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/options.h"
59 #include "gromacs/domdec/partition.h"
60 #include "gromacs/gmxlib/network.h"
61 #include "gromacs/gmxlib/nrnb.h"
62 #include "gromacs/gpu_utils/gpu_utils.h"
63 #include "gromacs/hardware/hw_info.h"
64 #include "gromacs/listed_forces/manage_threading.h"
65 #include "gromacs/math/vec.h"
66 #include "gromacs/math/vectypes.h"
67 #include "gromacs/mdlib/calc_verletbuf.h"
68 #include "gromacs/mdlib/constr.h"
69 #include "gromacs/mdlib/constraintrange.h"
70 #include "gromacs/mdlib/updategroups.h"
71 #include "gromacs/mdlib/vsite.h"
72 #include "gromacs/mdtypes/commrec.h"
73 #include "gromacs/mdtypes/forceoutput.h"
74 #include "gromacs/mdtypes/inputrec.h"
75 #include "gromacs/mdtypes/mdrunoptions.h"
76 #include "gromacs/mdtypes/state.h"
77 #include "gromacs/pbcutil/ishift.h"
78 #include "gromacs/pbcutil/pbc.h"
79 #include "gromacs/pulling/pull.h"
80 #include "gromacs/timing/wallcycle.h"
81 #include "gromacs/topology/block.h"
82 #include "gromacs/topology/idef.h"
83 #include "gromacs/topology/ifunc.h"
84 #include "gromacs/topology/mtop_lookup.h"
85 #include "gromacs/topology/mtop_util.h"
86 #include "gromacs/topology/topology.h"
87 #include "gromacs/utility/basedefinitions.h"
88 #include "gromacs/utility/basenetwork.h"
89 #include "gromacs/utility/cstringutil.h"
90 #include "gromacs/utility/exceptions.h"
91 #include "gromacs/utility/fatalerror.h"
92 #include "gromacs/utility/gmxmpi.h"
93 #include "gromacs/utility/logger.h"
94 #include "gromacs/utility/real.h"
95 #include "gromacs/utility/smalloc.h"
96 #include "gromacs/utility/strconvert.h"
97 #include "gromacs/utility/stringstream.h"
98 #include "gromacs/utility/stringutil.h"
99 #include "gromacs/utility/textwriter.h"
101 #include "atomdistribution.h"
103 #include "cellsizes.h"
104 #include "distribute.h"
105 #include "domdec_constraints.h"
106 #include "domdec_internal.h"
107 #include "domdec_vsite.h"
108 #include "redistribute.h"
111 // TODO remove this when moving domdec into gmx namespace
112 using gmx::DdRankOrder;
113 using gmx::DlbOption;
114 using gmx::DomdecOptions;
116 static const char *edlbs_names[int(DlbState::nr)] = { "off", "auto", "locked", "on", "on" };
118 /* The size per atom group of the cggl_flag buffer in gmx_domdec_comm_t */
121 /* The flags for the cggl_flag buffer in gmx_domdec_comm_t */
122 #define DD_FLAG_NRCG 65535
123 #define DD_FLAG_FW(d) (1<<(16+(d)*2))
124 #define DD_FLAG_BW(d) (1<<(16+(d)*2+1))
126 /* The DD zone order */
127 static const ivec dd_zo[DD_MAXZONE] =
128 {{0, 0, 0}, {1, 0, 0}, {1, 1, 0}, {0, 1, 0}, {0, 1, 1}, {0, 0, 1}, {1, 0, 1}, {1, 1, 1}};
130 /* The non-bonded zone-pair setup for domain decomposition
131 * The first number is the i-zone, the second number the first j-zone seen by
132 * this i-zone, the third number the last+1 j-zone seen by this i-zone.
133 * As is, this is for 3D decomposition, where there are 4 i-zones.
134 * With 2D decomposition use only the first 2 i-zones and a last+1 j-zone of 4.
135 * With 1D decomposition use only the first i-zone and a last+1 j-zone of 2.
138 ddNonbondedZonePairRanges[DD_MAXIZONE][3] = {{0, 0, 8},
145 #define dd_index(n,i) ((((i)[ZZ]*(n)[YY] + (i)[YY])*(n)[XX]) + (i)[XX])
147 static void index2xyz(ivec nc,int ind,ivec xyz)
149 xyz[XX] = ind % nc[XX];
150 xyz[YY] = (ind / nc[XX]) % nc[YY];
151 xyz[ZZ] = ind / (nc[YY]*nc[XX]);
155 static void ddindex2xyz(const ivec nc, int ind, ivec xyz)
157 xyz[XX] = ind / (nc[YY]*nc[ZZ]);
158 xyz[YY] = (ind / nc[ZZ]) % nc[YY];
159 xyz[ZZ] = ind % nc[ZZ];
162 static int ddcoord2ddnodeid(gmx_domdec_t *dd, ivec c)
167 ddindex = dd_index(dd->nc, c);
168 if (dd->comm->bCartesianPP_PME)
170 ddnodeid = dd->comm->ddindex2ddnodeid[ddindex];
172 else if (dd->comm->bCartesianPP)
175 MPI_Cart_rank(dd->mpi_comm_all, c, &ddnodeid);
186 int ddglatnr(const gmx_domdec_t *dd, int i)
196 if (i >= dd->comm->atomRanges.numAtomsTotal())
198 gmx_fatal(FARGS, "glatnr called with %d, which is larger than the local number of atoms (%d)", i, dd->comm->atomRanges.numAtomsTotal());
200 atnr = dd->globalAtomIndices[i] + 1;
206 t_block *dd_charge_groups_global(gmx_domdec_t *dd)
208 return &dd->comm->cgs_gl;
211 gmx::ArrayRef<const gmx::RangePartitioning> getUpdateGroupingPerMoleculetype(const gmx_domdec_t &dd)
213 GMX_RELEASE_ASSERT(dd.comm, "Need a valid dd.comm");
214 return dd.comm->systemInfo.updateGroupingPerMoleculetype;
217 void dd_store_state(gmx_domdec_t *dd, t_state *state)
221 if (state->ddp_count != dd->ddp_count)
223 gmx_incons("The MD state does not match the domain decomposition state");
226 state->cg_gl.resize(dd->ncg_home);
227 for (i = 0; i < dd->ncg_home; i++)
229 state->cg_gl[i] = dd->globalAtomGroupIndices[i];
232 state->ddp_count_cg_gl = dd->ddp_count;
235 gmx_domdec_zones_t *domdec_zones(gmx_domdec_t *dd)
237 return &dd->comm->zones;
240 void dd_get_ns_ranges(const gmx_domdec_t *dd, int icg,
241 int *jcg0, int *jcg1, ivec shift0, ivec shift1)
243 gmx_domdec_zones_t *zones;
246 zones = &dd->comm->zones;
249 while (icg >= zones->izone[izone].cg1)
258 else if (izone < zones->nizone)
260 *jcg0 = zones->izone[izone].jcg0;
264 gmx_fatal(FARGS, "DD icg %d out of range: izone (%d) >= nizone (%d)",
265 icg, izone, zones->nizone);
268 *jcg1 = zones->izone[izone].jcg1;
270 for (d = 0; d < dd->ndim; d++)
273 shift0[dim] = zones->izone[izone].shift0[dim];
274 shift1[dim] = zones->izone[izone].shift1[dim];
275 if (dd->comm->tric_dir[dim] || (isDlbOn(dd->comm) && d > 0))
277 /* A conservative approach, this can be optimized */
284 int dd_numHomeAtoms(const gmx_domdec_t &dd)
286 return dd.comm->atomRanges.numHomeAtoms();
289 int dd_natoms_mdatoms(const gmx_domdec_t *dd)
291 /* We currently set mdatoms entries for all atoms:
292 * local + non-local + communicated for vsite + constraints
295 return dd->comm->atomRanges.numAtomsTotal();
298 int dd_natoms_vsite(const gmx_domdec_t *dd)
300 return dd->comm->atomRanges.end(DDAtomRanges::Type::Vsites);
303 void dd_get_constraint_range(const gmx_domdec_t *dd, int *at_start, int *at_end)
305 *at_start = dd->comm->atomRanges.start(DDAtomRanges::Type::Constraints);
306 *at_end = dd->comm->atomRanges.end(DDAtomRanges::Type::Constraints);
309 void dd_move_x(gmx_domdec_t *dd,
311 gmx::ArrayRef<gmx::RVec> x,
312 gmx_wallcycle *wcycle)
314 wallcycle_start(wcycle, ewcMOVEX);
317 gmx_domdec_comm_t *comm;
318 gmx_domdec_comm_dim_t *cd;
319 rvec shift = {0, 0, 0};
320 gmx_bool bPBC, bScrew;
325 nat_tot = comm->atomRanges.numHomeAtoms();
326 for (int d = 0; d < dd->ndim; d++)
328 bPBC = (dd->ci[dd->dim[d]] == 0);
329 bScrew = (bPBC && dd->unitCellInfo.haveScrewPBC && dd->dim[d] == XX);
332 copy_rvec(box[dd->dim[d]], shift);
335 for (const gmx_domdec_ind_t &ind : cd->ind)
337 DDBufferAccess<gmx::RVec> sendBufferAccess(comm->rvecBuffer, ind.nsend[nzone + 1]);
338 gmx::ArrayRef<gmx::RVec> &sendBuffer = sendBufferAccess.buffer;
342 for (int j : ind.index)
344 sendBuffer[n] = x[j];
350 for (int j : ind.index)
352 /* We need to shift the coordinates */
353 for (int d = 0; d < DIM; d++)
355 sendBuffer[n][d] = x[j][d] + shift[d];
362 for (int j : ind.index)
365 sendBuffer[n][XX] = x[j][XX] + shift[XX];
367 * This operation requires a special shift force
368 * treatment, which is performed in calc_vir.
