requires atoms not to be more than
2217 * 2/3 of comm->cutoff, which is >=rlist, outside of their domain.
2218 * In nearly all cases, limitForAtomDisplacement will be smaller
2219 * than 2/3*rlist, so the PME requirement is satisfied.
2220 * But it would be better for both correctness and performance
2221 * to use limitForAtomDisplacement instead of 2/3*comm->cutoff.
2222 * Note that we would need to improve the pairlist buffer case.
2225 if (systemInfo.haveInterDomainBondeds)
2227 if (options.minimumCommunicationRange > 0)
2229 systemInfo.minCutoffForMultiBody =
2230 atomToAtomIntoDomainToDomainCutoff(systemInfo, options.minimumCommunicationRange);
2231 if (options.useBondedCommunication)
2233 systemInfo.filterBondedCommunication = (systemInfo.minCutoffForMultiBody > systemInfo.cutoff);
2237 systemInfo.cutoff = std::max(systemInfo.cutoff,
2238 systemInfo.minCutoffForMultiBody);
2241 else if (ir->bPeriodicMols)
2243 /* Can not easily determine the required cut-off */
2244 GMX_LOG(mdlog.warning).appendText("NOTE: Periodic molecules are present in this system. Because of this, the domain decomposition algorithm cannot easily determine the minimum cell size that it requires for treating bonded interactions. Instead, domain decomposition will assume that half the non-bonded cut-off will be a suitable lower bound.");
2245 systemInfo.minCutoffForMultiBody = systemInfo.cutoff/2;
2253 dd_bonded_cg_distance(mdlog, mtop, ir, as_rvec_array(xGlobal.data()), box,
2254 options.checkBondedInteractions,
2257 gmx_bcast(sizeof(r_2b), &r_2b, cr);
2258 gmx_bcast(sizeof(r_mb), &r_mb, cr);
2260 /* We use an initial margin of 10% for the minimum cell size,
2261 * except when we are just below the non-bonded cut-off.
2263 if (options.useBondedCommunication)
2265 if (std::max(r_2b, r_mb) > systemInfo.cutoff)
2267 const real r_bonded = std::max(r_2b, r_mb);
2268 systemInfo.minCutoffForMultiBody = tenPercentMargin*r_bonded;
2269 /* This is the (only) place where we turn on the filtering */
2270 systemInfo.filterBondedCommunication = true;
2274 const real r_bonded = r_mb;
2275 systemInfo.minCutoffForMultiBody = std::min(tenPercentMargin*r_bonded,
2278 /* We determine cutoff_mbody later */
2279 systemInfo.increaseMultiBodyCutoff = true;
2283 /* No special bonded communication,
2284 * simply increase the DD cut-off.
2286 systemInfo.minCutoffForMultiBody = tenPercentMargin*std::max(r_2b, r_mb);
2287 systemInfo.cutoff = std::max(systemInfo.cutoff,
2288 systemInfo.minCutoffForMultiBody);
2291 GMX_LOG(mdlog.info).appendTextFormatted(
2292 "Minimum cell size due to bonded interactions: %.3f nm",
2293 systemInfo.minCutoffForMultiBody);
2295 systemInfo.cellsizeLimit = std::max(systemInfo.cellsizeLimit,
2296 systemInfo.minCutoffForMultiBody);
2299 systemInfo.constraintCommunicationRange = 0;
2300 if (systemInfo.haveSplitConstraints && options.constraintCommunicationRange <= 0)
2302 /* There is a cell size limit due to the constraints (P-LINCS) */
2303 systemInfo.constraintCommunicationRange = gmx::constr_r_max(mdlog, mtop, ir);
2304 GMX_LOG(mdlog.info).appendTextFormatted(
2305 "Estimated maximum distance required for P-LINCS: %.3f nm",
2306 systemInfo.constraintCommunicationRange);
2307 if (systemInfo.constraintCommunicationRange > systemInfo.cellsizeLimit)
2309 GMX_LOG(mdlog.info).appendText("This distance will limit the DD cell size, you can override this with -rcon");
2312 else if (options.constraintCommunicationRange > 0)
2314 /* Here we do not check for dd->splitConstraints.
2315 * because one can also set a cell size limit for virtual sites only
2316 * and at this point we don't know yet if there are intercg v-sites.
2318 GMX_LOG(mdlog.info).appendTextFormatted(
2319 "User supplied maximum distance required for P-LINCS: %.3f nm",
2320 options.constraintCommunicationRange);
2321 systemInfo.constraintCommunicationRange = options.constraintCommunicationRange;
2323 systemInfo.cellsizeLimit = std::max(systemInfo.cellsizeLimit,
2324 systemInfo.constraintCommunicationRange);
2329 /*! \brief Set the cell size and interaction limits, as well as the DD grid
2331 * Also computes the initial ddbox.
