2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2005,2006,2007,2008,2009 by the GROMACS development team.
5 * Copyright (c) 2010,2011,2012,2013,2014 by the GROMACS development team.
6 * Copyright (c) 2015,2016,2017,2018,2019 by the GROMACS development team.
7 * Copyright (c) 2020, by the GROMACS development team, led by
8 * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
9 * and including many others, as listed in the AUTHORS file in the
10 * top-level source directory and at http://www.gromacs.org.
12 * GROMACS is free software; you can redistribute it and/or
13 * modify it under the terms of the GNU Lesser General Public License
14 * as published by the Free Software Foundation; either version 2.1
15 * of the License, or (at your option) any later version.
17 * GROMACS is distributed in the hope that it will be useful,
18 * but WITHOUT ANY WARRANTY; without even the implied warranty of
19 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
20 * Lesser General Public License for more details.
22 * You should have received a copy of the GNU Lesser General Public
23 * License along with GROMACS; if not, see
24 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
25 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
27 * If you want to redistribute modifications to GROMACS, please
28 * consider that scientific software is very special. Version
29 * control is crucial - bugs must be traceable. We will be happy to
30 * consider code for inclusion in the official distribution, but
31 * derived work must not be called official GROMACS. Details are found
32 * in the README & COPYING files - if they are missing, get the
33 * official version at http://www.gromacs.org.
35 * To help us fund GROMACS development, we humbly ask that you cite
36 * the research papers on the package. Check out http://www.gromacs.org.
56 #include "gromacs/domdec/builder.h"
57 #include "gromacs/domdec/collect.h"
58 #include "gromacs/domdec/dlb.h"
59 #include "gromacs/domdec/dlbtiming.h"
60 #include "gromacs/domdec/domdec_network.h"
61 #include "gromacs/domdec/ga2la.h"
62 #include "gromacs/domdec/gpuhaloexchange.h"
63 #include "gromacs/domdec/options.h"
64 #include "gromacs/domdec/partition.h"
65 #include "gromacs/gmxlib/network.h"
66 #include "gromacs/gmxlib/nrnb.h"
67 #include "gromacs/gpu_utils/gpu_utils.h"
68 #include "gromacs/hardware/hw_info.h"
69 #include "gromacs/listed_forces/manage_threading.h"
70 #include "gromacs/math/vec.h"
71 #include "gromacs/math/vectypes.h"
72 #include "gromacs/mdlib/calc_verletbuf.h"
73 #include "gromacs/mdlib/constr.h"
74 #include "gromacs/mdlib/constraintrange.h"
75 #include "gromacs/mdlib/updategroups.h"
76 #include "gromacs/mdlib/vsite.h"
77 #include "gromacs/mdtypes/commrec.h"
78 #include "gromacs/mdtypes/forceoutput.h"
79 #include "gromacs/mdtypes/inputrec.h"
80 #include "gromacs/mdtypes/mdrunoptions.h"
81 #include "gromacs/mdtypes/state.h"
82 #include "gromacs/pbcutil/ishift.h"
83 #include "gromacs/pbcutil/pbc.h"
84 #include "gromacs/pulling/pull.h"
85 #include "gromacs/timing/wallcycle.h"
86 #include "gromacs/topology/block.h"
87 #include "gromacs/topology/idef.h"
88 #include "gromacs/topology/ifunc.h"
89 #include "gromacs/topology/mtop_lookup.h"
90 #include "gromacs/topology/mtop_util.h"
91 #include "gromacs/topology/topology.h"
92 #include "gromacs/utility/basedefinitions.h"
93 #include "gromacs/utility/basenetwork.h"
94 #include "gromacs/utility/cstringutil.h"
95 #include "gromacs/utility/exceptions.h"
96 #include "gromacs/utility/fatalerror.h"
97 #include "gromacs/utility/gmxmpi.h"
98 #include "gromacs/utility/logger.h"
99 #include "gromacs/utility/real.h"
100 #include "gromacs/utility/smalloc.h"
101 #include "gromacs/utility/strconvert.h"
102 #include "gromacs/utility/stringstream.h"
103 #include "gromacs/utility/stringutil.h"
104 #include "gromacs/utility/textwriter.h"
106 #include "atomdistribution.h"
108 #include "cellsizes.h"
109 #include "distribute.h"
110 #include "domdec_constraints.h"
111 #include "domdec_internal.h"
112 #include "domdec_setup.h"
113 #include "domdec_vsite.h"
114 #include "redistribute.h"
117 // TODO remove this when moving domdec into gmx namespace
118 using gmx::DdRankOrder;
119 using gmx::DlbOption;
120 using gmx::DomdecOptions;
122 static const char* edlbs_names[int(DlbState::nr)] = { "off", "auto", "locked", "on", "on" };
124 /* The size per atom group of the cggl_flag buffer in gmx_domdec_comm_t */
127 /* The flags for the cggl_flag buffer in gmx_domdec_comm_t */
128 #define DD_FLAG_NRCG 65535
129 #define DD_FLAG_FW(d) (1 << (16 + (d)*2))
130 #define DD_FLAG_BW(d) (1 << (16 + (d)*2 + 1))
132 /* The DD zone order */
133 static const ivec dd_zo[DD_MAXZONE] = { { 0, 0, 0 }, { 1, 0, 0 }, { 1, 1, 0 }, { 0, 1, 0 },
134 { 0, 1, 1 }, { 0, 0, 1 }, { 1, 0, 1 }, { 1, 1, 1 } };
136 /* The non-bonded zone-pair setup for domain decomposition
137 * The first number is the i-zone, the second number the first j-zone seen by
138 * this i-zone, the third number the last+1 j-zone seen by this i-zone.
139 * As is, this is for 3D decomposition, where there are 4 i-zones.
140 * With 2D decomposition use only the first 2 i-zones and a last+1 j-zone of 4.
141 * With 1D decomposition use only the first i-zone and a last+1 j-zone of 2.