370 sendBuffer[n][YY] = box[YY][YY] - x[j][YY];
371 sendBuffer[n][ZZ] = box[ZZ][ZZ] - x[j][ZZ];
376 DDBufferAccess<gmx::RVec> receiveBufferAccess(comm->rvecBuffer2, cd->receiveInPlace ? 0 : ind.nrecv[nzone + 1]);
378 gmx::ArrayRef<gmx::RVec> receiveBuffer;
379 if (cd->receiveInPlace)
381 receiveBuffer = gmx::arrayRefFromArray(x.data() + nat_tot, ind.nrecv[nzone + 1]);
385 receiveBuffer = receiveBufferAccess.buffer;
387 /* Send and receive the coordinates */
388 ddSendrecv(dd, d, dddirBackward,
389 sendBuffer, receiveBuffer);
391 if (!cd->receiveInPlace)
394 for (int zone = 0; zone < nzone; zone++)
396 for (int i = ind.cell2at0[zone]; i < ind.cell2at1[zone]; i++)
398 x[i] = receiveBuffer[j++];
402 nat_tot += ind.nrecv[nzone+1];
407 wallcycle_stop(wcycle, ewcMOVEX);
410 void dd_move_f(gmx_domdec_t *dd,
411 gmx::ForceWithShiftForces *forceWithShiftForces,
412 gmx_wallcycle *wcycle)
414 wallcycle_start(wcycle, ewcMOVEF);
416 gmx::ArrayRef<gmx::RVec> f = forceWithShiftForces->force();
417 gmx::ArrayRef<gmx::RVec> fshift = forceWithShiftForces->shiftForces();
419 gmx_domdec_comm_t &comm = *dd->comm;
420 int nzone = comm.zones.n/2;
421 int nat_tot = comm.atomRanges.end(DDAtomRanges::Type::Zones);
422 for (int d = dd->ndim-1; d >= 0; d--)
424 /* Only forces in domains near the PBC boundaries need to
425 consider PBC in the treatment of fshift */
426 const bool shiftForcesNeedPbc = (forceWithShiftForces->computeVirial() && dd->ci[dd->dim[d]] == 0);
427 const bool applyScrewPbc = (shiftForcesNeedPbc && dd->unitCellInfo.haveScrewPBC && dd->dim[d] == XX);
428 /* Determine which shift vector we need */
429 ivec vis = { 0, 0, 0 };
431 const int is = IVEC2IS(vis);
433 /* Loop over the pulses */
434 const gmx_domdec_comm_dim_t &cd = comm.cd[d];
435 for (int p = cd.numPulses() - 1; p >= 0; p--)
437 const gmx_domdec_ind_t &ind = cd.ind[p];
438 DDBufferAccess<gmx::RVec> receiveBufferAccess(comm.rvecBuffer, ind.nsend[nzone + 1]);
439 gmx::ArrayRef<gmx::RVec> &receiveBuffer = receiveBufferAccess.buffer;
441 nat_tot -= ind.nrecv[nzone+1];
443 DDBufferAccess<gmx::RVec> sendBufferAccess(comm.rvecBuffer2, cd.receiveInPlace ? 0 : ind.nrecv[nzone + 1]);
445 gmx::ArrayRef<gmx::RVec> sendBuffer;
446 if (cd.receiveInPlace)
448 sendBuffer = gmx::arrayRefFromArray(f.data() + nat_tot, ind.nrecv[nzone + 1]);
452 sendBuffer = sendBufferAccess.buffer;
454 for (int zone = 0; zone < nzone; zone++)
456 for (int i = ind.cell2at0[zone]; i < ind.cell2at1[zone]; i++)
458 sendBuffer[j++] = f[i];
462 /* Communicate the forces */
463 ddSendrecv(dd, d, dddirForward,
464 sendBuffer, receiveBuffer);
465 /* Add the received forces */
467 if (!shiftForcesNeedPbc)
469 for (int j : ind.index)
471 for (int d = 0; d < DIM; d++)
473 f[j][d] += receiveBuffer[n][d];
478 else if (!applyScrewPbc)
480 for (int j : ind.index)
482 for (int d = 0; d < DIM; d++)
484 f[j][d] += receiveBuffer[n][d];
486 /* Add this force to the shift force */
487 for (int d = 0; d < DIM; d++)
489 fshift[is][d] += receiveBuffer[n][d];
496 for (int j : ind.index)
498 /* Rotate the force */
499 f[j][XX] += receiveBuffer[n][XX];
500 f[j][YY] -= receiveBuffer[n][YY];
501 f[j][ZZ] -= receiveBuffer[n][ZZ];
502 if (shiftForcesNeedPbc)
504 /* Add this force to the shift force */
505 for (int d = 0; d < DIM; d++)
507 fshift[is][d] += receiveBuffer[n][d];
516 wallcycle_stop(wcycle, ewcMOVEF);
519 /* Convenience function for extracting a real buffer from an rvec buffer
521 * To reduce the number of temporary communication buffers and avoid
522 * cache polution, we reuse gmx::RVec buffers for storing reals.
523 * This functions return a real buffer reference with the same number
524 * of elements as the gmx::RVec buffer (so 1/3 of the size in bytes).
526 static gmx::ArrayRef<real>
527 realArrayRefFromRvecArrayRef(gmx::ArrayRef<gmx::RVec> arrayRef)
529 return gmx::arrayRefFromArray(as_rvec_array(arrayRef.data())[0],
533 void dd_atom_spread_real(gmx_domdec_t *dd, real v[])
536 gmx_domdec_comm_t *comm;
537 gmx_domdec_comm_dim_t *cd;
542 nat_tot = comm->atomRanges.numHomeAtoms();
543 for (int d = 0; d < dd->ndim; d++)
546 for (const gmx_domdec_ind_t &ind : cd->ind)
548 /* Note: We provision for RVec instead of real, so a factor of 3
549 * more than needed. The buffer actually already has this size
550 * and we pass a plain pointer below, so this does not matter.
552 DDBufferAccess<gmx::RVec> sendBufferAccess(comm->rvecBuffer, ind.nsend[nzone + 1]);
553 gmx::ArrayRef<real> sendBuffer = realArrayRefFromRvecArrayRef(sendBufferAccess.buffer);
555 for (int j : ind.index)
557 sendBuffer[n++] = v[j];
560 DDBufferAccess<gmx::RVec> receiveBufferAccess(comm->rvecBuffer2, cd->receiveInPlace ? 0 : ind.nrecv[nzone + 1]);
562 gmx::ArrayRef<real> receiveBuffer;
563 if (cd->receiveInPlace)
565 receiveBuffer = gmx::arrayRefFromArray(v + nat_tot, ind.nrecv[nzone + 1]);
569 receiveBuffer = realArrayRefFromRvecArrayRef(receiveBufferAccess.buffer);
571 /* Send and receive the data */
572 ddSendrecv(dd, d, dddirBackward,
573 sendBuffer, receiveBuffer);
574 if (!cd->receiveInPlace)
577 for (int zone = 0; zone < nzone; zone++)
579 for (int i = ind.cell2at0[zone]; i < ind.cell2at1[zone]; i++)
581 v[i] = receiveBuffer[j++];
585 nat_tot += ind.nrecv[nzone+1];
591 void dd_atom_sum_real(gmx_domdec_t *dd, real v[])
594 gmx_domdec_comm_t *comm;
595 gmx_domdec_comm_dim_t *cd;
599 nzone = comm->zones.n/2;
600 nat_tot = comm->atomRanges.end(DDAtomRanges::Type::Zones);
601 for (int d = dd->ndim-1; d >= 0; d--)
604 for (int p = cd->numPulses() - 1; p >= 0; p--)
606 const gmx_domdec_ind_t &ind = cd->ind[p];
608 /* Note: We provision for RVec instead of real, so a factor of 3
609 * more than needed. The buffer actually already has this size
610 * and we typecast, so this works as intended.
612 DDBufferAccess<gmx::RVec> receiveBufferAccess(comm->rvecBuffer, ind.nsend[nzone + 1]);
613 gmx::ArrayRef<real> receiveBuffer = realArrayRefFromRvecArrayRef(receiveBufferAccess.buffer);
614 nat_tot -= ind.nrecv[nzone + 1];
616 DDBufferAccess<gmx::RVec> sendBufferAccess(comm->rvecBuffer2, cd->receiveInPlace ? 0 : ind.nrecv[nzone + 1]);
618 gmx::ArrayRef<real> sendBuffer;
619 if (cd->receiveInPlace)
621 sendBuffer = gmx::arrayRefFromArray(v + nat_tot, ind.nrecv[nzone + 1]);
625 sendBuffer = realArrayRefFromRvecArrayRef(sendBufferAccess.buffer);
627 for (int zone = 0; zone < nzone; zone++)
629 for (int i = ind.cell2at0[zone]; i < ind.cell2at1[zone]; i++)
631 sendBuffer[j++] = v[i];
635 /* Communicate the forces */
636 ddSendrecv(dd, d, dddirForward,
637 sendBuffer, receiveBuffer);
638 /* Add the received forces */
640 for (int j : ind.index)
642 v[j] += receiveBuffer[n];
650 real dd_cutoff_multibody(const gmx_domdec_t *dd)
652 const gmx_domdec_comm_t &comm = *dd->comm;
653 const DDSystemInfo &systemInfo = comm.systemInfo;
656 if (systemInfo.haveInterDomainMultiBodyBondeds)
658 if (comm.cutoff_mbody > 0)
660 r = comm.cutoff_mbody;
664 /* cutoff_mbody=0 means we do not have DLB */
665 r = comm.cellsize_min[dd->dim[0]];
666 for (int di = 1; di < dd->ndim; di++)
668 r = std::min(r, comm.cellsize_min[dd->dim[di]]);
670 if (comm.systemInfo.filterBondedCommunication)
672 r = std::max(r, comm.cutoff_mbody);
676 r = std::min(r, systemInfo.cutoff);
684 real dd_cutoff_twobody(const gmx_domdec_t *dd)
688 r_mb = dd_cutoff_multibody(dd);
690 return std::max(dd->comm->systemInfo.cutoff, r_mb);
694 static void dd_cart_coord2pmecoord(const gmx_domdec_t *dd, const ivec coord,
699 nc = dd->nc[dd->comm->cartpmedim];
700 ntot = dd->comm->ntot[dd->comm->cartpmedim];
701 copy_ivec(coord, coord_pme);
702 coord_pme[dd->comm->cartpmedim] =
703 nc + (coord[dd->comm->cartpmedim]*(ntot - nc) + (ntot - nc)/2)/nc;
706 static int ddindex2pmeindex(const gmx_domdec_t *dd, int ddindex)
711 npme = dd->comm->npmenodes;
713 /* Here we assign a PME node to communicate with this DD node
714 * by assuming that the major index of both is x.
715 * We add cr->npmenodes/2 to obtain an even distribution.
717 return (ddindex*npme + npme/2)/npp;
720 static int *dd_interleaved_pme_ranks(const gmx_domdec_t *dd)
725 snew(pme_rank, dd->comm->npmenodes);
727 for (i = 0; i < dd->nnodes; i++)
729 p0 = ddindex2pmeindex(dd, i);
730 p1 = ddindex2pmeindex(dd, i+1);
731 if (i+1 == dd->nnodes || p1 > p0)
735 fprintf(debug, "pme_rank[%d] = %d\n", n, i+1+n);
737 pme_rank[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, dd_index(dd->nc, coords));
773 static int ddcoord2simnodeid(const t_commrec *cr, int x, int y, int z)
775 gmx_domdec_comm_t *comm;
777 int ddindex, nodeid = -1;
784 if (comm->bCartesianPP_PME)
787 MPI_Cart_rank(cr->mpi_comm_mysim, coords, &nodeid);
792 ddindex = dd_index(cr->dd->nc, coords);
793 if (comm->bCartesianPP)
795 nodeid = comm->ddindex2simnodeid[ddindex];
801 nodeid = ddindex + gmx_ddcoord2pmeindex(cr, x, y, z);
813 static int dd_simnode2pmenode(const gmx_domdec_t *dd,
814 const t_commrec gmx_unused *cr,
819 const gmx_domdec_comm_t *comm = dd->comm;
821 /* This assumes a uniform x domain decomposition grid cell size */
822 if (comm->bCartesianPP_PME)
825 ivec coord, coord_pme;
826 MPI_Cart_coords(cr->mpi_comm_mysim, sim_nodeid, DIM, coord);
827 if (coord[comm->cartpmedim] < dd->nc[comm->cartpmedim])
829 /* This is a PP node */
830 dd_cart_coord2pmecoord(dd, coord, coord_pme);
831 MPI_Cart_rank(cr->mpi_comm_mysim, coord_pme, &pmenode);
835 else if (comm->bCartesianPP)
837 if (sim_nodeid < dd->nnodes)
839 pmenode = dd->nnodes + ddindex2pmeindex(dd, sim_nodeid);
844 /* This assumes DD cells with identical x coordinates
845 * are numbered sequentially.