2334 getDDGridSetup(const gmx::MDLogger &mdlog,
2336 const DomdecOptions &options,
2337 const DDSettings &ddSettings,
2338 const DDSystemInfo &systemInfo,
2339 const gmx_mtop_t *mtop,
2340 const t_inputrec *ir,
2342 gmx::ArrayRef<const gmx::RVec> xGlobal,
2345 DDGridSetup ddGridSetup;
2347 if (options.numCells[XX] > 0)
2349 copy_ivec(options.numCells, ddGridSetup.numDomains);
2350 set_ddbox_cr(*cr, &ddGridSetup.numDomains, *ir, box, xGlobal, ddbox);
2352 if (options.numPmeRanks >= 0)
2354 ddGridSetup.numPmeOnlyRanks = options.numPmeRanks;
2358 /* When the DD grid is set explicitly and -npme is set to auto,
2359 * don't use PME ranks. We check later if the DD grid is
2360 * compatible with the total number of ranks.
2362 ddGridSetup.numPmeOnlyRanks = 0;
2367 set_ddbox_cr(*cr, nullptr, *ir, box, xGlobal, ddbox);
2369 /* We need to choose the optimal DD grid and possibly PME nodes */
2371 dd_choose_grid(mdlog, cr, ir, mtop, box, ddbox,
2372 options.numPmeRanks,
2373 ddSettings.request1DAnd1Pulse,
2374 !isDlbDisabled(ddSettings.initialDlbState),
2378 if (ddGridSetup.numDomains[XX] == 0)
2381 gmx_bool bC = (systemInfo.haveSplitConstraints &&
2382 systemInfo.constraintCommunicationRange > systemInfo.minCutoffForMultiBody);
2383 sprintf(buf, "Change the number of ranks or mdrun option %s%s%s",
2384 !bC ? "-rdd" : "-rcon",
2385 ddSettings.initialDlbState != DlbState::offUser ? " or -dds" : "",
2386 bC ? " or your LINCS settings" : "");
2388 gmx_fatal_collective(FARGS, cr->mpi_comm_mysim, MASTER(cr),
2389 "There is no domain decomposition for %d ranks that is compatible with the given box and a minimum cell size of %g nm\n"
2391 "Look in the log file for details on the domain decomposition",
2392 cr->nnodes - ddGridSetup.numPmeOnlyRanks,
2393 ddGridSetup.cellsizeLimit,
2398 const real acs = average_cellsize_min(*ddbox, ddGridSetup.numDomains);
2399 if (acs < systemInfo.cellsizeLimit)
2401 if (options.numCells[XX] <= 0)
2403 GMX_RELEASE_ASSERT(false, "dd_choose_grid() should return a grid that satisfies the cell size limits");
2407 gmx_fatal_collective(FARGS, cr->mpi_comm_mysim, MASTER(cr),
2408 "The initial cell size (%f) is smaller than the cell size limit (%f), change options -dd, -rdd or -rcon, see the log file for details",
2409 acs, systemInfo.cellsizeLimit);
2413 const int numPPRanks = ddGridSetup.numDomains[XX]*ddGridSetup.numDomains[YY]*ddGridSetup.numDomains[ZZ];
2414 if (cr->nnodes - numPPRanks != ddGridSetup.numPmeOnlyRanks)
2416 gmx_fatal_collective(FARGS, cr->mpi_comm_mysim, MASTER(cr),
2417 "The size of the domain decomposition grid (%d) does not match the number of ranks (%d). The total number of ranks is %d",
2418 numPPRanks, cr->nnodes - ddGridSetup.numPmeOnlyRanks, cr->nnodes);
2420 if (ddGridSetup.numPmeOnlyRanks > numPPRanks)
2422 gmx_fatal_collective(FARGS, cr->mpi_comm_mysim, MASTER(cr),
2423 "The number of separate PME ranks (%d) is larger than the number of PP ranks (%d), this is not supported.", ddGridSetup.numPmeOnlyRanks, numPPRanks);
2426 ddGridSetup.numDDDimensions = set_dd_dim(ddGridSetup.numDomains, ddSettings,
2427 ddGridSetup.ddDimensions);
2432 /*! \brief Set the cell size and interaction limits, as well as the DD grid */
2434 getDDRankSetup(const gmx::MDLogger &mdlog,
2436 const DDGridSetup &ddGridSetup,
2437 const t_inputrec &ir)
2439 GMX_LOG(mdlog.info).appendTextFormatted(
2440 "Domain decomposition grid %d x %d x %d, separate PME ranks %d",
2441 ddGridSetup.numDomains[XX], ddGridSetup.numDomains[YY], ddGridSetup.numDomains[ZZ],
2442 ddGridSetup.numPmeOnlyRanks);
2444 DDRankSetup ddRankSetup;
2446 ddRankSetup.numPPRanks = cr->nnodes - ddGridSetup.numPmeOnlyRanks;
2447 copy_ivec(ddGridSetup.numDomains, ddRankSetup.numPPCells);
2449 ddRankSetup.usePmeOnlyRanks = (ddGridSetup.numPmeOnlyRanks > 0);
2450 if (ddRankSetup.usePmeOnlyRanks)
2452 ddRankSetup.numRanksDoingPme = ddGridSetup.numPmeOnlyRanks;
2456 ddRankSetup.numRanksDoingPme = ddGridSetup.numDomains[XX]*ddGridSetup.numDomains[YY]*ddGridSetup.numDomains[ZZ];
2459 if (EEL_PME(ir.coulombtype) || EVDW_PME(ir.vdwtype))
2461 /* The following choices should match those
2462 * in comm_cost_est in domdec_setup.c.