143 static const int ddNonbondedZonePairRanges[DD_MAXIZONE][3] = { { 0, 0, 8 },
150 #define dd_index(n,i) ((((i)[ZZ]*(n)[YY] + (i)[YY])*(n)[XX]) + (i)[XX])
152 static void index2xyz(ivec nc,int ind,ivec xyz)
154 xyz[XX] = ind % nc[XX];
155 xyz[YY] = (ind / nc[XX]) % nc[YY];
156 xyz[ZZ] = ind / (nc[YY]*nc[XX]);
160 static void ddindex2xyz(const ivec nc, int ind, ivec xyz)
162 xyz[XX] = ind / (nc[YY] * nc[ZZ]);
163 xyz[YY] = (ind / nc[ZZ]) % nc[YY];
164 xyz[ZZ] = ind % nc[ZZ];
167 static int ddcoord2ddnodeid(gmx_domdec_t* dd, ivec c)
171 const CartesianRankSetup& cartSetup = dd->comm->cartesianRankSetup;
172 const int ddindex = dd_index(dd->nc, c);
173 if (cartSetup.bCartesianPP_PME)
175 ddnodeid = cartSetup.ddindex2ddnodeid[ddindex];
177 else if (cartSetup.bCartesianPP)
180 MPI_Cart_rank(dd->mpi_comm_all, c, &ddnodeid);
191 int ddglatnr(const gmx_domdec_t* dd, int i)
201 if (i >= dd->comm->atomRanges.numAtomsTotal())
204 "glatnr called with %d, which is larger than the local number of atoms (%d)",
205 i, dd->comm->atomRanges.numAtomsTotal());
207 atnr = dd->globalAtomIndices[i] + 1;
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 int dd_numAtomsZones(const gmx_domdec_t& dd)
244 return dd.comm->atomRanges.end(DDAtomRanges::Type::Zones);
247 int dd_numHomeAtoms(const gmx_domdec_t& dd)
249 return dd.comm->atomRanges.numHomeAtoms();
252 int dd_natoms_mdatoms(const gmx_domdec_t* dd)
254 /* We currently set mdatoms entries for all atoms:
255 * local + non-local + communicated for vsite + constraints
258 return dd->comm->atomRanges.numAtomsTotal();
261 int dd_natoms_vsite(const gmx_domdec_t* dd)
263 return dd->comm->atomRanges.end(DDAtomRanges::Type::Vsites);
266 void dd_get_constraint_range(const gmx_domdec_t* dd, int* at_start, int* at_end)
268 *at_start = dd->comm->atomRanges.start(DDAtomRanges::Type::Constraints);
269 *at_end = dd->comm->atomRanges.end(DDAtomRanges::Type::Constraints);
272 void dd_move_x(gmx_domdec_t* dd, const matrix box, gmx::ArrayRef<gmx::RVec> x, gmx_wallcycle* wcycle)
274 wallcycle_start(wcycle, ewcMOVEX);
277 gmx_domdec_comm_t* comm;
278 gmx_domdec_comm_dim_t* cd;
279 rvec shift = { 0, 0, 0 };
280 gmx_bool bPBC, bScrew;
285 nat_tot = comm->atomRanges.numHomeAtoms();
286 for (int d = 0; d < dd->ndim; d++)
288 bPBC = (dd->ci[dd->dim[d]] == 0);
289 bScrew = (bPBC && dd->unitCellInfo.haveScrewPBC && dd->dim[d] == XX);
292 copy_rvec(box[dd->dim[d]], shift);
295 for (const gmx_domdec_ind_t& ind : cd->ind)
297 DDBufferAccess<gmx::RVec> sendBufferAccess(comm->rvecBuffer, ind.nsend[nzone + 1]);
298 gmx::ArrayRef<gmx::RVec>& sendBuffer = sendBufferAccess.buffer;
302 for (int j : ind.index)
304 sendBuffer[n] = x[j];
310 for (int j : ind.index)
312 /* We need to shift the coordinates */
313 for (int d = 0; d < DIM; d++)
315 sendBuffer[n][d] = x[j][d] + shift[d];
322 for (int j : ind.index)
325 sendBuffer[n][XX] = x[j][XX] + shift[XX];
327 * This operation requires a special shift force
328 * treatment, which is performed in calc_vir.
330 sendBuffer[n][YY] = box[YY][YY] - x[j][YY];
331 sendBuffer[n][ZZ] = box[ZZ][ZZ] - x[j][ZZ];
336 DDBufferAccess<gmx::RVec> receiveBufferAccess(
337 comm->rvecBuffer2, cd->receiveInPlace ? 0 : ind.nrecv[nzone + 1]);
339 gmx::ArrayRef<gmx::RVec> receiveBuffer;
340 if (cd->receiveInPlace)
342 receiveBuffer = gmx::arrayRefFromArray(x.data() + nat_tot, ind.nrecv[nzone + 1]);
346 receiveBuffer = receiveBufferAccess.buffer;
348 /* Send and receive the coordinates */
349 ddSendrecv(dd, d, dddirBackward, sendBuffer, receiveBuffer);
351 if (!cd->receiveInPlace)
354 for (int zone = 0; zone < nzone; zone++)
356 for (int i = ind.cell2at0[zone]; i < ind.cell2at1[zone]; i++)
358 x[i] = receiveBuffer[j++];
362 nat_tot += ind.nrecv[nzone + 1];
367 wallcycle_stop(wcycle, ewcMOVEX);
370 void dd_move_f(gmx_domdec_t* dd, gmx::ForceWithShiftForces* forceWithShiftForces, gmx_wallcycle* wcycle)
372 wallcycle_start(wcycle, ewcMOVEF);
374 gmx::ArrayRef<gmx::RVec> f = forceWithShiftForces->force();
375 gmx::ArrayRef<gmx::RVec> fshift = forceWithShiftForces->shiftForces();
377 gmx_domdec_comm_t& comm = *dd->comm;
378 int nzone = comm.zones.n / 2;
379 int nat_tot = comm.atomRanges.end(DDAtomRanges::Type::Zones);
380 for (int d = dd->ndim - 1; d >= 0; d--)
382 /* Only forces in domains near the PBC boundaries need to
383 consider PBC in the treatment of fshift */
384 const bool shiftForcesNeedPbc =