847 if (dd->comm->pmenodes == nullptr)
849 if (sim_nodeid < dd->nnodes)
851 /* The DD index equals the nodeid */
852 pmenode = dd->nnodes + ddindex2pmeindex(dd, sim_nodeid);
858 while (sim_nodeid > dd->comm->pmenodes[i])
862 if (sim_nodeid < dd->comm->pmenodes[i])
864 pmenode = dd->comm->pmenodes[i];
872 NumPmeDomains getNumPmeDomains(const gmx_domdec_t *dd)
877 dd->comm->npmenodes_x, dd->comm->npmenodes_y
888 std::vector<int> get_pme_ddranks(const t_commrec *cr, int pmenodeid)
892 ivec coord, coord_pme;
896 std::vector<int> ddranks;
897 ddranks.reserve((dd->nnodes+cr->npmenodes-1)/cr->npmenodes);
899 for (x = 0; x < dd->nc[XX]; x++)
901 for (y = 0; y < dd->nc[YY]; y++)
903 for (z = 0; z < dd->nc[ZZ]; z++)
905 if (dd->comm->bCartesianPP_PME)
910 dd_cart_coord2pmecoord(dd, coord, coord_pme);
911 if (dd->ci[XX] == coord_pme[XX] &&
912 dd->ci[YY] == coord_pme[YY] &&
913 dd->ci[ZZ] == coord_pme[ZZ])
915 ddranks.push_back(ddcoord2simnodeid(cr, x, y, z));
920 /* The slab corresponds to the nodeid in the PME group */
921 if (gmx_ddcoord2pmeindex(cr, x, y, z) == pmenodeid)
923 ddranks.push_back(ddcoord2simnodeid(cr, x, y, z));
932 static gmx_bool receive_vir_ener(const gmx_domdec_t *dd, const t_commrec *cr)
934 gmx_bool bReceive = TRUE;
936 if (cr->npmenodes < dd->nnodes)
938 gmx_domdec_comm_t *comm = dd->comm;
939 if (comm->bCartesianPP_PME)
942 int pmenode = dd_simnode2pmenode(dd, cr, cr->sim_nodeid);
944 MPI_Cart_coords(cr->mpi_comm_mysim, cr->sim_nodeid, DIM, coords);
945 coords[comm->cartpmedim]++;
946 if (coords[comm->cartpmedim] < dd->nc[comm->cartpmedim])
949 MPI_Cart_rank(cr->mpi_comm_mysim, coords, &rank);
950 if (dd_simnode2pmenode(dd, cr, rank) == pmenode)
952 /* This is not the last PP node for pmenode */
957 GMX_RELEASE_ASSERT(false, "Without MPI we should not have Cartesian PP-PME with #PMEnodes < #DDnodes");
962 int pmenode = dd_simnode2pmenode(dd, cr, cr->sim_nodeid);
963 if (cr->sim_nodeid+1 < cr->nnodes &&
964 dd_simnode2pmenode(dd, cr, cr->sim_nodeid+1) == pmenode)
966 /* This is not the last PP node for pmenode */
975 static void set_slb_pme_dim_f(gmx_domdec_t *dd, int dim, real **dim_f)
977 gmx_domdec_comm_t *comm;
982 snew(*dim_f, dd->nc[dim]+1);
984 for (i = 1; i < dd->nc[dim]; i++)
986 if (comm->slb_frac[dim])
988 (*dim_f)[i] = (*dim_f)[i-1] + comm->slb_frac[dim][i-1];
992 (*dim_f)[i] = static_cast<real>(i)/static_cast<real>(dd->nc[dim]);
995 (*dim_f)[dd->nc[dim]] = 1;
998 static void init_ddpme(gmx_domdec_t *dd, gmx_ddpme_t *ddpme, int dimind)
1000 int pmeindex, slab, nso, i;
1003 if (dimind == 0 && dd->dim[0] == YY && dd->comm->npmenodes_x == 1)
1009 ddpme->dim = dimind;
1011 ddpme->dim_match = (ddpme->dim == dd->dim[dimind]);
1013 ddpme->nslab = (ddpme->dim == 0 ?
1014 dd->comm->npmenodes_x :
1015 dd->comm->npmenodes_y);
1017 if (ddpme->nslab <= 1)
1022 nso = dd->comm->npmenodes/ddpme->nslab;
1023 /* Determine for each PME slab the PP location range for dimension dim */
1024 snew(ddpme->pp_min, ddpme->nslab);
1025 snew(ddpme->pp_max, ddpme->nslab);
1026 for (slab = 0; slab < ddpme->nslab; slab++)
1028 ddpme->pp_min[slab] = dd->nc[dd->dim[dimind]] - 1;
1029 ddpme->pp_max[slab] = 0;
1031 for (i = 0; i < dd->nnodes; i++)
1033 ddindex2xyz(dd->nc, i, xyz);
1034 /* For y only use our y/z slab.
1035 * This assumes that the PME x grid size matches the DD grid size.
1037 if (dimind == 0 || xyz[XX] == dd->ci[XX])
1039 pmeindex = ddindex2pmeindex(dd, i);
1042 slab = pmeindex/nso;
1046 slab = pmeindex % ddpme->nslab;
1048 ddpme->pp_min[slab] = std::min(ddpme->pp_min[slab], xyz[dimind]);
1049 ddpme->pp_max[slab] = std::max(ddpme->pp_max[slab], xyz[dimind]);
1053 set_slb_pme_dim_f(dd, ddpme->dim, &ddpme->slb_dim_f);
1056 int dd_pme_maxshift_x(const gmx_domdec_t *dd)
1058 if (dd->comm->ddpme[0].dim == XX)
1060 return dd->comm->ddpme[0].maxshift;
1068 int dd_pme_maxshift_y(const gmx_domdec_t *dd)
1070 if (dd->comm->ddpme[0].dim == YY)
1072 return dd->comm->ddpme[0].maxshift;
1074 else if (dd->comm->npmedecompdim >= 2 && dd->comm->ddpme[1].dim == YY)
1076 return dd->comm->ddpme[1].maxshift;
1084 bool ddHaveSplitConstraints(const gmx_domdec_t &dd)
1086 return dd.comm->systemInfo.haveSplitConstraints;
1089 void dd_cycles_add(const gmx_domdec_t *dd, float cycles, int ddCycl)
1091 /* Note that the cycles value can be incorrect, either 0 or some
1092 * extremely large value, when our thread migrated to another core
1093 * with an unsynchronized cycle counter. If this happens less often
1094 * that once per nstlist steps, this will not cause issues, since
1095 * we later subtract the maximum value from the sum over nstlist steps.
1096 * A zero count will slightly lower the total, but that's a small effect.
1097 * Note that the main purpose of the subtraction of the maximum value
1098 * is to avoid throwing off the load balancing when stalls occur due
1099 * e.g. system activity or network congestion.
1101 dd->comm->cycl[ddCycl] += cycles;
1102 dd->comm->cycl_n[ddCycl]++;
1103 if (cycles > dd->comm->cycl_max[ddCycl])
1105 dd->comm->cycl_max[ddCycl] = cycles;
1110 static void make_load_communicator(gmx_domdec_t *dd, int dim_ind, ivec loc)
1115 gmx_bool bPartOfGroup = FALSE;
1117 dim = dd->dim[dim_ind];
1118 copy_ivec(loc, loc_c);
1119 for (i = 0; i < dd->nc[dim]; i++)
1122 rank = dd_index(dd->nc, loc_c);
1123 if (rank == dd->rank)
1125 /* This process is part of the group */
1126 bPartOfGroup = TRUE;
1129 MPI_Comm_split(dd->mpi_comm_all, bPartOfGroup ? 0 : MPI_UNDEFINED, dd->rank,
1133 dd->comm->mpi_comm_load[dim_ind] = c_row;
1134 if (!isDlbDisabled(dd->comm))
1136 DDCellsizesWithDlb &cellsizes = dd->comm->cellsizesWithDlb[dim_ind];
1138 if (dd->ci[dim] == dd->master_ci[dim])
1140 /* This is the root process of this row */
1141 cellsizes.rowMaster = std::make_unique<RowMaster>();
1143 RowMaster &rowMaster = *cellsizes.rowMaster;
1144 rowMaster.cellFrac.resize(ddCellFractionBufferSize(dd, dim_ind));
1145 rowMaster.oldCellFrac.resize(dd->nc[dim] + 1);
1146 rowMaster.isCellMin.resize(dd->nc[dim]);
1149 rowMaster.bounds.resize(dd->nc[dim]);
1151 rowMaster.buf_ncd.resize(dd->nc[dim]);
1155 /* This is not a root process, we only need to receive cell_f */
1156 cellsizes.fracRow.resize(ddCellFractionBufferSize(dd, dim_ind));
1159 if (dd->ci[dim] == dd->master_ci[dim])
1161 snew(dd->comm->load[dim_ind].load, dd->nc[dim]*DD_NLOAD_MAX);
1167 void dd_setup_dlb_resource_sharing(t_commrec *cr,
1171 int physicalnode_id_hash;
1173 MPI_Comm mpi_comm_pp_physicalnode;
1175 if (!thisRankHasDuty(cr, DUTY_PP) || gpu_id < 0)
1177 /* Only ranks with short-ranged tasks (currently) use GPUs.
1178 * If we don't have GPUs assigned, there are no resources to share.
1183 physicalnode_id_hash = gmx_physicalnode_id_hash();
1189 fprintf(debug, "dd_setup_dd_dlb_gpu_sharing:\n");
1190 fprintf(debug, "DD PP rank %d physical node hash %d gpu_id %d\n",
1191 dd->rank, physicalnode_id_hash, gpu_id);
1193 /* Split the PP communicator over the physical nodes */
1194 /* TODO: See if we should store this (before), as it's also used for
1195 * for the nodecomm summation.
1197 // TODO PhysicalNodeCommunicator could be extended/used to handle
1198 // the need for per-node per-group communicators.
1199 MPI_Comm_split(dd->mpi_comm_all, physicalnode_id_hash, dd->rank,
1200 &mpi_comm_pp_physicalnode);
1201 MPI_Comm_split(mpi_comm_pp_physicalnode, gpu_id, dd->rank,
1202 &dd->comm->mpi_comm_gpu_shared);
1203 MPI_Comm_free(&mpi_comm_pp_physicalnode);
1204 MPI_Comm_size(dd->comm->mpi_comm_gpu_shared, &dd->comm->nrank_gpu_shared);
1208 fprintf(debug, "nrank_gpu_shared %d\n", dd->comm->nrank_gpu_shared);
1211 /* Note that some ranks could share a GPU, while others don't */
1213 if (dd->comm->nrank_gpu_shared == 1)
1215 MPI_Comm_free(&dd->comm->mpi_comm_gpu_shared);
1218 GMX_UNUSED_VALUE(cr);
1219 GMX_UNUSED_VALUE(gpu_id);
1223 static void make_load_communicators(gmx_domdec_t gmx_unused *dd)
1226 int dim0, dim1, i, j;
1231 fprintf(debug, "Making load communicators\n");
1234 dd->comm->load = new domdec_load_t[std::max(dd->ndim, 1)];
1235 snew(dd->comm->mpi_comm_load, std::max(dd->ndim, 1));
1243 make_load_communicator(dd, 0, loc);
1247 for (i = 0; i < dd->nc[dim0]; i++)
1250 make_load_communicator(dd, 1, loc);
1256 for (i = 0; i < dd->nc[dim0]; i++)
1260 for (j = 0; j < dd->nc[dim1]; j++)
1263 make_load_communicator(dd, 2, loc);
1270 fprintf(debug, "Finished making load communicators\n");
1275 /*! \brief Sets up the relation between neighboring domains and zones */
1276 static void setup_neighbor_relations(gmx_domdec_t *dd)
1278 int d, dim, i, j, m;
1280 gmx_domdec_zones_t *zones;
1281 gmx_domdec_ns_ranges_t *izone;
1282 GMX_ASSERT(dd->ndim >= 0, "Must have non-negative number of dimensions for DD");
1284 for (d = 0; d < dd->ndim; d++)
1287 copy_ivec(dd->ci, tmp);
1288 tmp[dim] = (tmp[dim] + 1) % dd->nc[dim];
1289 dd->neighbor[d][0] = ddcoord2ddnodeid(dd, tmp);
1290 copy_ivec(dd->ci, tmp);
1291 tmp[dim] = (tmp[dim] - 1 + dd->nc[dim]) % dd->nc[dim];
1292 dd->neighbor[d][1] = ddcoord2ddnodeid(dd, tmp);
1295 fprintf(debug, "DD rank %d neighbor ranks in dir %d are + %d - %d\n",
1298 dd->neighbor[d][1]);
1302 int nzone = (1 << dd->ndim);
1303 GMX_ASSERT(dd->ndim < DIM, "Invalid number of dimensions");
1304 int nizone = (1 << std::max(dd->ndim - 1, 0));
1305 assert(nizone >= 1 && nizone <= DD_MAXIZONE);
1307 zones = &dd->comm->zones;
1309 for (i = 0; i < nzone; i++)
1312 clear_ivec(zones->shift[i]);
1313 for (d = 0; d < dd->ndim; d++)
1315 zones->shift[i][dd->dim[d]] = dd_zo[i][m++];
1320 for (i = 0; i < nzone; i++)
1322 for (d = 0; d < DIM; d++)
1324 s[d] = dd->ci[d] - zones->shift[i][d];
1329 else if (s[d] >= dd->nc[d])
1335 zones->nizone = nizone;
1336 for (i = 0; i < zones->nizone; i++)
1338 assert(ddNonbondedZonePairRanges[i][0] == i);
1340 izone = &zones->izone[i];
1341 /* dd_zp3 is for 3D decomposition, for fewer dimensions use only
1342 * j-zones up to nzone.