2463 * Note that here the checks have to take into account
2464 * that the decomposition might occur in a different order than xyz
2465 * (for instance through the env.var. GMX_DD_ORDER_ZYX),
2466 * in which case they will not match those in comm_cost_est,
2467 * but since that is mainly for testing purposes that's fine.
2469 if (ddGridSetup.numDDDimensions >= 2 &&
2470 ddGridSetup.ddDimensions[0] == XX &&
2471 ddGridSetup.ddDimensions[1] == YY &&
2472 ddRankSetup.numRanksDoingPme > ddGridSetup.numDomains[XX] &&
2473 ddRankSetup.numRanksDoingPme % ddGridSetup.numDomains[XX] == 0 &&
2474 getenv("GMX_PMEONEDD") == nullptr)
2476 ddRankSetup.npmedecompdim = 2;
2477 ddRankSetup.npmenodes_x = ddGridSetup.numDomains[XX];
2478 ddRankSetup.npmenodes_y = ddRankSetup.numRanksDoingPme/ddRankSetup.npmenodes_x;
2482 /* In case nc is 1 in both x and y we could still choose to
2483 * decompose pme in y instead of x, but we use x for simplicity.
2485 ddRankSetup.npmedecompdim = 1;
2486 if (ddGridSetup.ddDimensions[0] == YY)
2488 ddRankSetup.npmenodes_x = 1;
2489 ddRankSetup.npmenodes_y = ddRankSetup.numRanksDoingPme;
2493 ddRankSetup.npmenodes_x = ddRankSetup.numRanksDoingPme;
2494 ddRankSetup.npmenodes_y = 1;
2497 GMX_LOG(mdlog.info).appendTextFormatted(
2498 "PME domain decomposition: %d x %d x %d",
2499 ddRankSetup.npmenodes_x, ddRankSetup.npmenodes_y, 1);
2503 ddRankSetup.npmedecompdim = 0;
2504 ddRankSetup.npmenodes_x = 0;
2505 ddRankSetup.npmenodes_y = 0;
2511 /*! \brief Set the cell size and interaction limits */
2512 static void set_dd_limits(const gmx::MDLogger &mdlog,
2513 t_commrec *cr, gmx_domdec_t *dd,
2514 const DomdecOptions &options,
2515 const DDSettings &ddSettings,
2516 const DDSystemInfo &systemInfo,
2517 const DDGridSetup &ddGridSetup,
2518 const gmx_mtop_t *mtop,
2519 const t_inputrec *ir,
2520 const gmx_ddbox_t &ddbox)
2522 gmx_domdec_comm_t *comm = dd->comm;
2523 comm->ddSettings = ddSettings;
2525 /* Initialize to GPU share count to 0, might change later */
2526 comm->nrank_gpu_shared = 0;
2528 comm->dlbState = comm->ddSettings.initialDlbState;
2529 dd_dlb_set_should_check_whether_to_turn_dlb_on(dd, TRUE);
2530 /* To consider turning DLB on after 2*nstlist steps we need to check
2531 * at partitioning count 3. Thus we need to increase the first count by 2.
2533 comm->ddPartioningCountFirstDlbOff += 2;
2535 comm->bPMELoadBalDLBLimits = FALSE;
2537 /* Allocate the charge group/atom sorting struct */
2538 comm->sort = std::make_unique<gmx_domdec_sort_t>();
2540 comm->systemInfo = systemInfo;
2542 if (systemInfo.useUpdateGroups)
2544 /* Note: We would like to use dd->nnodes for the atom count estimate,
2545 * but that is not yet available here. But this anyhow only
2546 * affect performance up to the second dd_partition_system call.