385 (forceWithShiftForces->computeVirial() && dd->ci[dd->dim[d]] == 0);
386 const bool applyScrewPbc =
387 (shiftForcesNeedPbc && dd->unitCellInfo.haveScrewPBC && dd->dim[d] == XX);
388 /* Determine which shift vector we need */
389 ivec vis = { 0, 0, 0 };
391 const int is = IVEC2IS(vis);
393 /* Loop over the pulses */
394 const gmx_domdec_comm_dim_t& cd = comm.cd[d];
395 for (int p = cd.numPulses() - 1; p >= 0; p--)
397 const gmx_domdec_ind_t& ind = cd.ind[p];
398 DDBufferAccess<gmx::RVec> receiveBufferAccess(comm.rvecBuffer, ind.nsend[nzone + 1]);
399 gmx::ArrayRef<gmx::RVec>& receiveBuffer = receiveBufferAccess.buffer;
401 nat_tot -= ind.nrecv[nzone + 1];
403 DDBufferAccess<gmx::RVec> sendBufferAccess(
404 comm.rvecBuffer2, cd.receiveInPlace ? 0 : ind.nrecv[nzone + 1]);
406 gmx::ArrayRef<gmx::RVec> sendBuffer;
407 if (cd.receiveInPlace)
409 sendBuffer = gmx::arrayRefFromArray(f.data() + nat_tot, ind.nrecv[nzone + 1]);
413 sendBuffer = sendBufferAccess.buffer;
415 for (int zone = 0; zone < nzone; zone++)
417 for (int i = ind.cell2at0[zone]; i < ind.cell2at1[zone]; i++)
419 sendBuffer[j++] = f[i];
423 /* Communicate the forces */
424 ddSendrecv(dd, d, dddirForward, sendBuffer, receiveBuffer);
425 /* Add the received forces */
427 if (!shiftForcesNeedPbc)
429 for (int j : ind.index)
431 for (int d = 0; d < DIM; d++)
433 f[j][d] += receiveBuffer[n][d];
438 else if (!applyScrewPbc)
440 for (int j : ind.index)
442 for (int d = 0; d < DIM; d++)
444 f[j][d] += receiveBuffer[n][d];
446 /* Add this force to the shift force */
447 for (int d = 0; d < DIM; d++)
449 fshift[is][d] += receiveBuffer[n][d];
456 for (int j : ind.index)
458 /* Rotate the force */
459 f[j][XX] += receiveBuffer[n][XX];
460 f[j][YY] -= receiveBuffer[n][YY];
461 f[j][ZZ] -= receiveBuffer[n][ZZ];
462 if (shiftForcesNeedPbc)
464 /* Add this force to the shift force */
465 for (int d = 0; d < DIM; d++)
467 fshift[is][d] += receiveBuffer[n][d];
476 wallcycle_stop(wcycle, ewcMOVEF);
479 /* Convenience function for extracting a real buffer from an rvec buffer
481 * To reduce the number of temporary communication buffers and avoid
482 * cache polution, we reuse gmx::RVec buffers for storing reals.
483 * This functions return a real buffer reference with the same number
484 * of elements as the gmx::RVec buffer (so 1/3 of the size in bytes).
486 static gmx::ArrayRef<real> realArrayRefFromRvecArrayRef(gmx::ArrayRef<gmx::RVec> arrayRef)
488 return gmx::arrayRefFromArray(as_rvec_array(arrayRef.data())[0], arrayRef.size());
491 void dd_atom_spread_real(gmx_domdec_t* dd, real v[])
494 gmx_domdec_comm_t* comm;
495 gmx_domdec_comm_dim_t* cd;
500 nat_tot = comm->atomRanges.numHomeAtoms();
501 for (int d = 0; d < dd->ndim; d++)
504 for (const gmx_domdec_ind_t& ind : cd->ind)
506 /* Note: We provision for RVec instead of real, so a factor of 3
507 * more than needed. The buffer actually already has this size
508 * and we pass a plain pointer below, so this does not matter.
510 DDBufferAccess<gmx::RVec> sendBufferAccess(comm->rvecBuffer, ind.nsend[nzone + 1]);
511 gmx::ArrayRef<real> sendBuffer = realArrayRefFromRvecArrayRef(sendBufferAccess.buffer);
513 for (int j : ind.index)
515 sendBuffer[n++] = v[j];
518 DDBufferAccess<gmx::RVec> receiveBufferAccess(
519 comm->rvecBuffer2, cd->receiveInPlace ? 0 : ind.nrecv[nzone + 1]);
521 gmx::ArrayRef<real> receiveBuffer;
522 if (cd->receiveInPlace)
524 receiveBuffer = gmx::arrayRefFromArray(v + nat_tot, ind.nrecv[nzone + 1]);
528 receiveBuffer = realArrayRefFromRvecArrayRef(receiveBufferAccess.buffer);
530 /* Send and receive the data */
531 ddSendrecv(dd, d, dddirBackward, sendBuffer, receiveBuffer);
532 if (!cd->receiveInPlace)
535 for (int zone = 0; zone < nzone; zone++)
537 for (int i = ind.cell2at0[zone]; i < ind.cell2at1[zone]; i++)
539 v[i] = receiveBuffer[j++];
543 nat_tot += ind.nrecv[nzone + 1];
549 void dd_atom_sum_real(gmx_domdec_t* dd, real v[])
552 gmx_domdec_comm_t* comm;
553 gmx_domdec_comm_dim_t* cd;
557 nzone = comm->zones.n / 2;
558 nat_tot = comm->atomRanges.end(DDAtomRanges::Type::Zones);
559 for (int d = dd->ndim - 1; d >= 0; d--)
562 for (int p = cd->numPulses() - 1; p >= 0; p--)
564 const gmx_domdec_ind_t& ind = cd->ind[p];
566 /* Note: We provision for RVec instead of real, so a factor of 3
567 * more than needed. The buffer actually already has this size
568 * and we typecast, so this works as intended.