1344 izone->j0 = std::min(ddNonbondedZonePairRanges[i][1], nzone);
1345 izone->j1 = std::min(ddNonbondedZonePairRanges[i][2], nzone);
1346 for (dim = 0; dim < DIM; dim++)
1348 if (dd->nc[dim] == 1)
1350 /* All shifts should be allowed */
1351 izone->shift0[dim] = -1;
1352 izone->shift1[dim] = 1;
1356 /* Determine the min/max j-zone shift wrt the i-zone */
1357 izone->shift0[dim] = 1;
1358 izone->shift1[dim] = -1;
1359 for (j = izone->j0; j < izone->j1; j++)
1361 int shift_diff = zones->shift[j][dim] - zones->shift[i][dim];
1362 if (shift_diff < izone->shift0[dim])
1364 izone->shift0[dim] = shift_diff;
1366 if (shift_diff > izone->shift1[dim])
1368 izone->shift1[dim] = shift_diff;
1375 if (!isDlbDisabled(dd->comm))
1377 dd->comm->cellsizesWithDlb.resize(dd->ndim);
1380 if (dd->comm->ddSettings.recordLoad)
1382 make_load_communicators(dd);
1386 static void make_pp_communicator(const gmx::MDLogger &mdlog,
1388 t_commrec gmx_unused *cr,
1389 bool gmx_unused reorder)
1392 gmx_domdec_comm_t *comm;
1399 if (comm->bCartesianPP)
1401 /* Set up cartesian communication for the particle-particle part */
1402 GMX_LOG(mdlog.info).appendTextFormatted(
1403 "Will use a Cartesian communicator: %d x %d x %d",
1404 dd->nc[XX], dd->nc[YY], dd->nc[ZZ]);
1406 for (int i = 0; i < DIM; i++)
1410 MPI_Cart_create(cr->mpi_comm_mygroup, DIM, dd->nc, periods, static_cast<int>(reorder),
1412 /* We overwrite the old communicator with the new cartesian one */
1413 cr->mpi_comm_mygroup = comm_cart;
1416 dd->mpi_comm_all = cr->mpi_comm_mygroup;
1417 MPI_Comm_rank(dd->mpi_comm_all, &dd->rank);
1419 if (comm->bCartesianPP_PME)
1421 /* Since we want to use the original cartesian setup for sim,
1422 * and not the one after split, we need to make an index.
1424 snew(comm->ddindex2ddnodeid, dd->nnodes);
1425 comm->ddindex2ddnodeid[dd_index(dd->nc, dd->ci)] = dd->rank;
1426 gmx_sumi(dd->nnodes, comm->ddindex2ddnodeid, cr);
1427 /* Get the rank of the DD master,
1428 * 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 (comm->bCartesianPP)
1442 if (cr->npmenodes == 0)
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 snew(comm->ddindex2simnodeid, dd->nnodes);
1457 snew(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, comm->ddindex2simnodeid, dd->nnodes, MPI_INT, MPI_SUM,
1464 cr->mpi_comm_mysim);
1467 /* Determine the master coordinates and rank.
1468 * The DD master should be the same node as the master of this sim.
1470 for (int i = 0; i < dd->nnodes; i++)
1472 if (comm->ddindex2simnodeid[i] == 0)
1474 ddindex2xyz(dd->nc, i, dd->master_ci);
1475 MPI_Cart_rank(dd->mpi_comm_all, dd->master_ci, &dd->masterrank);
1480 fprintf(debug, "The master rank is %d\n", dd->masterrank);
1485 /* No Cartesian communicators */
1486 /* We use the rank in dd->comm->all as DD index */
1487 ddindex2xyz(dd->nc, dd->rank, dd->ci);
1488 /* The simulation master nodeid is 0, so the DD master rank is also 0 */
1490 clear_ivec(dd->master_ci);
1494 GMX_LOG(mdlog.info).appendTextFormatted(
1495 "Domain decomposition rank %d, coordinates %d %d %d\n",
1496 dd->rank, dd->ci[XX], dd->ci[YY], dd->ci[ZZ]);
1500 "Domain decomposition rank %d, coordinates %d %d %d\n\n",
1501 dd->rank, dd->ci[XX], dd->ci[YY], dd->ci[ZZ]);
1505 static void receive_ddindex2simnodeid(gmx_domdec_t *dd,
1509 gmx_domdec_comm_t *comm = dd->comm;
1511 if (!comm->bCartesianPP_PME && comm->bCartesianPP)
1514 snew(comm->ddindex2simnodeid, dd->nnodes);
1515 snew(buf, dd->nnodes);
1516 if (thisRankHasDuty(cr, DUTY_PP))
1518 buf[dd_index(dd->nc, dd->ci)] = cr->sim_nodeid;
1520 /* Communicate the ddindex to simulation nodeid index */
1521 MPI_Allreduce(buf, comm->ddindex2simnodeid, dd->nnodes, MPI_INT, MPI_SUM,
1522 cr->mpi_comm_mysim);
1526 GMX_UNUSED_VALUE(dd);
1527 GMX_UNUSED_VALUE(cr);
1531 static void split_communicator(const gmx::MDLogger &mdlog,
1532 t_commrec *cr, gmx_domdec_t *dd,
1533 DdRankOrder gmx_unused rankOrder,
1534 bool gmx_unused reorder)
1536 gmx_domdec_comm_t *comm;
1545 if (comm->bCartesianPP)
1547 for (i = 1; i < DIM; i++)
1549 bDiv[i] = ((cr->npmenodes*dd->nc[i]) % (dd->nnodes) == 0);
1551 if (bDiv[YY] || bDiv[ZZ])
1553 comm->bCartesianPP_PME = TRUE;
1554 /* If we have 2D PME decomposition, which is always in x+y,
1555 * we stack the PME only nodes in z.
1556 * Otherwise we choose the direction that provides the thinnest slab
1557 * of PME only nodes as this will have the least effect
1558 * on the PP communication.
1559 * But for the PME communication the opposite might be better.
1561 if (bDiv[ZZ] && (comm->npmenodes_y > 1 ||
1563 dd->nc[YY] > dd->nc[ZZ]))
1565 comm->cartpmedim = ZZ;
1569 comm->cartpmedim = YY;
1571 comm->ntot[comm->cartpmedim]
1572 += (cr->npmenodes*dd->nc[comm->cartpmedim])/dd->nnodes;
1576 GMX_LOG(mdlog.info).appendTextFormatted(
1577 "Number of PME-only ranks (%d) is not a multiple of nx*ny (%d*%d) or nx*nz (%d*%d)",
1578 cr->npmenodes, dd->nc[XX], dd->nc[YY], dd->nc[XX], dd->nc[ZZ]);
1579 GMX_LOG(mdlog.info).appendText("Will not use a Cartesian communicator for PP <-> PME\n");
1583 if (comm->bCartesianPP_PME)
1589 GMX_LOG(mdlog.info).appendTextFormatted(
1590 "Will use a Cartesian communicator for PP <-> PME: %d x %d x %d",
1591 comm->ntot[XX], comm->ntot[YY], comm->ntot[ZZ]);
1593 for (i = 0; i < DIM; i++)
1597 MPI_Cart_create(cr->mpi_comm_mysim, DIM, comm->ntot, periods, static_cast<int>(reorder),
1599 MPI_Comm_rank(comm_cart, &rank);
1600 if (MASTER(cr) && rank != 0)
1602 gmx_fatal(FARGS, "MPI rank 0 was renumbered by MPI_Cart_create, we do not allow this");
1605 /* With this assigment we loose the link to the original communicator
1606 * which will usually be MPI_COMM_WORLD, unless have multisim.
1608 cr->mpi_comm_mysim = comm_cart;
1609 cr->sim_nodeid = rank;
1611 MPI_Cart_coords(cr->mpi_comm_mysim, cr->sim_nodeid, DIM, dd->ci);
1613 GMX_LOG(mdlog.info).appendTextFormatted(
1614 "Cartesian rank %d, coordinates %d %d %d\n",
1615 cr->sim_nodeid, dd->ci[XX], dd->ci[YY], dd->ci[ZZ]);
1617 if (dd->ci[comm->cartpmedim] < dd->nc[comm->cartpmedim])
1621 if (cr->npmenodes == 0 ||
1622 dd->ci[comm->cartpmedim] >= dd->nc[comm->cartpmedim])
1624 cr->duty = DUTY_PME;
1627 /* Split the sim communicator into PP and PME only nodes */
1628 MPI_Comm_split(cr->mpi_comm_mysim,
1629 getThisRankDuties(cr),
1630 dd_index(comm->ntot, dd->ci),
1631 &cr->mpi_comm_mygroup);
1638 case DdRankOrder::pp_pme:
1639 GMX_LOG(mdlog.info).appendText("Order of the ranks: PP first, PME last");
1641 case DdRankOrder::interleave:
1642 /* Interleave the PP-only and PME-only ranks */
1643 GMX_LOG(mdlog.info).appendText("Interleaving PP and PME ranks");
1644 comm->pmenodes = dd_interleaved_pme_ranks(dd);
1646 case DdRankOrder::cartesian:
1649 gmx_fatal(FARGS, "Invalid ddRankOrder=%d", static_cast<int>(rankOrder));
1652 if (dd_simnode2pmenode(dd, cr, cr->sim_nodeid) == -1)
1654 cr->duty = DUTY_PME;
1661 /* Split the sim communicator into PP and PME only nodes */
1662 MPI_Comm_split(cr->mpi_comm_mysim,
1663 getThisRankDuties(cr),
1665 &cr->mpi_comm_mygroup);
1666 MPI_Comm_rank(cr->mpi_comm_mygroup, &cr->nodeid);
1670 GMX_LOG(mdlog.info).appendTextFormatted(
1671 "This rank does only %s work.\n",
1672 thisRankHasDuty(cr, DUTY_PP) ? "particle-particle" : "PME-mesh");
1675 /*! \brief Generates the MPI communicators for domain decomposition */
1676 static void make_dd_communicators(const gmx::MDLogger &mdlog,
1678 gmx_domdec_t *dd, DdRankOrder ddRankOrder)
1680 gmx_domdec_comm_t *comm;
1685 copy_ivec(dd->nc, comm->ntot);
1687 comm->bCartesianPP = (ddRankOrder == DdRankOrder::cartesian);
1688 comm->bCartesianPP_PME = FALSE;
1690 /* Reorder the nodes by default. This might change the MPI ranks.
1691 * Real reordering is only supported on very few architectures,
1692 * Blue Gene is one of them.
1694 CartReorder = getenv("GMX_NO_CART_REORDER") == nullptr;
1696 if (cr->npmenodes > 0)
1698 /* Split the communicator into a PP and PME part */
1699 split_communicator(mdlog, cr, dd, ddRankOrder, CartReorder);
1700 if (comm->bCartesianPP_PME)
1702 /* We (possibly) reordered the nodes in split_communicator,
1703 * so it is no longer required in make_pp_communicator.