2548 const int homeAtomCountEstimate = mtop->natoms/cr->nnodes;
2549 comm->updateGroupsCog =
2550 std::make_unique<gmx::UpdateGroupsCog>(*mtop,
2551 systemInfo.updateGroupingPerMoleculetype,
2552 maxReferenceTemperature(*ir),
2553 homeAtomCountEstimate);
2556 comm->cgs_gl = gmx_mtop_global_cgs(mtop);
2558 /* Set the DD setup given by ddGridSetup */
2559 copy_ivec(ddGridSetup.numDomains, dd->nc);
2560 dd->ndim = ddGridSetup.numDDDimensions;
2561 copy_ivec(ddGridSetup.ddDimensions, dd->dim);
2563 dd->nnodes = dd->nc[XX]*dd->nc[YY]*dd->nc[ZZ];
2565 snew(comm->slb_frac, DIM);
2566 if (isDlbDisabled(comm))
2568 comm->slb_frac[XX] = get_slb_frac(mdlog, "x", dd->nc[XX], options.cellSizeX);
2569 comm->slb_frac[YY] = get_slb_frac(mdlog, "y", dd->nc[YY], options.cellSizeY);
2570 comm->slb_frac[ZZ] = get_slb_frac(mdlog, "z", dd->nc[ZZ], options.cellSizeZ);
2573 /* Set the multi-body cut-off and cellsize limit for DLB */
2574 comm->cutoff_mbody = systemInfo.minCutoffForMultiBody;
2575 comm->cellsize_limit = systemInfo.cellsizeLimit;
2576 if (systemInfo.haveInterDomainBondeds && systemInfo.increaseMultiBodyCutoff)
2578 if (systemInfo.filterBondedCommunication || !isDlbDisabled(comm))
2580 /* Set the bonded communication distance to halfway
2581 * the minimum and the maximum,
2582 * since the extra communication cost is nearly zero.
2584 real acs = average_cellsize_min(ddbox, dd->nc);
2585 comm->cutoff_mbody = 0.5*(systemInfo.minCutoffForMultiBody + acs);
2586 if (!isDlbDisabled(comm))
2588 /* Check if this does not limit the scaling */
2589 comm->cutoff_mbody = std::min(comm->cutoff_mbody,
2590 options.dlbScaling*acs);
2592 if (!systemInfo.filterBondedCommunication)
2594 /* Without bBondComm do not go beyond the n.b. cut-off */
2595 comm->cutoff_mbody = std::min(comm->cutoff_mbody, systemInfo.cutoff);
2596 if (comm->cellsize_limit >= systemInfo.cutoff)
2598 /* We don't loose a lot of efficieny
2599 * when increasing it to the n.b. cut-off.
2600 * It can even be slightly faster, because we need
2601 * less checks for the communication setup.
2603 comm->cutoff_mbody = systemInfo.cutoff;
2606 /* Check if we did not end up below our original limit */
2607 comm->cutoff_mbody = std::max(comm->cutoff_mbody,
2608 systemInfo.minCutoffForMultiBody);
2610 if (comm->cutoff_mbody > comm->cellsize_limit)
2612 comm->cellsize_limit = comm->cutoff_mbody;
2615 /* Without DLB and cutoff_mbody<cutoff, cutoff_mbody is dynamic */
2620 fprintf(debug, "Bonded atom communication beyond the cut-off: %s\n"
2621 "cellsize limit %f\n",
2622 gmx::boolToString(systemInfo.filterBondedCommunication),
2623 comm->cellsize_limit);
2628 check_dd_restrictions(dd, ir, mdlog);
2632 static char *init_bLocalCG(const gmx_mtop_t *mtop)
2637 ncg = ncg_mtop(mtop);
2638 snew(bLocalCG, ncg);
2639 for (cg = 0; cg < ncg; cg++)
2641 bLocalCG[cg] = FALSE;
2647 void dd_init_bondeds(FILE *fplog,
2649 const gmx_mtop_t *mtop,
2650 const gmx_vsite_t *vsite,
2651 const t_inputrec *ir,
2652 gmx_bool bBCheck, cginfo_mb_t *cginfo_mb)
2654 gmx_domdec_comm_t *comm;
2656 dd_make_reverse_top(fplog, dd, mtop, vsite, ir, bBCheck);
2660 if (comm->systemInfo.filterBondedCommunication)
2662 /* Communicate atoms beyond the cut-off for bonded interactions */
2665 comm->cglink = make_charge_group_links(mtop, dd, cginfo_mb);
2667 comm->bLocalCG = init_bLocalCG(mtop);
2671 /* Only communicate atoms based on cut-off */
2672 comm->cglink = nullptr;
2673 comm->bLocalCG = nullptr;
2677 static void writeSettings(gmx::TextWriter *log,
2679 const gmx_mtop_t *mtop,
2680 const t_inputrec *ir,
2681 gmx_bool bDynLoadBal,
2683 const gmx_ddbox_t *ddbox)
2685 gmx_domdec_comm_t *comm;
2694 log->writeString("The maximum number of communication pulses is:");
2695 for (d = 0; d < dd->ndim; d++)
2697 log->writeStringFormatted(" %c %d", dim2char(dd->dim[d]), comm->cd[d].np_dlb);
2699 log->ensureLineBreak();
2700 log->writeLineFormatted("The minimum size for domain decomposition cells is %.3f nm", comm->cellsize_limit);
2701 log->writeLineFormatted("The requested allowed shrink of DD cells (option -dds) is: %.2f", dlb_scale);