570 DDBufferAccess<gmx::RVec> receiveBufferAccess(comm->rvecBuffer, ind.nsend[nzone + 1]);
571 gmx::ArrayRef<real> receiveBuffer = realArrayRefFromRvecArrayRef(receiveBufferAccess.buffer);
572 nat_tot -= ind.nrecv[nzone + 1];
574 DDBufferAccess<gmx::RVec> sendBufferAccess(
575 comm->rvecBuffer2, cd->receiveInPlace ? 0 : ind.nrecv[nzone + 1]);
577 gmx::ArrayRef<real> sendBuffer;
578 if (cd->receiveInPlace)
580 sendBuffer = gmx::arrayRefFromArray(v + nat_tot, ind.nrecv[nzone + 1]);
584 sendBuffer = realArrayRefFromRvecArrayRef(sendBufferAccess.buffer);
586 for (int zone = 0; zone < nzone; zone++)
588 for (int i = ind.cell2at0[zone]; i < ind.cell2at1[zone]; i++)
590 sendBuffer[j++] = v[i];
594 /* Communicate the forces */
595 ddSendrecv(dd, d, dddirForward, sendBuffer, receiveBuffer);
596 /* Add the received forces */
598 for (int j : ind.index)
600 v[j] += receiveBuffer[n];
608 real dd_cutoff_multibody(const gmx_domdec_t* dd)
610 const gmx_domdec_comm_t& comm = *dd->comm;
611 const DDSystemInfo& systemInfo = comm.systemInfo;
614 if (systemInfo.haveInterDomainMultiBodyBondeds)
616 if (comm.cutoff_mbody > 0)
618 r = comm.cutoff_mbody;
622 /* cutoff_mbody=0 means we do not have DLB */
623 r = comm.cellsize_min[dd->dim[0]];
624 for (int di = 1; di < dd->ndim; di++)
626 r = std::min(r, comm.cellsize_min[dd->dim[di]]);
628 if (comm.systemInfo.filterBondedCommunication)
630 r = std::max(r, comm.cutoff_mbody);
634 r = std::min(r, systemInfo.cutoff);
642 real dd_cutoff_twobody(const gmx_domdec_t* dd)
646 r_mb = dd_cutoff_multibody(dd);
648 return std::max(dd->comm->systemInfo.cutoff, r_mb);
652 static void dd_cart_coord2pmecoord(const DDRankSetup& ddRankSetup,
653 const CartesianRankSetup& cartSetup,
657 const int nc = ddRankSetup.numPPCells[cartSetup.cartpmedim];
658 const int ntot = cartSetup.ntot[cartSetup.cartpmedim];
659 copy_ivec(coord, coord_pme);
660 coord_pme[cartSetup.cartpmedim] =
661 nc + (coord[cartSetup.cartpmedim] * (ntot - nc) + (ntot - nc) / 2) / nc;
664 /* Returns the PME rank index in 0...npmenodes-1 for the PP cell with index ddCellIndex */
665 static int ddindex2pmeindex(const DDRankSetup& ddRankSetup, const int ddCellIndex)
667 const int npp = ddRankSetup.numPPRanks;
668 const int npme = ddRankSetup.numRanksDoingPme;
670 /* Here we assign a PME node to communicate with this DD node
671 * by assuming that the major index of both is x.
672 * We add npme/2 to obtain an even distribution.
674 return (ddCellIndex * npme + npme / 2) / npp;
677 static std::vector<int> dd_interleaved_pme_ranks(const DDRankSetup& ddRankSetup)
679 std::vector<int> pmeRanks(ddRankSetup.numRanksDoingPme);
682 for (int i = 0; i < ddRankSetup.numPPRanks; i++)
684 const int p0 = ddindex2pmeindex(ddRankSetup, i);
685 const int p1 = ddindex2pmeindex(ddRankSetup, i + 1);
686 if (i + 1 == ddRankSetup.numPPRanks || p1 > p0)
690 fprintf(debug, "pme_rank[%d] = %d\n", n, i + 1 + n);
692 pmeRanks[n] = i + 1 + n;
700 static int gmx_ddcoord2pmeindex(const t_commrec* cr, int x, int y, int z)
708 if (dd->comm->bCartesian) {
709 gmx_ddindex2xyz(dd->nc,ddindex,coords);
710 dd_coords2pmecoords(dd,coords,coords_pme);
711 copy_ivec(dd->ntot,nc);
712 nc[dd->cartpmedim] -= dd->nc[dd->cartpmedim];
713 coords_pme[dd->cartpmedim] -= dd->nc[dd->cartpmedim];
715 slab = (coords_pme[XX]*nc[YY] + coords_pme[YY])*nc[ZZ] + coords_pme[ZZ];
717 slab = (ddindex*cr->npmenodes + cr->npmenodes/2)/dd->nnodes;
723 slab = ddindex2pmeindex(dd->comm->ddRankSetup, dd_index(dd->nc, coords));
728 static int ddcoord2simnodeid(const t_commrec* cr, int x, int y, int z)
730 const CartesianRankSetup& cartSetup = cr->dd->comm->cartesianRankSetup;
731 ivec coords = { x, y, z };
734 if (cartSetup.bCartesianPP_PME)
737 MPI_Cart_rank(cr->mpi_comm_mysim, coords, &nodeid);
742 int ddindex = dd_index(cr->dd->nc, coords);
743 if (cartSetup.bCartesianPP)
745 nodeid = cartSetup.ddindex2simnodeid[ddindex];
749 if (cr->dd->comm->ddRankSetup.usePmeOnlyRanks)
751 nodeid = ddindex + gmx_ddcoord2pmeindex(cr, x, y, z);
763 static int dd_simnode2pmenode(const DDRankSetup& ddRankSetup,
764 const CartesianRankSetup& cartSetup,
765 gmx::ArrayRef<const int> pmeRanks,
766 const t_commrec gmx_unused* cr,
767 const int sim_nodeid)
771 /* This assumes a uniform x domain decomposition grid cell size */
772 if (cartSetup.bCartesianPP_PME)
775 ivec coord, coord_pme;
776 MPI_Cart_coords(cr->mpi_comm_mysim, sim_nodeid, DIM, coord);
777 if (coord[cartSetup.cartpmedim] < ddRankSetup.numPPCells[cartSetup.cartpmedim])
779 /* This is a PP rank */
780 dd_cart_coord2pmecoord(ddRankSetup, cartSetup, coord, coord_pme);
781 MPI_Cart_rank(cr->mpi_comm_mysim, coord_pme, &pmenode);
785 else if (cartSetup.bCartesianPP)
787 if (sim_nodeid < ddRankSetup.numPPRanks)
789 pmenode = ddRankSetup.numPPRanks + ddindex2pmeindex(ddRankSetup, sim_nodeid);
794 /* This assumes DD cells with identical x coordinates
795 * are numbered sequentially.