1705 CartReorder = FALSE;
1710 /* All nodes do PP and PME */
1712 /* We do not require separate communicators */
1713 cr->mpi_comm_mygroup = cr->mpi_comm_mysim;
1717 if (thisRankHasDuty(cr, DUTY_PP))
1719 /* Copy or make a new PP communicator */
1720 make_pp_communicator(mdlog, dd, cr, CartReorder);
1724 receive_ddindex2simnodeid(dd, cr);
1727 if (!thisRankHasDuty(cr, DUTY_PME))
1729 /* Set up the commnuication to our PME node */
1730 dd->pme_nodeid = dd_simnode2pmenode(dd, cr, cr->sim_nodeid);
1731 dd->pme_receive_vir_ener = receive_vir_ener(dd, cr);
1734 fprintf(debug, "My pme_nodeid %d receive ener %s\n",
1735 dd->pme_nodeid, gmx::boolToString(dd->pme_receive_vir_ener));
1740 dd->pme_nodeid = -1;
1743 /* We can not use DDMASTER(dd), because dd->masterrank is set later */
1746 dd->ma = std::make_unique<AtomDistribution>(dd->nc,
1748 comm->cgs_gl.index[comm->cgs_gl.nr]);
1752 static real *get_slb_frac(const gmx::MDLogger &mdlog,
1753 const char *dir, int nc, const char *size_string)
1755 real *slb_frac, tot;
1760 if (nc > 1 && size_string != nullptr)
1762 GMX_LOG(mdlog.info).appendTextFormatted(
1763 "Using static load balancing for the %s direction", dir);
1766 for (i = 0; i < nc; i++)
1769 sscanf(size_string, "%20lf%n", &dbl, &n);
1772 gmx_fatal(FARGS, "Incorrect or not enough DD cell size entries for direction %s: '%s'", dir, size_string);
1779 std::string relativeCellSizes = "Relative cell sizes:";
1780 for (i = 0; i < nc; i++)
1783 relativeCellSizes += gmx::formatString(" %5.3f", slb_frac[i]);
1785 GMX_LOG(mdlog.info).appendText(relativeCellSizes);
1791 static int multi_body_bondeds_count(const gmx_mtop_t *mtop)
1794 gmx_mtop_ilistloop_t iloop = gmx_mtop_ilistloop_init(mtop);
1796 while (const InteractionLists *ilists = gmx_mtop_ilistloop_next(iloop, &nmol))
1798 for (auto &ilist : extractILists(*ilists, IF_BOND))
1800 if (NRAL(ilist.functionType) > 2)
1802 n += nmol*(ilist.iatoms.size()/ilistStride(ilist));
1810 static int dd_getenv(const gmx::MDLogger &mdlog,
1811 const char *env_var, int def)
1817 val = getenv(env_var);
1820 if (sscanf(val, "%20d", &nst) <= 0)
1824 GMX_LOG(mdlog.info).appendTextFormatted(
1825 "Found env.var. %s = %s, using value %d",
1832 static void check_dd_restrictions(const gmx_domdec_t *dd,
1833 const t_inputrec *ir,
1834 const gmx::MDLogger &mdlog)
1836 if (ir->ePBC == epbcSCREW &&
1837 (dd->nc[XX] == 1 || dd->nc[YY] > 1 || dd->nc[ZZ] > 1))
1839 gmx_fatal(FARGS, "With pbc=%s can only do domain decomposition in the x-direction", epbc_names[ir->ePBC]);
1842 if (ir->ns_type == ensSIMPLE)
1844 gmx_fatal(FARGS, "Domain decomposition does not support simple neighbor searching, use grid searching or run with one MPI rank");
1847 if (ir->nstlist == 0)
1849 gmx_fatal(FARGS, "Domain decomposition does not work with nstlist=0");
1852 if (ir->comm_mode == ecmANGULAR && ir->ePBC != epbcNONE)
1854 GMX_LOG(mdlog.warning).appendText("comm-mode angular will give incorrect results when the comm group partially crosses a periodic boundary");
1858 static real average_cellsize_min(const gmx_ddbox_t &ddbox,
1859 const ivec numDomains)
1861 real r = ddbox.box_size[XX];
1862 for (int d = 0; d < DIM; d++)
1864 if (numDomains[d] > 1)
1866 /* Check using the initial average cell size */
1867 r = std::min(r, ddbox.box_size[d]*ddbox.skew_fac[d]/numDomains[d]);
1874 /*! \brief Depending on the DLB initial value return the DLB switched off state or issue an error.
1876 static DlbState forceDlbOffOrBail(DlbState cmdlineDlbState,
1877 const std::string &reasonStr,
1878 const gmx::MDLogger &mdlog)
1880 std::string dlbNotSupportedErr = "Dynamic load balancing requested, but ";
1881 std::string dlbDisableNote = "NOTE: disabling dynamic load balancing as ";
1883 if (cmdlineDlbState == DlbState::onUser)
1885 gmx_fatal(FARGS, "%s", (dlbNotSupportedErr + reasonStr).c_str());
1887 else if (cmdlineDlbState == DlbState::offCanTurnOn)
1889 GMX_LOG(mdlog.info).appendText(dlbDisableNote + reasonStr);
1891 return DlbState::offForever;
1894 /*! \brief Return the dynamic load balancer's initial state based on initial conditions and user inputs.
1896 * This function parses the parameters of "-dlb" command line option setting
1897 * corresponding state values. Then it checks the consistency of the determined
1898 * state with other run parameters and settings. As a result, the initial state
1899 * may be altered or an error may be thrown if incompatibility of options is detected.
1901 * \param [in] mdlog Logger.
1902 * \param [in] dlbOption Enum value for the DLB option.
1903 * \param [in] bRecordLoad True if the load balancer is recording load information.
1904 * \param [in] mdrunOptions Options for mdrun.
1905 * \param [in] ir Pointer mdrun to input parameters.
1906 * \returns DLB initial/startup state.
1908 static DlbState determineInitialDlbState(const gmx::MDLogger &mdlog,
1909 DlbOption dlbOption, gmx_bool bRecordLoad,
1910 const gmx::MdrunOptions &mdrunOptions,
1911 const t_inputrec *ir)
1913 DlbState dlbState = DlbState::offCanTurnOn;
1917 case DlbOption::turnOnWhenUseful: dlbState = DlbState::offCanTurnOn; break;
1918 case DlbOption::no: dlbState = DlbState::offUser; break;
1919 case DlbOption::yes: dlbState = DlbState::onUser; break;
1920 default: gmx_incons("Invalid dlbOption enum value");
1923 /* Reruns don't support DLB: bail or override auto mode */
1924 if (mdrunOptions.rerun)
1926 std::string reasonStr = "it is not supported in reruns.";
1927 return forceDlbOffOrBail(dlbState, reasonStr, mdlog);
1930 /* Unsupported integrators */
1931 if (!EI_DYNAMICS(ir->eI))
1933 auto reasonStr = gmx::formatString("it is only supported with dynamics, not with integrator '%s'.", EI(ir->eI));
1934 return forceDlbOffOrBail(dlbState, reasonStr, mdlog);
1937 /* Without cycle counters we can't time work to balance on */
1940 std::string reasonStr = "cycle counters unsupported or not enabled in the operating system kernel.";
1941 return forceDlbOffOrBail(dlbState, reasonStr, mdlog);
1944 if (mdrunOptions.reproducible)
1946 std::string reasonStr = "you started a reproducible run.";
1949 case DlbState::offUser:
1951 case DlbState::offForever:
1952 GMX_RELEASE_ASSERT(false, "DlbState::offForever is not a valid initial state");
1954 case DlbState::offCanTurnOn:
1955 return forceDlbOffOrBail(dlbState, reasonStr, mdlog);
1956 case DlbState::onCanTurnOff:
1957 GMX_RELEASE_ASSERT(false, "DlbState::offCanTurnOff is not a valid initial state");
1959 case DlbState::onUser:
1960 return forceDlbOffOrBail(dlbState, reasonStr + " In load balanced runs binary reproducibility cannot be ensured.", mdlog);
1962 gmx_fatal(FARGS, "Death horror: undefined case (%d) for load balancing choice", static_cast<int>(dlbState));
1969 static void set_dd_dim(const gmx::MDLogger &mdlog, gmx_domdec_t *dd)
1972 if (getenv("GMX_DD_ORDER_ZYX") != nullptr)
1974 /* Decomposition order z,y,x */
1975 GMX_LOG(mdlog.info).appendText("Using domain decomposition order z, y, x");
1976 for (int dim = DIM-1; dim >= 0; dim--)
1978 if (dd->nc[dim] > 1)
1980 dd->dim[dd->ndim++] = dim;
1986 /* Decomposition order x,y,z */
1987 for (int dim = 0; dim < DIM; dim++)
1989 if (dd->nc[dim] > 1)
1991 dd->dim[dd->ndim++] = dim;
1998 /* Set dim[0] to avoid extra checks on ndim in several places */
2003 static gmx_domdec_comm_t *init_dd_comm()
2005 gmx_domdec_comm_t *comm = new gmx_domdec_comm_t;
2007 comm->n_load_have = 0;
2008 comm->n_load_collect = 0;
2010 comm->haveTurnedOffDlb = false;
2012 for (int i = 0; i < static_cast<int>(DDAtomRanges::Type::Number); i++)
2014 comm->sum_nat[i] = 0;
2018 comm->load_step = 0;
2021 clear_ivec(comm->load_lim);
2025 /* This should be replaced by a unique pointer */
2026 comm->balanceRegion = ddBalanceRegionAllocate();
2031 /* Returns whether mtop contains constraints and/or vsites */
2032 static bool systemHasConstraintsOrVsites(const gmx_mtop_t &mtop)
2034 auto ilistLoop = gmx_mtop_ilistloop_init(mtop);
2036 while (const InteractionLists *ilists = gmx_mtop_ilistloop_next(ilistLoop, &nmol))
2038 if (!extractILists(*ilists, IF_CONSTRAINT | IF_VSITE).empty())
2047 static void setupUpdateGroups(const gmx::MDLogger &mdlog,
2048 const gmx_mtop_t &mtop,
2049 const t_inputrec &inputrec,
2050 const real cutoffMargin,
2051 DDSystemInfo *systemInfo)
2053 /* When we have constraints and/or vsites, it is beneficial to use
2054 * update groups (when possible) to allow independent update of groups.
2056 if (!systemHasConstraintsOrVsites(mtop))
2058 /* No constraints or vsites, atoms can be updated independently */
2062 systemInfo->updateGroupingPerMoleculetype = gmx::makeUpdateGroups(mtop);
2063 systemInfo->useUpdateGroups =
2064 (!systemInfo->updateGroupingPerMoleculetype.empty() &&
2065 getenv("GMX_NO_UPDATEGROUPS") == nullptr);
2067 if (systemInfo->useUpdateGroups)
2069 int numUpdateGroups = 0;
2070 for (const auto &molblock : mtop.molblock)
2072 numUpdateGroups += molblock.nmol*systemInfo->updateGroupingPerMoleculetype[molblock.type].numBlocks();
2075 systemInfo->maxUpdateGroupRadius =
2076 computeMaxUpdateGroupRadius(mtop,
2077 systemInfo->updateGroupingPerMoleculetype,
2078 maxReferenceTemperature(inputrec));
2080 /* To use update groups, the large domain-to-domain cutoff distance
2081 * should be compatible with the box size.