2702 log->writeString("The allowed shrink of domain decomposition cells is:");
2703 for (d = 0; d < DIM; d++)
2707 if (d >= ddbox->npbcdim && dd->nc[d] == 2)
2714 comm->cellsize_min_dlb[d]/
2715 (ddbox->box_size[d]*ddbox->skew_fac[d]/dd->nc[d]);
2717 log->writeStringFormatted(" %c %.2f", dim2char(d), shrink);
2720 log->ensureLineBreak();
2724 set_dd_cell_sizes_slb(dd, ddbox, setcellsizeslbPULSE_ONLY, np);
2725 log->writeString("The initial number of communication pulses is:");
2726 for (d = 0; d < dd->ndim; d++)
2728 log->writeStringFormatted(" %c %d", dim2char(dd->dim[d]), np[dd->dim[d]]);
2730 log->ensureLineBreak();
2731 log->writeString("The initial domain decomposition cell size is:");
2732 for (d = 0; d < DIM; d++)
2736 log->writeStringFormatted(" %c %.2f nm",
2737 dim2char(d), dd->comm->cellsize_min[d]);
2740 log->ensureLineBreak();
2744 const bool haveInterDomainVsites =
2745 (countInterUpdategroupVsites(*mtop, comm->systemInfo.updateGroupingPerMoleculetype) != 0);
2747 if (comm->systemInfo.haveInterDomainBondeds ||
2748 haveInterDomainVsites ||
2749 comm->systemInfo.haveSplitConstraints ||
2750 comm->systemInfo.haveSplitSettles)
2752 std::string decompUnits;
2753 if (comm->systemInfo.useUpdateGroups)
2755 decompUnits = "atom groups";
2759 decompUnits = "atoms";
2762 log->writeLineFormatted("The maximum allowed distance for %s involved in interactions is:", decompUnits.c_str());
2763 log->writeLineFormatted("%40s %-7s %6.3f nm", "non-bonded interactions", "", comm->systemInfo.cutoff);
2767 limit = dd->comm->cellsize_limit;
2771 if (dd->unitCellInfo.ddBoxIsDynamic)
2773 log->writeLine("(the following are initial values, they could change due to box deformation)");
2775 limit = dd->comm->cellsize_min[XX];
2776 for (d = 1; d < DIM; d++)
2778 limit = std::min(limit, dd->comm->cellsize_min[d]);
2782 if (comm->systemInfo.haveInterDomainBondeds)
2784 log->writeLineFormatted("%40s %-7s %6.3f nm",
2785 "two-body bonded interactions", "(-rdd)",
2786 std::max(comm->systemInfo.cutoff, comm->cutoff_mbody));
2787 log->writeLineFormatted("%40s %-7s %6.3f nm",
2788 "multi-body bonded interactions", "(-rdd)",
2789 (comm->systemInfo.filterBondedCommunication || isDlbOn(dd->comm)) ? comm->cutoff_mbody : std::min(comm->systemInfo.cutoff, limit));
2791 if (haveInterDomainVsites)
2793 log->writeLineFormatted("%40s %-7s %6.3f nm",
2794 "virtual site constructions", "(-rcon)", limit);
2796 if (comm->systemInfo.haveSplitConstraints || comm->systemInfo.haveSplitSettles)
2798 std::string separation = gmx::formatString("atoms separated by up to %d constraints",
2800 log->writeLineFormatted("%40s %-7s %6.3f nm\n",
2801 separation.c_str(), "(-rcon)", limit);
2803 log->ensureLineBreak();
2807 static void logSettings(const gmx::MDLogger &mdlog,
2809 const gmx_mtop_t *mtop,
2810 const t_inputrec *ir,
2812 const gmx_ddbox_t *ddbox)
2814 gmx::StringOutputStream stream;
2815 gmx::TextWriter log(&stream);
2816 writeSettings(&log, dd, mtop, ir, isDlbOn(dd->comm), dlb_scale, ddbox);
2817 if (dd->comm->dlbState == DlbState::offCanTurnOn)
2820 log.ensureEmptyLine();
2821 log.writeLine("When dynamic load balancing gets turned on, these settings will change to:");
2823 writeSettings(&log, dd, mtop, ir, true, dlb_scale, ddbox);
2825 GMX_LOG(mdlog.info).asParagraph().appendText(stream.toString());
2828 static void set_cell_limits_dlb(const gmx::MDLogger &mdlog,
2831 const t_inputrec *ir,
2832 const gmx_ddbox_t *ddbox)
2834 gmx_domdec_comm_t *comm;
2835 int d, dim, npulse, npulse_d_max, npulse_d;
2840 bNoCutOff = (ir->rvdw == 0 || ir->rcoulomb == 0);
2842 /* Determine the maximum number of comm. pulses in one dimension */
2844 comm->cellsize_limit = std::max(comm->cellsize_limit, comm->cutoff_mbody);
2846 /* Determine the maximum required number of grid pulses */
2847 if (comm->cellsize_limit >= comm->systemInfo.cutoff)
2849 /* Only a single pulse is required */
2852 else if (!bNoCutOff && comm->cellsize_limit > 0)
2854 /* We round down slightly here to avoid overhead due to the latency
2855 * of extra communication calls when the cut-off
2856 * would be only slightly longer than the cell size.