797 if (pmeRanks.empty())
799 if (sim_nodeid < ddRankSetup.numPPRanks)
801 /* The DD index equals the nodeid */
802 pmenode = ddRankSetup.numPPRanks + ddindex2pmeindex(ddRankSetup, sim_nodeid);
808 while (sim_nodeid > pmeRanks[i])
812 if (sim_nodeid < pmeRanks[i])
814 pmenode = pmeRanks[i];
822 NumPmeDomains getNumPmeDomains(const gmx_domdec_t* dd)
826 return { dd->comm->ddRankSetup.npmenodes_x, dd->comm->ddRankSetup.npmenodes_y };
834 std::vector<int> get_pme_ddranks(const t_commrec* cr, const int pmenodeid)
836 const DDRankSetup& ddRankSetup = cr->dd->comm->ddRankSetup;
837 const CartesianRankSetup& cartSetup = cr->dd->comm->cartesianRankSetup;
838 GMX_RELEASE_ASSERT(ddRankSetup.usePmeOnlyRanks,
839 "This function should only be called when PME-only ranks are in use");
840 std::vector<int> ddranks;
841 ddranks.reserve((ddRankSetup.numPPRanks + ddRankSetup.numRanksDoingPme - 1) / ddRankSetup.numRanksDoingPme);
843 for (int x = 0; x < ddRankSetup.numPPCells[XX]; x++)
845 for (int y = 0; y < ddRankSetup.numPPCells[YY]; y++)
847 for (int z = 0; z < ddRankSetup.numPPCells[ZZ]; z++)
849 if (cartSetup.bCartesianPP_PME)
851 ivec coord = { x, y, z };
853 dd_cart_coord2pmecoord(ddRankSetup, cartSetup, coord, coord_pme);
854 if (cr->dd->ci[XX] == coord_pme[XX] && cr->dd->ci[YY] == coord_pme[YY]
855 && cr->dd->ci[ZZ] == coord_pme[ZZ])
857 ddranks.push_back(ddcoord2simnodeid(cr, x, y, z));
862 /* The slab corresponds to the nodeid in the PME group */
863 if (gmx_ddcoord2pmeindex(cr, x, y, z) == pmenodeid)
865 ddranks.push_back(ddcoord2simnodeid(cr, x, y, z));
874 static gmx_bool receive_vir_ener(const gmx_domdec_t* dd, gmx::ArrayRef<const int> pmeRanks, const t_commrec* cr)
876 gmx_bool bReceive = TRUE;
878 const DDRankSetup& ddRankSetup = dd->comm->ddRankSetup;
879 if (ddRankSetup.usePmeOnlyRanks)
881 const CartesianRankSetup& cartSetup = dd->comm->cartesianRankSetup;
882 if (cartSetup.bCartesianPP_PME)
885 int pmenode = dd_simnode2pmenode(ddRankSetup, cartSetup, pmeRanks, cr, cr->sim_nodeid);
887 MPI_Cart_coords(cr->mpi_comm_mysim, cr->sim_nodeid, DIM, coords);
888 coords[cartSetup.cartpmedim]++;
889 if (coords[cartSetup.cartpmedim] < dd->nc[cartSetup.cartpmedim])
892 MPI_Cart_rank(cr->mpi_comm_mysim, coords, &rank);
893 if (dd_simnode2pmenode(ddRankSetup, cartSetup, pmeRanks, cr, rank) == pmenode)
895 /* This is not the last PP node for pmenode */
902 "Without MPI we should not have Cartesian PP-PME with #PMEnodes < #DDnodes");
907 int pmenode = dd_simnode2pmenode(ddRankSetup, cartSetup, pmeRanks, cr, cr->sim_nodeid);
908 if (cr->sim_nodeid + 1 < cr->nnodes
909 && dd_simnode2pmenode(ddRankSetup, cartSetup, pmeRanks, cr, cr->sim_nodeid + 1) == pmenode)
911 /* This is not the last PP node for pmenode */
920 static void set_slb_pme_dim_f(gmx_domdec_t* dd, int dim, real** dim_f)
922 gmx_domdec_comm_t* comm;
927 snew(*dim_f, dd->nc[dim] + 1);
929 for (i = 1; i < dd->nc[dim]; i++)
931 if (comm->slb_frac[dim])
933 (*dim_f)[i] = (*dim_f)[i - 1] + comm->slb_frac[dim][i - 1];
937 (*dim_f)[i] = static_cast<real>(i) / static_cast<real>(dd->nc[dim]);
940 (*dim_f)[dd->nc[dim]] = 1;
943 static void init_ddpme(gmx_domdec_t* dd, gmx_ddpme_t* ddpme, int dimind)
945 const DDRankSetup& ddRankSetup = dd->comm->ddRankSetup;
947 if (dimind == 0 && dd->dim[0] == YY && ddRankSetup.npmenodes_x == 1)
955 ddpme->dim_match = (ddpme->dim == dd->dim[dimind]);
957 ddpme->nslab = (ddpme->dim == 0 ? ddRankSetup.npmenodes_x : ddRankSetup.npmenodes_y);
959 if (ddpme->nslab <= 1)
964 const int nso = ddRankSetup.numRanksDoingPme / ddpme->nslab;
965 /* Determine for each PME slab the PP location range for dimension dim */
966 snew(ddpme->pp_min, ddpme->nslab);
967 snew(ddpme->pp_max, ddpme->nslab);
968 for (int slab = 0; slab < ddpme->nslab; slab++)
970 ddpme->pp_min[slab] = dd->nc[dd->dim[dimind]] - 1;
971 ddpme->pp_max[slab] = 0;
973 for (int i = 0; i < dd->nnodes; i++)
976 ddindex2xyz(dd->nc, i, xyz);
977 /* For y only use our y/z slab.