2083 systemInfo->useUpdateGroups = (atomToAtomIntoDomainToDomainCutoff(*systemInfo, 0) < cutoffMargin);
2085 if (systemInfo->useUpdateGroups)
2087 GMX_LOG(mdlog.info).appendTextFormatted(
2088 "Using update groups, nr %d, average size %.1f atoms, max. radius %.3f nm\n",
2090 mtop.natoms/static_cast<double>(numUpdateGroups),
2091 systemInfo->maxUpdateGroupRadius);
2095 GMX_LOG(mdlog.info).appendTextFormatted("The combination of rlist and box size prohibits the use of update groups\n");
2096 systemInfo->updateGroupingPerMoleculetype.clear();
2101 UnitCellInfo::UnitCellInfo(const t_inputrec &ir) :
2102 npbcdim(ePBC2npbcdim(ir.ePBC)),
2103 numBoundedDimensions(inputrec2nboundeddim(&ir)),
2104 ddBoxIsDynamic(numBoundedDimensions < DIM || inputrecDynamicBox(&ir)),
2105 haveScrewPBC(ir.ePBC == epbcSCREW)
2109 /*! \brief Generate the simulation system information */
2111 getSystemInfo(const gmx::MDLogger &mdlog,
2113 const DomdecOptions &options,
2114 const gmx_mtop_t *mtop,
2115 const t_inputrec *ir,
2117 gmx::ArrayRef<const gmx::RVec> xGlobal)
2119 const real tenPercentMargin = 1.1;
2121 DDSystemInfo systemInfo;
2123 /* We need to decide on update groups early, as this affects communication distances */
2124 systemInfo.useUpdateGroups = false;
2125 if (ir->cutoff_scheme == ecutsVERLET)
2127 real cutoffMargin = std::sqrt(max_cutoff2(ir->ePBC, box)) - ir->rlist;
2128 setupUpdateGroups(mdlog, *mtop, *ir, cutoffMargin, &systemInfo);
2131 // TODO: Check whether all bondeds are within update groups
2132 systemInfo.haveInterDomainBondeds = (mtop->natoms > gmx_mtop_num_molecules(*mtop) ||
2133 mtop->bIntermolecularInteractions);
2134 systemInfo.haveInterDomainMultiBodyBondeds = (multi_body_bondeds_count(mtop) > 0);
2136 if (systemInfo.useUpdateGroups)
2138 systemInfo.haveSplitConstraints = false;
2139 systemInfo.haveSplitSettles = false;
2143 systemInfo.haveSplitConstraints = gmx::inter_charge_group_constraints(*mtop);
2144 systemInfo.haveSplitSettles = gmx::inter_charge_group_settles(*mtop);
2149 /* Set the cut-off to some very large value,
2150 * so we don't need if statements everywhere in the code.
2151 * We use sqrt, since the cut-off is squared in some places.
2153 systemInfo.cutoff = GMX_CUTOFF_INF;
2157 systemInfo.cutoff = atomToAtomIntoDomainToDomainCutoff(systemInfo, ir->rlist);
2159 systemInfo.minCutoffForMultiBody = 0;
2161 /* Determine the minimum cell size limit, affected by many factors */
2162 systemInfo.cellsizeLimit = 0;
2163 systemInfo.filterBondedCommunication = false;
2165 /* We do not allow home atoms to move beyond the neighboring domain
2166 * between domain decomposition steps, which limits the cell size.
2167 * Get an estimate of cell size limit due to atom displacement.
2168 * In most cases this is a large overestimate, because it assumes
2169 * non-interaction atoms.
2170 * We set the chance to 1 in a trillion steps.
2172 constexpr real c_chanceThatAtomMovesBeyondDomain = 1e-12;
2173 const real limitForAtomDisplacement =
2174 minCellSizeForAtomDisplacement(*mtop, *ir,
2175 systemInfo.updateGroupingPerMoleculetype,
2176 c_chanceThatAtomMovesBeyondDomain);
2177 GMX_LOG(mdlog.info).appendTextFormatted(
2178 "Minimum cell size due to atom displacement: %.3f nm",
2179 limitForAtomDisplacement);
2181 systemInfo.cellsizeLimit = std::max(systemInfo.cellsizeLimit,
2182 limitForAtomDisplacement);
2184 /* TODO: PME decomposition currently requires atoms not to be more than
2185 * 2/3 of comm->cutoff, which is >=rlist, outside of their domain.
2186 * In nearly all cases, limitForAtomDisplacement will be smaller
2187 * than 2/3*rlist, so the PME requirement is satisfied.
2188 * But it would be better for both correctness and performance
2189 * to use limitForAtomDisplacement instead of 2/3*comm->cutoff.
2190 * Note that we would need to improve the pairlist buffer case.
2193 if (systemInfo.haveInterDomainBondeds)
2195 if (options.minimumCommunicationRange > 0)
2197 systemInfo.minCutoffForMultiBody =
2198 atomToAtomIntoDomainToDomainCutoff(systemInfo, options.minimumCommunicationRange);
2199 if (options.useBondedCommunication)
2201 systemInfo.filterBondedCommunication = (systemInfo.minCutoffForMultiBody > systemInfo.cutoff);
2205 systemInfo.cutoff = std::max(systemInfo.cutoff,
2206 systemInfo.minCutoffForMultiBody);
2209 else if (ir->bPeriodicMols)
2211 /* Can not easily determine the required cut-off */
2212 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.");
2213 systemInfo.minCutoffForMultiBody = systemInfo.cutoff/2;
2221 dd_bonded_cg_distance(mdlog, mtop, ir, as_rvec_array(xGlobal.data()), box,
2222 options.checkBondedInteractions,
2225 gmx_bcast(sizeof(r_2b), &r_2b, cr);
2226 gmx_bcast(sizeof(r_mb), &r_mb, cr);
2228 /* We use an initial margin of 10% for the minimum cell size,
2229 * except when we are just below the non-bonded cut-off.
2231 if (options.useBondedCommunication)
2233 if (std::max(r_2b, r_mb) > systemInfo.cutoff)
2235 const real r_bonded = std::max(r_2b, r_mb);
2236 systemInfo.minCutoffForMultiBody = tenPercentMargin*r_bonded;
2237 /* This is the (only) place where we turn on the filtering */
2238 systemInfo.filterBondedCommunication = true;
2242 const real r_bonded = r_mb;
2243 systemInfo.minCutoffForMultiBody = std::min(tenPercentMargin*r_bonded,
2246 /* We determine cutoff_mbody later */
2247 systemInfo.increaseMultiBodyCutoff = true;
2251 /* No special bonded communication,
2252 * simply increase the DD cut-off.
2254 systemInfo.minCutoffForMultiBody = tenPercentMargin*std::max(r_2b, r_mb);
2255 systemInfo.cutoff = std::max(systemInfo.cutoff,
2256 systemInfo.minCutoffForMultiBody);
2259 GMX_LOG(mdlog.info).appendTextFormatted(
2260 "Minimum cell size due to bonded interactions: %.3f nm",
2261 systemInfo.minCutoffForMultiBody);
2263 systemInfo.cellsizeLimit = std::max(systemInfo.cellsizeLimit,
2264 systemInfo.minCutoffForMultiBody);
2267 systemInfo.constraintCommunicationRange = 0;
2268 if (systemInfo.haveSplitConstraints && options.constraintCommunicationRange <= 0)
2270 /* There is a cell size limit due to the constraints (P-LINCS) */
2271 systemInfo.constraintCommunicationRange = gmx::constr_r_max(mdlog, mtop, ir);
2272 GMX_LOG(mdlog.info).appendTextFormatted(
2273 "Estimated maximum distance required for P-LINCS: %.3f nm",
2274 systemInfo.constraintCommunicationRange);
2275 if (systemInfo.constraintCommunicationRange > systemInfo.cellsizeLimit)
2277 GMX_LOG(mdlog.info).appendText("This distance will limit the DD cell size, you can override this with -rcon");
2280 else if (options.constraintCommunicationRange > 0)
2282 /* Here we do not check for dd->splitConstraints.
2283 * because one can also set a cell size limit for virtual sites only
2284 * and at this point we don't know yet if there are intercg v-sites.
2286 GMX_LOG(mdlog.info).appendTextFormatted(
2287 "User supplied maximum distance required for P-LINCS: %.3f nm",
2288 options.constraintCommunicationRange);
2289 systemInfo.constraintCommunicationRange = options.constraintCommunicationRange;
2291 systemInfo.cellsizeLimit = std::max(systemInfo.cellsizeLimit,
2292 systemInfo.constraintCommunicationRange);
2297 /*! \brief Set the cell size and interaction limits, as well as the DD grid */
2298 static void set_dd_limits_and_grid(const gmx::MDLogger &mdlog,
2299 t_commrec *cr, gmx_domdec_t *dd,
2300 const DomdecOptions &options,
2301 const DDSettings &ddSettings,
2302 const gmx_mtop_t *mtop,
2303 const t_inputrec *ir,
2305 gmx::ArrayRef<const gmx::RVec> xGlobal,
2308 gmx_domdec_comm_t *comm = dd->comm;
2309 comm->ddSettings = ddSettings;
2311 /* Initialize to GPU share count to 0, might change later */
2312 comm->nrank_gpu_shared = 0;
2314 comm->dlbState = comm->ddSettings.initialDlbState;
2315 dd_dlb_set_should_check_whether_to_turn_dlb_on(dd, TRUE);
2316 /* To consider turning DLB on after 2*nstlist steps we need to check
2317 * at partitioning count 3. Thus we need to increase the first count by 2.
2319 comm->ddPartioningCountFirstDlbOff += 2;
2321 GMX_LOG(mdlog.info).appendTextFormatted(
2322 "Dynamic load balancing: %s", edlbs_names[int(comm->dlbState)]);
2324 comm->bPMELoadBalDLBLimits = FALSE;
2326 /* Allocate the charge group/atom sorting struct */
2327 comm->sort = std::make_unique<gmx_domdec_sort_t>();
2329 /* Generate the simulation system information */
2330 comm->systemInfo = getSystemInfo(mdlog, cr, options, mtop, ir, box, xGlobal);
2331 const DDSystemInfo &systemInfo = comm->systemInfo;
2333 if (systemInfo.useUpdateGroups)
2335 /* Note: We would like to use dd->nnodes for the atom count estimate,
2336 * but that is not yet available here. But this anyhow only
2337 * affect performance up to the second dd_partition_system call.
2339 const int homeAtomCountEstimate = mtop->natoms/cr->nnodes;
2340 comm->updateGroupsCog =
2341 std::make_unique<gmx::UpdateGroupsCog>(*mtop,
2342 systemInfo.updateGroupingPerMoleculetype,
2343 maxReferenceTemperature(*ir),
2344 homeAtomCountEstimate);
2347 comm->cgs_gl = gmx_mtop_global_cgs(mtop);
2350 if (options.numCells[XX] > 0)
2352 copy_ivec(options.numCells, ddSetup.numDomains);
2353 set_ddbox_cr(*cr, &ddSetup.numDomains, *ir, box, xGlobal, ddbox);
2355 if (options.numPmeRanks >= 0)
2357 ddSetup.numPmeRanks = options.numPmeRanks;
2361 /* When the DD grid is set explicitly and -npme is set to auto,
2362 * don't use PME ranks. We check later if the DD grid is
2363 * compatible with the total number of ranks.