2857 * Later cellsize_limit is redetermined,
2858 * so we can not miss interactions due to this rounding.
2860 npulse = static_cast<int>(0.96 + comm->systemInfo.cutoff/comm->cellsize_limit);
2864 /* There is no cell size limit */
2865 npulse = std::max(dd->nc[XX]-1, std::max(dd->nc[YY]-1, dd->nc[ZZ]-1));
2868 if (!bNoCutOff && npulse > 1)
2870 /* See if we can do with less pulses, based on dlb_scale */
2872 for (d = 0; d < dd->ndim; d++)
2875 npulse_d = static_cast<int>(1 + dd->nc[dim]*comm->systemInfo.cutoff
2876 /(ddbox->box_size[dim]*ddbox->skew_fac[dim]*dlb_scale));
2877 npulse_d_max = std::max(npulse_d_max, npulse_d);
2879 npulse = std::min(npulse, npulse_d_max);
2882 /* This env var can override npulse */
2883 d = dd_getenv(mdlog, "GMX_DD_NPULSE", 0);
2890 comm->bVacDLBNoLimit = (ir->ePBC == epbcNONE);
2891 for (d = 0; d < dd->ndim; d++)
2893 if (comm->ddSettings.request1DAnd1Pulse)
2895 comm->cd[d].np_dlb = 1;
2899 comm->cd[d].np_dlb = std::min(npulse, dd->nc[dd->dim[d]]-1);
2900 comm->maxpulse = std::max(comm->maxpulse, comm->cd[d].np_dlb);
2902 if (comm->cd[d].np_dlb < dd->nc[dd->dim[d]]-1)
2904 comm->bVacDLBNoLimit = FALSE;
2908 /* cellsize_limit is set for LINCS in init_domain_decomposition */
2909 if (!comm->bVacDLBNoLimit)
2911 comm->cellsize_limit = std::max(comm->cellsize_limit,
2912 comm->systemInfo.cutoff/comm->maxpulse);
2914 comm->cellsize_limit = std::max(comm->cellsize_limit, comm->cutoff_mbody);
2915 /* Set the minimum cell size for each DD dimension */
2916 for (d = 0; d < dd->ndim; d++)
2918 if (comm->bVacDLBNoLimit ||
2919 comm->cd[d].np_dlb*comm->cellsize_limit >= comm->systemInfo.cutoff)
2921 comm->cellsize_min_dlb[dd->dim[d]] = comm->cellsize_limit;
2925 comm->cellsize_min_dlb[dd->dim[d]] =
2926 comm->systemInfo.cutoff/comm->cd[d].np_dlb;
2929 if (comm->cutoff_mbody <= 0)
2931 comm->cutoff_mbody = std::min(comm->systemInfo.cutoff, comm->cellsize_limit);
2939 gmx_bool dd_bonded_molpbc(const gmx_domdec_t *dd, int ePBC)
2941 /* If each molecule is a single charge group
2942 * or we use domain decomposition for each periodic dimension,
2943 * we do not need to take pbc into account for the bonded interactions.