978 * This assumes that the PME x grid size matches the DD grid size.
980 if (dimind == 0 || xyz[XX] == dd->ci[XX])
982 const int pmeindex = ddindex2pmeindex(ddRankSetup, i);
986 slab = pmeindex / nso;
990 slab = pmeindex % ddpme->nslab;
992 ddpme->pp_min[slab] = std::min(ddpme->pp_min[slab], xyz[dimind]);
993 ddpme->pp_max[slab] = std::max(ddpme->pp_max[slab], xyz[dimind]);
997 set_slb_pme_dim_f(dd, ddpme->dim, &ddpme->slb_dim_f);
1000 int dd_pme_maxshift_x(const gmx_domdec_t* dd)
1002 const DDRankSetup& ddRankSetup = dd->comm->ddRankSetup;
1004 if (ddRankSetup.ddpme[0].dim == XX)
1006 return ddRankSetup.ddpme[0].maxshift;
1014 int dd_pme_maxshift_y(const gmx_domdec_t* dd)
1016 const DDRankSetup& ddRankSetup = dd->comm->ddRankSetup;
1018 if (ddRankSetup.ddpme[0].dim == YY)
1020 return ddRankSetup.ddpme[0].maxshift;
1022 else if (ddRankSetup.npmedecompdim >= 2 && ddRankSetup.ddpme[1].dim == YY)
1024 return ddRankSetup.ddpme[1].maxshift;
1032 bool ddHaveSplitConstraints(const gmx_domdec_t& dd)
1034 return dd.comm->systemInfo.haveSplitConstraints;
1037 bool ddUsesUpdateGroups(const gmx_domdec_t& dd)
1039 return dd.comm->systemInfo.useUpdateGroups;
1042 void dd_cycles_add(const gmx_domdec_t* dd, float cycles, int ddCycl)
1044 /* Note that the cycles value can be incorrect, either 0 or some
1045 * extremely large value, when our thread migrated to another core
1046 * with an unsynchronized cycle counter. If this happens less often
1047 * that once per nstlist steps, this will not cause issues, since
1048 * we later subtract the maximum value from the sum over nstlist steps.
1049 * A zero count will slightly lower the total, but that's a small effect.
1050 * Note that the main purpose of the subtraction of the maximum value
1051 * is to avoid throwing off the load balancing when stalls occur due
1052 * e.g. system activity or network congestion.
1054 dd->comm->cycl[ddCycl] += cycles;
1055 dd->comm->cycl_n[ddCycl]++;
1056 if (cycles > dd->comm->cycl_max[ddCycl])
1058 dd->comm->cycl_max[ddCycl] = cycles;
1063 static void make_load_communicator(gmx_domdec_t* dd, int dim_ind, ivec loc)
1068 gmx_bool bPartOfGroup = FALSE;
1070 dim = dd->dim[dim_ind];
1071 copy_ivec(loc, loc_c);
1072 for (i = 0; i < dd->nc[dim]; i++)
1075 rank = dd_index(dd->nc, loc_c);
1076 if (rank == dd->rank)
1078 /* This process is part of the group */
1079 bPartOfGroup = TRUE;
1082 MPI_Comm_split(dd->mpi_comm_all, bPartOfGroup ? 0 : MPI_UNDEFINED, dd->rank, &c_row);
1085 dd->comm->mpi_comm_load[dim_ind] = c_row;
1086 if (!isDlbDisabled(dd->comm))
1088 DDCellsizesWithDlb& cellsizes = dd->comm->cellsizesWithDlb[dim_ind];
1090 if (dd->ci[dim] == dd->master_ci[dim])
1092 /* This is the root process of this row */
1093 cellsizes.rowMaster = std::make_unique<RowMaster>();
1095 RowMaster& rowMaster = *cellsizes.rowMaster;
1096 rowMaster.cellFrac.resize(ddCellFractionBufferSize(dd, dim_ind));
1097 rowMaster.oldCellFrac.resize(dd->nc[dim] + 1);
1098 rowMaster.isCellMin.resize(dd->nc[dim]);
1101 rowMaster.bounds.resize(dd->nc[dim]);
1103 rowMaster.buf_ncd.resize(dd->nc[dim]);
1107 /* This is not a root process, we only need to receive cell_f */
1108 cellsizes.fracRow.resize(ddCellFractionBufferSize(dd, dim_ind));
1111 if (dd->ci[dim] == dd->master_ci[dim])
1113 snew(dd->comm->load[dim_ind].load, dd->nc[dim] * DD_NLOAD_MAX);
1119 void dd_setup_dlb_resource_sharing(const t_commrec* cr, int gpu_id)
1122 int physicalnode_id_hash;
1124 MPI_Comm mpi_comm_pp_physicalnode;
1126 if (!thisRankHasDuty(cr, DUTY_PP) || gpu_id < 0)
1128 /* Only ranks with short-ranged tasks (currently) use GPUs.
1129 * If we don't have GPUs assigned, there are no resources to share.
1134 physicalnode_id_hash = gmx_physicalnode_id_hash();
1140 fprintf(debug, "dd_setup_dd_dlb_gpu_sharing:\n");
1141 fprintf(debug, "DD PP rank %d physical node hash %d gpu_id %d\n", dd->rank,
1142 physicalnode_id_hash, gpu_id);
1144 /* Split the PP communicator over the physical nodes */
1145 /* TODO: See if we should store this (before), as it's also used for
1146 * for the nodecomm summation.
1148 // TODO PhysicalNodeCommunicator could be extended/used to handle
1149 // the need for per-node per-group communicators.