2365 ddSetup.numPmeRanks = 0;
2370 set_ddbox_cr(*cr, nullptr, *ir, box, xGlobal, ddbox);
2372 /* We need to choose the optimal DD grid and possibly PME nodes */
2374 dd_choose_grid(mdlog, cr, ir, mtop, box, ddbox,
2375 options.numPmeRanks,
2376 !isDlbDisabled(comm),
2380 if (ddSetup.numDomains[XX] == 0)
2383 gmx_bool bC = (systemInfo.haveSplitConstraints &&
2384 systemInfo.constraintCommunicationRange > systemInfo.minCutoffForMultiBody);
2385 sprintf(buf, "Change the number of ranks or mdrun option %s%s%s",
2386 !bC ? "-rdd" : "-rcon",
2387 comm->dlbState != DlbState::offUser ? " or -dds" : "",
2388 bC ? " or your LINCS settings" : "");
2390 gmx_fatal_collective(FARGS, cr->mpi_comm_mysim, MASTER(cr),
2391 "There is no domain decomposition for %d ranks that is compatible with the given box and a minimum cell size of %g nm\n"
2393 "Look in the log file for details on the domain decomposition",
2394 cr->nnodes - ddSetup.numPmeRanks,
2395 ddSetup.cellsizeLimit,
2400 const real acs = average_cellsize_min(*ddbox, ddSetup.numDomains);
2401 if (acs < systemInfo.cellsizeLimit)
2403 if (options.numCells[XX] <= 0)
2405 GMX_RELEASE_ASSERT(false, "dd_choose_grid() should return a grid that satisfies the cell size limits");
2409 gmx_fatal_collective(FARGS, cr->mpi_comm_mysim, MASTER(cr),
2410 "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",
2411 acs, systemInfo.cellsizeLimit);
2415 /* Set the DD setup given by ddSetup */
2416 cr->npmenodes = ddSetup.numPmeRanks;
2417 copy_ivec(ddSetup.numDomains, dd->nc);
2418 set_dd_dim(mdlog, dd);
2420 GMX_LOG(mdlog.info).appendTextFormatted(
2421 "Domain decomposition grid %d x %d x %d, separate PME ranks %d",
2422 dd->nc[XX], dd->nc[YY], dd->nc[ZZ], cr->npmenodes);
2424 dd->nnodes = dd->nc[XX]*dd->nc[YY]*dd->nc[ZZ];
2425 if (cr->nnodes - dd->nnodes != cr->npmenodes)
2427 gmx_fatal_collective(FARGS, cr->mpi_comm_mysim, MASTER(cr),
2428 "The size of the domain decomposition grid (%d) does not match the number of ranks (%d). The total number of ranks is %d",
2429 dd->nnodes, cr->nnodes - cr->npmenodes, cr->nnodes);
2431 if (cr->npmenodes > dd->nnodes)
2433 gmx_fatal_collective(FARGS, cr->mpi_comm_mysim, MASTER(cr),
2434 "The number of separate PME ranks (%d) is larger than the number of PP ranks (%d), this is not supported.", cr->npmenodes, dd->nnodes);
2436 if (cr->npmenodes > 0)
2438 comm->npmenodes = cr->npmenodes;
2442 comm->npmenodes = dd->nnodes;
2445 if (EEL_PME(ir->coulombtype) || EVDW_PME(ir->vdwtype))
2447 /* The following choices should match those
2448 * in comm_cost_est in domdec_setup.c.
2449 * Note that here the checks have to take into account
2450 * that the decomposition might occur in a different order than xyz
2451 * (for instance through the env.var. GMX_DD_ORDER_ZYX),
2452 * in which case they will not match those in comm_cost_est,
2453 * but since that is mainly for testing purposes that's fine.
2455 if (dd->ndim >= 2 && dd->dim[0] == XX && dd->dim[1] == YY &&
2456 comm->npmenodes > dd->nc[XX] && comm->npmenodes % dd->nc[XX] == 0 &&
2457 getenv("GMX_PMEONEDD") == nullptr)
2459 comm->npmedecompdim = 2;
2460 comm->npmenodes_x = dd->nc[XX];
2461 comm->npmenodes_y = comm->npmenodes/comm->npmenodes_x;
2465 /* In case nc is 1 in both x and y we could still choose to
2466 * decompose pme in y instead of x, but we use x for simplicity.
2468 comm->npmedecompdim = 1;
2469 if (dd->dim[0] == YY)
2471 comm->npmenodes_x = 1;
2472 comm->npmenodes_y = comm->npmenodes;
2476 comm->npmenodes_x = comm->npmenodes;
2477 comm->npmenodes_y = 1;
2480 GMX_LOG(mdlog.info).appendTextFormatted(
2481 "PME domain decomposition: %d x %d x %d",
2482 comm->npmenodes_x, comm->npmenodes_y, 1);
2486 comm->npmedecompdim = 0;
2487 comm->npmenodes_x = 0;
2488 comm->npmenodes_y = 0;
2491 snew(comm->slb_frac, DIM);
2492 if (isDlbDisabled(comm))
2494 comm->slb_frac[XX] = get_slb_frac(mdlog, "x", dd->nc[XX], options.cellSizeX);
2495 comm->slb_frac[YY] = get_slb_frac(mdlog, "y", dd->nc[YY], options.cellSizeY);
2496 comm->slb_frac[ZZ] = get_slb_frac(mdlog, "z", dd->nc[ZZ], options.cellSizeZ);
2499 /* Set the multi-body cut-off and cellsize limit for DLB */
2500 comm->cutoff_mbody = systemInfo.minCutoffForMultiBody;
2501 comm->cellsize_limit = systemInfo.cellsizeLimit;
2502 if (systemInfo.haveInterDomainBondeds && systemInfo.increaseMultiBodyCutoff)
2504 if (systemInfo.filterBondedCommunication || !isDlbDisabled(comm))
2506 /* Set the bonded communication distance to halfway
2507 * the minimum and the maximum,
2508 * since the extra communication cost is nearly zero.
2510 real acs = average_cellsize_min(*ddbox, dd->nc);
2511 comm->cutoff_mbody = 0.5*(systemInfo.minCutoffForMultiBody + acs);
2512 if (!isDlbDisabled(comm))
2514 /* Check if this does not limit the scaling */
2515 comm->cutoff_mbody = std::min(comm->cutoff_mbody,
2516 options.dlbScaling*acs);
2518 if (!systemInfo.filterBondedCommunication)
2520 /* Without bBondComm do not go beyond the n.b. cut-off */
2521 comm->cutoff_mbody = std::min(comm->cutoff_mbody, systemInfo.cutoff);
2522 if (comm->cellsize_limit >= systemInfo.cutoff)
2524 /* We don't loose a lot of efficieny
2525 * when increasing it to the n.b. cut-off.
2526 * It can even be slightly faster, because we need
2527 * less checks for the communication setup.
2529 comm->cutoff_mbody = systemInfo.cutoff;
2532 /* Check if we did not end up below our original limit */
2533 comm->cutoff_mbody = std::max(comm->cutoff_mbody,
2534 systemInfo.minCutoffForMultiBody);
2536 if (comm->cutoff_mbody > comm->cellsize_limit)
2538 comm->cellsize_limit = comm->cutoff_mbody;
2541 /* Without DLB and cutoff_mbody<cutoff, cutoff_mbody is dynamic */
2546 fprintf(debug, "Bonded atom communication beyond the cut-off: %s\n"
2547 "cellsize limit %f\n",
2548 gmx::boolToString(systemInfo.filterBondedCommunication),
2549 comm->cellsize_limit);
2554 check_dd_restrictions(dd, ir, mdlog);
2558 static char *init_bLocalCG(const gmx_mtop_t *mtop)
2563 ncg = ncg_mtop(mtop);
2564 snew(bLocalCG, ncg);
2565 for (cg = 0; cg < ncg; cg++)
2567 bLocalCG[cg] = FALSE;
2573 void dd_init_bondeds(FILE *fplog,
2575 const gmx_mtop_t *mtop,
2576 const gmx_vsite_t *vsite,
2577 const t_inputrec *ir,
2578 gmx_bool bBCheck, cginfo_mb_t *cginfo_mb)
2580 gmx_domdec_comm_t *comm;
2582 dd_make_reverse_top(fplog, dd, mtop, vsite, ir, bBCheck);
2586 if (comm->systemInfo.filterBondedCommunication)
2588 /* Communicate atoms beyond the cut-off for bonded interactions */
2591 comm->cglink = make_charge_group_links(mtop, dd, cginfo_mb);
2593 comm->bLocalCG = init_bLocalCG(mtop);
2597 /* Only communicate atoms based on cut-off */
2598 comm->cglink = nullptr;
2599 comm->bLocalCG = nullptr;
2603 static void writeSettings(gmx::TextWriter *log,
2605 const gmx_mtop_t *mtop,
2606 const t_inputrec *ir,
2607 gmx_bool bDynLoadBal,
2609 const gmx_ddbox_t *ddbox)
2611 gmx_domdec_comm_t *comm;
2620 log->writeString("The maximum number of communication pulses is:");
2621 for (d = 0; d < dd->ndim; d++)
2623 log->writeStringFormatted(" %c %d", dim2char(dd->dim[d]), comm->cd[d].np_dlb);
2625 log->ensureLineBreak();
2626 log->writeLineFormatted("The minimum size for domain decomposition cells is %.3f nm", comm->cellsize_limit);
2627 log->writeLineFormatted("The requested allowed shrink of DD cells (option -dds) is: %.2f", dlb_scale);
2628 log->writeString("The allowed shrink of domain decomposition cells is:");
2629 for (d = 0; d < DIM; d++)
2633 if (d >= ddbox->npbcdim && dd->nc[d] == 2)
2640 comm->cellsize_min_dlb[d]/
2641 (ddbox->box_size[d]*ddbox->skew_fac[d]/dd->nc[d]);
2643 log->writeStringFormatted(" %c %.2f", dim2char(d), shrink);
2646 log->ensureLineBreak();
2650 set_dd_cell_sizes_slb(dd, ddbox, setcellsizeslbPULSE_ONLY, np);
2651 log->writeString("The initial number of communication pulses is:");
2652 for (d = 0; d < dd->ndim; d++)
2654 log->writeStringFormatted(" %c %d", dim2char(dd->dim[d]), np[dd->dim[d]]);
2656 log->ensureLineBreak();
2657 log->writeString("The initial domain decomposition cell size is:");
2658 for (d = 0; d < DIM; d++)
2662 log->writeStringFormatted(" %c %.2f nm",
2663 dim2char(d), dd->comm->cellsize_min[d]);
2666 log->ensureLineBreak();
2670 const bool haveInterDomainVsites =
2671 (countInterUpdategroupVsites(*mtop, comm->systemInfo.updateGroupingPerMoleculetype) != 0);
2673 if (comm->systemInfo.haveInterDomainBondeds ||
2674 haveInterDomainVsites ||
2675 comm->systemInfo.haveSplitConstraints ||
2676 comm->systemInfo.haveSplitSettles)
2678 std::string decompUnits;
2679 if (comm->systemInfo.useUpdateGroups)
2681 decompUnits = "atom groups";
2685 decompUnits = "atoms";
2688 log->writeLineFormatted("The maximum allowed distance for %s involved in interactions is:", decompUnits.c_str());
2689 log->writeLineFormatted("%40s %-7s %6.3f nm", "non-bonded interactions", "", comm->systemInfo.cutoff);
2693 limit = dd->comm->cellsize_limit;
2697 if (dd->unitCellInfo.ddBoxIsDynamic)
2699 log->writeLine("(the following are initial values, they could change due to box deformation)");
2701 limit = dd->comm->cellsize_min[XX];
2702 for (d = 1; d < DIM; d++)
2704 limit = std::min(limit, dd->comm->cellsize_min[d]);
2708 if (comm->systemInfo.haveInterDomainBondeds)
2710 log->writeLineFormatted("%40s %-7s %6.3f nm",
2711 "two-body bonded interactions", "(-rdd)",
2712 std::max(comm->systemInfo.cutoff, comm->cutoff_mbody));
2713 log->writeLineFormatted("%40s %-7s %6.3f nm",
2714 "multi-body bonded interactions", "(-rdd)",
2715 (comm->systemInfo.filterBondedCommunication || isDlbOn(dd->comm)) ? comm->cutoff_mbody : std::min(comm->systemInfo.cutoff, limit));
2717 if (haveInterDomainVsites)
2719 log->writeLineFormatted("%40s %-7s %6.3f nm",
2720 "virtual site constructions", "(-rcon)", limit);
2722 if (comm->systemInfo.haveSplitConstraints || comm->systemInfo.haveSplitSettles)
2724 std::string separation = gmx::formatString("atoms separated by up to %d constraints",
2726 log->writeLineFormatted("%40s %-7s %6.3f nm\n",
2727 separation.c_str(), "(-rcon)", limit);
2729 log->ensureLineBreak();
2733 static void logSettings(const gmx::MDLogger &mdlog,
2735 const gmx_mtop_t *mtop,
2736 const t_inputrec *ir,
2738 const gmx_ddbox_t *ddbox)
2740 gmx::StringOutputStream stream;
2741 gmx::TextWriter log(&stream);
2742 writeSettings(&log, dd, mtop, ir, isDlbOn(dd->comm), dlb_scale, ddbox);
2743 if (dd->comm->dlbState == DlbState::offCanTurnOn)
2746 log.ensureEmptyLine();
2747 log.writeLine("When dynamic load balancing gets turned on, these settings will change to:");
2749 writeSettings(&log, dd, mtop, ir, true, dlb_scale, ddbox);
2751 GMX_LOG(mdlog.info).asParagraph().appendText(stream.toString());
2754 static void set_cell_limits_dlb(const gmx::MDLogger &mdlog,
2757 const t_inputrec *ir,
2758 const gmx_ddbox_t *ddbox)
2760 gmx_domdec_comm_t *comm;
2761 int d, dim, npulse, npulse_d_max, npulse_d;
2766 bNoCutOff = (ir->rvdw == 0 || ir->rcoulomb == 0);
2768 /* Determine the maximum number of comm. pulses in one dimension */
2770 comm->cellsize_limit = std::max(comm->cellsize_limit, comm->cutoff_mbody);
2772 /* Determine the maximum required number of grid pulses */
2773 if (comm->cellsize_limit >= comm->systemInfo.cutoff)
2775 /* Only a single pulse is required */
2778 else if (!bNoCutOff && comm->cellsize_limit > 0)
2780 /* We round down slightly here to avoid overhead due to the latency
2781 * of extra communication calls when the cut-off
2782 * would be only slightly longer than the cell size.