2945 return (ePBC != epbcNONE && dd->comm->systemInfo.haveInterDomainBondeds &&
2948 (dd->nc[ZZ] > 1 || ePBC == epbcXY)));
2951 /*! \brief Sets grid size limits and PP-PME setup, prints settings to log */
2952 static void set_ddgrid_parameters(const gmx::MDLogger &mdlog,
2953 gmx_domdec_t *dd, real dlb_scale,
2954 const gmx_mtop_t *mtop, const t_inputrec *ir,
2955 const gmx_ddbox_t *ddbox)
2957 gmx_domdec_comm_t *comm = dd->comm;
2958 DDRankSetup &ddRankSetup = comm->ddRankSetup;
2960 if (EEL_PME(ir->coulombtype) || EVDW_PME(ir->vdwtype))
2962 init_ddpme(dd, &ddRankSetup.ddpme[0], 0);
2963 if (ddRankSetup.npmedecompdim >= 2)
2965 init_ddpme(dd, &ddRankSetup.ddpme[1], 1);
2970 ddRankSetup.numRanksDoingPme = 0;
2971 if (dd->pme_nodeid >= 0)
2973 gmx_fatal_collective(FARGS, dd->mpi_comm_all, DDMASTER(dd),
2974 "Can not have separate PME ranks without PME electrostatics");
2980 fprintf(debug, "The DD cut-off is %f\n", comm->systemInfo.cutoff);
2982 if (!isDlbDisabled(comm))
2984 set_cell_limits_dlb(mdlog, dd, dlb_scale, ir, ddbox);
2987 logSettings(mdlog, dd, mtop, ir, dlb_scale, ddbox);
2990 if (ir->ePBC == epbcNONE)
2992 vol_frac = 1 - 1/static_cast<double>(dd->nnodes);
2997 (1 + comm_box_frac(dd->nc, comm->systemInfo.cutoff, ddbox))/static_cast<double>(dd->nnodes);
3001 fprintf(debug, "Volume fraction for all DD zones: %f\n", vol_frac);
3003 int natoms_tot = comm->cgs_gl.index[comm->cgs_gl.nr];
3005 dd->ga2la = new gmx_ga2la_t(natoms_tot,
3006 static_cast<int>(vol_frac*natoms_tot));
3009 /*! \brief Get some important DD parameters which can be modified by env.vars */
3011 getDDSettings(const gmx::MDLogger &mdlog,
3012 const DomdecOptions &options,
3013 const gmx::MdrunOptions &mdrunOptions,
3014 const t_inputrec &ir)
3016 DDSettings ddSettings;
3018 ddSettings.useSendRecv2 = (dd_getenv(mdlog, "GMX_DD_USE_SENDRECV2", 0) != 0);
3019 ddSettings.dlb_scale_lim = dd_getenv(mdlog, "GMX_DLB_MAX_BOX_SCALING", 10);
3020 // TODO GPU halo exchange requires a 1D single-pulse DD, and when
3021 // it is properly integrated the hack with GMX_GPU_DD_COMMS should
3023 ddSettings.request1DAnd1Pulse = (bool(dd_getenv(mdlog, "GMX_DD_1D_1PULSE", 0)) ||
3024 (bool(getenv("GMX_GPU_DD_COMMS") != nullptr &&
3026 (GMX_GPU == GMX_GPU_CUDA))));
3027 ddSettings.useDDOrderZYX = bool(dd_getenv(mdlog, "GMX_DD_ORDER_ZYX", 0));
3028 ddSettings.useCartesianReorder = bool(dd_getenv(mdlog, "GMX_NO_CART_REORDER", 1));
3029 ddSettings.eFlop = dd_getenv(mdlog, "GMX_DLB_BASED_ON_FLOPS", 0);
3030 const int recload = dd_getenv(mdlog, "GMX_DD_RECORD_LOAD", 1);
3031 ddSettings.nstDDDump = dd_getenv(mdlog, "GMX_DD_NST_DUMP", 0);
3032 ddSettings.nstDDDumpGrid = dd_getenv(mdlog, "GMX_DD_NST_DUMP_GRID", 0);
3033 ddSettings.DD_debug = dd_getenv(mdlog, "GMX_DD_DEBUG", 0);
3035 if (ddSettings.useSendRecv2)
3037 GMX_LOG(mdlog.info).appendText("Will use two sequential MPI_Sendrecv calls instead of two simultaneous non-blocking MPI_Irecv and MPI_Isend pairs for constraint and vsite communication");
3040 if (ddSettings.eFlop)
3042 GMX_LOG(mdlog.info).appendText("Will load balance based on FLOP count");
3043 ddSettings.recordLoad = true;
3047 ddSettings.recordLoad = (wallcycle_have_counter() && recload > 0);
3050 ddSettings.initialDlbState =
3051 determineInitialDlbState(mdlog, options.dlbOption, ddSettings.recordLoad, mdrunOptions, &ir);
3052 GMX_LOG(mdlog.info).appendTextFormatted("Dynamic load balancing: %s",
3053 edlbs_names[static_cast<int>(ddSettings.initialDlbState)]);
3058 gmx_domdec_t::gmx_domdec_t(const t_inputrec &ir) :