1150 MPI_Comm_split(dd->mpi_comm_all, physicalnode_id_hash, dd->rank, &mpi_comm_pp_physicalnode);
1151 MPI_Comm_split(mpi_comm_pp_physicalnode, gpu_id, dd->rank, &dd->comm->mpi_comm_gpu_shared);
1152 MPI_Comm_free(&mpi_comm_pp_physicalnode);
1153 MPI_Comm_size(dd->comm->mpi_comm_gpu_shared, &dd->comm->nrank_gpu_shared);
1157 fprintf(debug, "nrank_gpu_shared %d\n", dd->comm->nrank_gpu_shared);
1160 /* Note that some ranks could share a GPU, while others don't */
1162 if (dd->comm->nrank_gpu_shared == 1)
1164 MPI_Comm_free(&dd->comm->mpi_comm_gpu_shared);
1167 GMX_UNUSED_VALUE(cr);
1168 GMX_UNUSED_VALUE(gpu_id);
1172 static void make_load_communicators(gmx_domdec_t gmx_unused* dd)
1175 int dim0, dim1, i, j;
1180 fprintf(debug, "Making load communicators\n");
1183 dd->comm->load = new domdec_load_t[std::max(dd->ndim, 1)];
1184 snew(dd->comm->mpi_comm_load, std::max(dd->ndim, 1));
1192 make_load_communicator(dd, 0, loc);
1196 for (i = 0; i < dd->nc[dim0]; i++)
1199 make_load_communicator(dd, 1, loc);
1205 for (i = 0; i < dd->nc[dim0]; i++)
1209 for (j = 0; j < dd->nc[dim1]; j++)
1212 make_load_communicator(dd, 2, loc);
1219 fprintf(debug, "Finished making load communicators\n");
1224 /*! \brief Sets up the relation between neighboring domains and zones */
1225 static void setup_neighbor_relations(gmx_domdec_t* dd)
1229 gmx_domdec_zones_t* zones;
1230 GMX_ASSERT((dd->ndim >= 0) && (dd->ndim <= DIM), "Must have valid number of dimensions for DD");
1232 for (d = 0; d < dd->ndim; d++)
1235 copy_ivec(dd->ci, tmp);
1236 tmp[dim] = (tmp[dim] + 1) % dd->nc[dim];
1237 dd->neighbor[d][0] = ddcoord2ddnodeid(dd, tmp);
1238 copy_ivec(dd->ci, tmp);
1239 tmp[dim] = (tmp[dim] - 1 + dd->nc[dim]) % dd->nc[dim];
1240 dd->neighbor[d][1] = ddcoord2ddnodeid(dd, tmp);
1243 fprintf(debug, "DD rank %d neighbor ranks in dir %d are + %d - %d\n", dd->rank, dim,
1244 dd->neighbor[d][0], dd->neighbor[d][1]);
1248 int nzone = (1 << dd->ndim);
1249 int nizone = (1 << std::max(dd->ndim - 1, 0));
1250 assert(nizone >= 1 && nizone <= DD_MAXIZONE);
1252 zones = &dd->comm->zones;
1254 for (int i = 0; i < nzone; i++)
1257 clear_ivec(zones->shift[i]);
1258 for (d = 0; d < dd->ndim; d++)
1260 zones->shift[i][dd->dim[d]] = dd_zo[i][m++];
1265 for (int i = 0; i < nzone; i++)
1267 for (d = 0; d < DIM; d++)
1269 s[d] = dd->ci[d] - zones->shift[i][d];
1274 else if (s[d] >= dd->nc[d])
1280 for (int iZoneIndex = 0; iZoneIndex < nizone; iZoneIndex++)
1283 ddNonbondedZonePairRanges[iZoneIndex][0] == iZoneIndex,
1284 "The first element for each ddNonbondedZonePairRanges should match its index");
1286 DDPairInteractionRanges iZone;
1287 iZone.iZoneIndex = iZoneIndex;
1288 /* dd_zp3 is for 3D decomposition, for fewer dimensions use only
1289 * j-zones up to nzone.
1291 iZone.jZoneRange = gmx::Range<int>(std::min(ddNonbondedZonePairRanges[iZoneIndex][1], nzone),
1292 std::min(ddNonbondedZonePairRanges[iZoneIndex][2], nzone));
1293 for (dim = 0; dim < DIM; dim++)
1295 if (dd->nc[dim] == 1)
1297 /* All shifts should be allowed */
1298 iZone.shift0[dim] = -1;
1299 iZone.shift1[dim] = 1;
1303 /* Determine the min/max j-zone shift wrt the i-zone */
1304 iZone.shift0[dim] = 1;
1305 iZone.shift1[dim] = -1;
1306 for (int jZone : iZone.jZoneRange)
1308 int shift_diff = zones->shift[jZone][dim] - zones->shift[iZoneIndex][dim];
1309 if (shift_diff < iZone.shift0[dim])
1311 iZone.shift0[dim] = shift_diff;
1313 if (shift_diff > iZone.shift1[dim])
1315 iZone.shift1[dim] = shift_diff;
1321 zones->iZones.push_back(iZone);
1324 if (!isDlbDisabled(dd->comm))
1326 dd->comm->cellsizesWithDlb.resize(dd->ndim);
1329 if (dd->comm->ddSettings.recordLoad)
1331 make_load_communicators(dd);
1335 static void make_pp_communicator(const gmx::MDLogger& mdlog,
1337 t_commrec gmx_unused* cr,
1338 bool gmx_unused reorder)
1341 gmx_domdec_comm_t* comm = dd->comm;
1342 CartesianRankSetup& cartSetup = comm->cartesianRankSetup;
1344 if (cartSetup.bCartesianPP)
1346 /* Set up cartesian communication for the particle-particle part */
1348 .appendTextFormatted("Will use a Cartesian communicator: %d x %d x %d", dd->nc[XX],
1349 dd->nc[YY], dd->nc[ZZ]);
1352 for (int i = 0; i < DIM; i++)
1357 MPI_Cart_create(cr->mpi_comm_mygroup, DIM, dd->nc, periods, static_cast<int>(reorder), &comm_cart);
1358 /* We overwrite the old communicator with the new cartesian one */
1359 cr->mpi_comm_mygroup = comm_cart;
1362 dd->mpi_comm_all = cr->mpi_comm_mygroup;
1363 MPI_Comm_rank(dd->mpi_comm_all, &dd->rank);
1365 if (cartSetup.bCartesianPP_PME)
1367 /* Since we want to use the original cartesian setup for sim,
1368 * and not the one after split, we need to make an index.