2783 * Later cellsize_limit is redetermined,
2784 * so we can not miss interactions due to this rounding.
2786 npulse = static_cast<int>(0.96 + comm->systemInfo.cutoff/comm->cellsize_limit);
2790 /* There is no cell size limit */
2791 npulse = std::max(dd->nc[XX]-1, std::max(dd->nc[YY]-1, dd->nc[ZZ]-1));
2794 if (!bNoCutOff && npulse > 1)
2796 /* See if we can do with less pulses, based on dlb_scale */
2798 for (d = 0; d < dd->ndim; d++)
2801 npulse_d = static_cast<int>(1 + dd->nc[dim]*comm->systemInfo.cutoff
2802 /(ddbox->box_size[dim]*ddbox->skew_fac[dim]*dlb_scale));
2803 npulse_d_max = std::max(npulse_d_max, npulse_d);
2805 npulse = std::min(npulse, npulse_d_max);
2808 /* This env var can override npulse */
2809 d = dd_getenv(mdlog, "GMX_DD_NPULSE", 0);
2816 comm->bVacDLBNoLimit = (ir->ePBC == epbcNONE);
2817 for (d = 0; d < dd->ndim; d++)
2819 comm->cd[d].np_dlb = std::min(npulse, dd->nc[dd->dim[d]]-1);
2820 comm->maxpulse = std::max(comm->maxpulse, comm->cd[d].np_dlb);
2821 if (comm->cd[d].np_dlb < dd->nc[dd->dim[d]]-1)
2823 comm->bVacDLBNoLimit = FALSE;
2827 /* cellsize_limit is set for LINCS in init_domain_decomposition */
2828 if (!comm->bVacDLBNoLimit)
2830 comm->cellsize_limit = std::max(comm->cellsize_limit,
2831 comm->systemInfo.cutoff/comm->maxpulse);
2833 comm->cellsize_limit = std::max(comm->cellsize_limit, comm->cutoff_mbody);
2834 /* Set the minimum cell size for each DD dimension */
2835 for (d = 0; d < dd->ndim; d++)
2837 if (comm->bVacDLBNoLimit ||
2838 comm->cd[d].np_dlb*comm->cellsize_limit >= comm->systemInfo.cutoff)
2840 comm->cellsize_min_dlb[dd->dim[d]] = comm->cellsize_limit;
2844 comm->cellsize_min_dlb[dd->dim[d]] =
2845 comm->systemInfo.cutoff/comm->cd[d].np_dlb;
2848 if (comm->cutoff_mbody <= 0)
2850 comm->cutoff_mbody = std::min(comm->systemInfo.cutoff, comm->cellsize_limit);
2858 gmx_bool dd_bonded_molpbc(const gmx_domdec_t *dd, int ePBC)
2860 /* If each molecule is a single charge group
2861 * or we use domain decomposition for each periodic dimension,
2862 * we do not need to take pbc into account for the bonded interactions.
2864 return (ePBC != epbcNONE && dd->comm->systemInfo.haveInterDomainBondeds &&
2867 (dd->nc[ZZ] > 1 || ePBC == epbcXY)));
2870 /*! \brief Sets grid size limits and PP-PME setup, prints settings to log */
2871 static void set_ddgrid_parameters(const gmx::MDLogger &mdlog,
2872 gmx_domdec_t *dd, real dlb_scale,
2873 const gmx_mtop_t *mtop, const t_inputrec *ir,
2874 const gmx_ddbox_t *ddbox)
2876 gmx_domdec_comm_t *comm;
2882 if (EEL_PME(ir->coulombtype) || EVDW_PME(ir->vdwtype))
2884 init_ddpme(dd, &comm->ddpme[0], 0);
2885 if (comm->npmedecompdim >= 2)
2887 init_ddpme(dd, &comm->ddpme[1], 1);
2892 comm->npmenodes = 0;
2893 if (dd->pme_nodeid >= 0)
2895 gmx_fatal_collective(FARGS, dd->mpi_comm_all, DDMASTER(dd),
2896 "Can not have separate PME ranks without PME electrostatics");
2902 fprintf(debug, "The DD cut-off is %f\n", comm->systemInfo.cutoff);
2904 if (!isDlbDisabled(comm))
2906 set_cell_limits_dlb(mdlog, dd, dlb_scale, ir, ddbox);
2909 logSettings(mdlog, dd, mtop, ir, dlb_scale, ddbox);
2911 if (ir->ePBC == epbcNONE)
2913 vol_frac = 1 - 1/static_cast<double>(dd->nnodes);
2918 (1 + comm_box_frac(dd->nc, comm->systemInfo.cutoff, ddbox))/static_cast<double>(dd->nnodes);
2922 fprintf(debug, "Volume fraction for all DD zones: %f\n", vol_frac);
2924 natoms_tot = comm->cgs_gl.index[comm->cgs_gl.nr];
2926 dd->ga2la = new gmx_ga2la_t(natoms_tot,
2927 static_cast<int>(vol_frac*natoms_tot));
2930 /*! \brief Get some important DD parameters which can be modified by env.vars */
2932 getDDSettings(const gmx::MDLogger &mdlog,
2933 const DomdecOptions &options,
2934 const gmx::MdrunOptions &mdrunOptions,
2935 const t_inputrec &ir)
2937 DDSettings ddSettings;
2939 ddSettings.useSendRecv2 = (dd_getenv(mdlog, "GMX_DD_USE_SENDRECV2", 0) != 0);
2940 ddSettings.dlb_scale_lim = dd_getenv(mdlog, "GMX_DLB_MAX_BOX_SCALING", 10);
2941 ddSettings.eFlop = dd_getenv(mdlog, "GMX_DLB_BASED_ON_FLOPS", 0);
2942 const int recload = dd_getenv(mdlog, "GMX_DD_RECORD_LOAD", 1);
2943 ddSettings.nstDDDump = dd_getenv(mdlog, "GMX_DD_NST_DUMP", 0);
2944 ddSettings.nstDDDumpGrid = dd_getenv(mdlog, "GMX_DD_NST_DUMP_GRID", 0);
2945 ddSettings.DD_debug = dd_getenv(mdlog, "GMX_DD_DEBUG", 0);
2947 if (ddSettings.useSendRecv2)
2949 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");
2952 if (ddSettings.eFlop)
2954 GMX_LOG(mdlog.info).appendText("Will load balance based on FLOP count");
2955 ddSettings.recordLoad = true;
2959 ddSettings.recordLoad = (wallcycle_have_counter() && recload > 0);
2962 ddSettings.initialDlbState =
2963 determineInitialDlbState(mdlog, options.dlbOption, ddSettings.recordLoad, mdrunOptions, &ir);
2968 gmx_domdec_t::gmx_domdec_t(const t_inputrec &ir) :
2973 gmx_domdec_t *init_domain_decomposition(const gmx::MDLogger &mdlog,
2975 const DomdecOptions &options,
2976 const gmx::MdrunOptions &mdrunOptions,
2977 const gmx_mtop_t *mtop,
2978 const t_inputrec *ir,
2980 gmx::ArrayRef<const gmx::RVec> xGlobal,
2981 gmx::LocalAtomSetManager *atomSets)
2985 GMX_LOG(mdlog.info).appendTextFormatted(
2986 "\nInitializing Domain Decomposition on %d ranks", cr->nnodes);
2988 dd = new gmx_domdec_t(*ir);
2990 dd->comm = init_dd_comm();
2992 DDSettings ddSettings = getDDSettings(mdlog, options, mdrunOptions, *ir);
2993 if (ddSettings.eFlop > 1)
2995 /* Ensure that we have different random flop counts on different ranks */
2996 srand(1 + cr->nodeid);
2999 gmx_ddbox_t ddbox = {0};
3000 set_dd_limits_and_grid(mdlog, cr, dd, options,
3006 make_dd_communicators(mdlog, cr, dd, options.rankOrder);
3008 if (thisRankHasDuty(cr, DUTY_PP))
3010 set_ddgrid_parameters(mdlog, dd, options.dlbScaling, mtop, ir, &ddbox);
3012 setup_neighbor_relations(dd);
3015 /* Set overallocation to avoid frequent reallocation of arrays */
3016 set_over_alloc_dd(TRUE);
3018 dd->atomSets = atomSets;
3023 static gmx_bool test_dd_cutoff(t_commrec *cr,
3024 const t_state &state,
3025 real cutoffRequested)
3035 set_ddbox(*dd, false, state.box, true, state.x, &ddbox);
3039 for (d = 0; d < dd->ndim; d++)
3043 inv_cell_size = DD_CELL_MARGIN*dd->nc[dim]/ddbox.box_size[dim];
3044 if (dd->unitCellInfo.ddBoxIsDynamic)
3046 inv_cell_size *= DD_PRES_SCALE_MARGIN;
3049 np = 1 + static_cast<int>(cutoffRequested*inv_cell_size*ddbox.skew_fac[dim]);
3051 if (!isDlbDisabled(dd->comm) && (dim < ddbox.npbcdim) && (dd->comm->cd[d].np_dlb > 0))
3053 if (np > dd->comm->cd[d].np_dlb)
3058 /* If a current local cell size is smaller than the requested
3059 * cut-off, we could still fix it, but this gets very complicated.
3060 * Without fixing here, we might actually need more checks.
3062 real cellSizeAlongDim = (dd->comm->cell_x1[dim] - dd->comm->cell_x0[dim])*ddbox.skew_fac[dim];
3063 if (cellSizeAlongDim*dd->comm->cd[d].np_dlb < cutoffRequested)
3070 if (!isDlbDisabled(dd->comm))
3072 /* If DLB is not active yet, we don't need to check the grid jumps.
3073 * Actually we shouldn't, because then the grid jump data is not set.
3075 if (isDlbOn(dd->comm) &&
3076 check_grid_jump(0, dd, cutoffRequested, &ddbox, FALSE))
3081 gmx_sumi(1, &LocallyLimited, cr);
3083 if (LocallyLimited > 0)
3092 gmx_bool change_dd_cutoff(t_commrec *cr,
3093 const t_state &state,
3094 real cutoffRequested)
3096 gmx_bool bCutoffAllowed;
3098 bCutoffAllowed = test_dd_cutoff(cr, state, cutoffRequested);
3102 cr->dd->comm->systemInfo.cutoff = cutoffRequested;
3105 return bCutoffAllowed;