3063 gmx_domdec_t *init_domain_decomposition(const gmx::MDLogger &mdlog,
3065 const DomdecOptions &options,
3066 const gmx::MdrunOptions &mdrunOptions,
3067 const gmx_mtop_t *mtop,
3068 const t_inputrec *ir,
3070 gmx::ArrayRef<const gmx::RVec> xGlobal,
3071 gmx::LocalAtomSetManager *atomSets)
3073 GMX_LOG(mdlog.info).appendTextFormatted(
3074 "\nInitializing Domain Decomposition on %d ranks", cr->nnodes);
3076 DDSettings ddSettings = getDDSettings(mdlog, options, mdrunOptions, *ir);
3077 if (ddSettings.eFlop > 1)
3079 /* Ensure that we have different random flop counts on different ranks */
3080 srand(1 + cr->nodeid);
3083 DDSystemInfo systemInfo = getSystemInfo(mdlog, cr, options, mtop, ir, box, xGlobal);
3085 gmx_ddbox_t ddbox = {0};
3086 DDGridSetup ddGridSetup = getDDGridSetup(mdlog, cr, options, ddSettings, systemInfo,
3087 mtop, ir, box, xGlobal, &ddbox);
3089 cr->npmenodes = ddGridSetup.numPmeOnlyRanks;
3091 DDRankSetup ddRankSetup = getDDRankSetup(mdlog, cr, ddGridSetup, *ir);
3093 /* Generate the group communicator, also decides the duty of each rank */
3094 ivec ddCellIndex = { 0, 0, 0 };
3095 std::vector<int> pmeRanks;
3096 CartesianRankSetup cartSetup =
3097 makeGroupCommunicators(mdlog, ddSettings, options.rankOrder,
3099 ddCellIndex, &pmeRanks);
3101 gmx_domdec_t *dd = new gmx_domdec_t(*ir);
3103 copy_ivec(ddCellIndex, dd->ci);
3105 dd->comm = init_dd_comm();
3107 dd->comm->ddRankSetup = ddRankSetup;
3108 dd->comm->cartesianRankSetup = cartSetup;
3110 set_dd_limits(mdlog, cr, dd, options,
3111 ddSettings, systemInfo, ddGridSetup,
3115 setupGroupCommunication(mdlog, ddSettings, pmeRanks, cr, dd);
3117 if (thisRankHasDuty(cr, DUTY_PP))
3119 set_ddgrid_parameters(mdlog, dd, options.dlbScaling, mtop, ir, &ddbox);
3121 setup_neighbor_relations(dd);
3124 /* Set overallocation to avoid frequent reallocation of arrays */
3125 set_over_alloc_dd(TRUE);
3127 dd->atomSets = atomSets;
3132 static gmx_bool test_dd_cutoff(t_commrec *cr,
3134 gmx::ArrayRef<const gmx::RVec> x,
3135 real cutoffRequested)
3145 set_ddbox(*dd, false, box, true, x, &ddbox);
3149 for (d = 0; d < dd->ndim; d++)
3153 inv_cell_size = DD_CELL_MARGIN*dd->nc[dim]/ddbox.box_size[dim];
3154 if (dd->unitCellInfo.ddBoxIsDynamic)
3156 inv_cell_size *= DD_PRES_SCALE_MARGIN;
3159 np = 1 + static_cast<int>(cutoffRequested*inv_cell_size*ddbox.skew_fac[dim]);
3161 if (!isDlbDisabled(dd->comm) && (dim < ddbox.npbcdim) && (dd->comm->cd[d].np_dlb > 0))
3163 if (np > dd->comm->cd[d].np_dlb)
3168 /* If a current local cell size is smaller than the requested
3169 * cut-off, we could still fix it, but this gets very complicated.
3170 * Without fixing here, we might actually need more checks.
3172 real cellSizeAlongDim = (dd->comm->cell_x1[dim] - dd->comm->cell_x0[dim])*ddbox.skew_fac[dim];
3173 if (cellSizeAlongDim*dd->comm->cd[d].np_dlb < cutoffRequested)
3180 if (!isDlbDisabled(dd->comm))
3182 /* If DLB is not active yet, we don't need to check the grid jumps.
3183 * Actually we shouldn't, because then the grid jump data is not set.
3185 if (isDlbOn(dd->comm) &&
3186 check_grid_jump(0, dd, cutoffRequested, &ddbox, FALSE))
3191 gmx_sumi(1, &LocallyLimited, cr);
3193 if (LocallyLimited > 0)
3202 gmx_bool change_dd_cutoff(t_commrec *cr,
3204 gmx::ArrayRef<const gmx::RVec> x,
3205 real cutoffRequested)
3207 gmx_bool bCutoffAllowed;
3209 bCutoffAllowed = test_dd_cutoff(cr, box, x, cutoffRequested);
3213 cr->dd->comm->systemInfo.cutoff = cutoffRequested;
3216 return bCutoffAllowed;