1370 cartSetup.ddindex2ddnodeid.resize(dd->nnodes);
1371 cartSetup.ddindex2ddnodeid[dd_index(dd->nc, dd->ci)] = dd->rank;
1372 gmx_sumi(dd->nnodes, cartSetup.ddindex2ddnodeid.data(), cr);
1373 /* Get the rank of the DD master,
1374 * above we made sure that the master node is a PP node.
1385 MPI_Allreduce(&rank, &dd->masterrank, 1, MPI_INT, MPI_SUM, dd->mpi_comm_all);
1387 else if (cartSetup.bCartesianPP)
1389 if (!comm->ddRankSetup.usePmeOnlyRanks)
1391 /* The PP communicator is also
1392 * the communicator for this simulation
1394 cr->mpi_comm_mysim = cr->mpi_comm_mygroup;
1396 cr->nodeid = dd->rank;
1398 MPI_Cart_coords(dd->mpi_comm_all, dd->rank, DIM, dd->ci);
1400 /* We need to make an index to go from the coordinates
1401 * to the nodeid of this simulation.
1403 cartSetup.ddindex2simnodeid.resize(dd->nnodes);
1404 std::vector<int> buf(dd->nnodes);
1405 if (thisRankHasDuty(cr, DUTY_PP))
1407 buf[dd_index(dd->nc, dd->ci)] = cr->sim_nodeid;
1409 /* Communicate the ddindex to simulation nodeid index */
1410 MPI_Allreduce(buf.data(), cartSetup.ddindex2simnodeid.data(), dd->nnodes, MPI_INT, MPI_SUM,
1411 cr->mpi_comm_mysim);
1413 /* Determine the master coordinates and rank.
1414 * The DD master should be the same node as the master of this sim.
1416 for (int i = 0; i < dd->nnodes; i++)
1418 if (cartSetup.ddindex2simnodeid[i] == 0)
1420 ddindex2xyz(dd->nc, i, dd->master_ci);
1421 MPI_Cart_rank(dd->mpi_comm_all, dd->master_ci, &dd->masterrank);
1426 fprintf(debug, "The master rank is %d\n", dd->masterrank);
1431 /* No Cartesian communicators */
1432 /* We use the rank in dd->comm->all as DD index */
1433 ddindex2xyz(dd->nc, dd->rank, dd->ci);
1434 /* The simulation master nodeid is 0, so the DD master rank is also 0 */
1436 clear_ivec(dd->master_ci);
1441 .appendTextFormatted("Domain decomposition rank %d, coordinates %d %d %d\n", dd->rank,
1442 dd->ci[XX], dd->ci[YY], dd->ci[ZZ]);
1445 fprintf(debug, "Domain decomposition rank %d, coordinates %d %d %d\n\n", dd->rank,
1446 dd->ci[XX], dd->ci[YY], dd->ci[ZZ]);
1450 static void receive_ddindex2simnodeid(gmx_domdec_t* dd, t_commrec* cr)
1453 CartesianRankSetup& cartSetup = dd->comm->cartesianRankSetup;
1455 if (!cartSetup.bCartesianPP_PME && cartSetup.bCartesianPP)
1457 cartSetup.ddindex2simnodeid.resize(dd->nnodes);
1458 std::vector<int> buf(dd->nnodes);
1459 if (thisRankHasDuty(cr, DUTY_PP))
1461 buf[dd_index(dd->nc, dd->ci)] = cr->sim_nodeid;
1463 /* Communicate the ddindex to simulation nodeid index */
1464 MPI_Allreduce(buf.data(), cartSetup.ddindex2simnodeid.data(), dd->nnodes, MPI_INT, MPI_SUM,
1465 cr->mpi_comm_mysim);
1468 GMX_UNUSED_VALUE(dd);
1469 GMX_UNUSED_VALUE(cr);
1473 static CartesianRankSetup split_communicator(const gmx::MDLogger& mdlog,
1475 const DdRankOrder ddRankOrder,
1476 bool gmx_unused reorder,
1477 const DDRankSetup& ddRankSetup,
1479 std::vector<int>* pmeRanks)
1481 CartesianRankSetup cartSetup;
1483 cartSetup.bCartesianPP = (ddRankOrder == DdRankOrder::cartesian);
1484 cartSetup.bCartesianPP_PME = false;
1486 const ivec& numDDCells = ddRankSetup.numPPCells;
1487 /* Initially we set ntot to the number of PP cells */
1488 copy_ivec(numDDCells, cartSetup.ntot);
1490 if (cartSetup.bCartesianPP)
1492 const int numDDCellsTot = ddRankSetup.numPPRanks;
1494 for (int i = 1; i < DIM; i++)
1496 bDiv[i] = ((ddRankSetup.numRanksDoingPme * numDDCells[i]) % numDDCellsTot == 0);
1498 if (bDiv[YY] || bDiv[ZZ])
1500 cartSetup.bCartesianPP_PME = TRUE;
1501 /* If we have 2D PME decomposition, which is always in x+y,
1502 * we stack the PME only nodes in z.
1503 * Otherwise we choose the direction that provides the thinnest slab
1504 * of PME only nodes as this will have the least effect
1505 * on the PP communication.
1506 * But for the PME communication the opposite might be better.
1508 if (bDiv[ZZ] && (ddRankSetup.npmenodes_y > 1 || !bDiv[YY] || numDDCells[YY] > numDDCells[ZZ]))
1510 cartSetup.cartpmedim = ZZ;
1514 cartSetup.cartpmedim = YY;
1516 cartSetup.ntot[cartSetup.cartpmedim] +=
1517 (ddRankSetup.numRanksDoingPme * numDDCells[cartSetup.cartpmedim]) / numDDCellsTot;
1522 .appendTextFormatted(
1523 "Number of PME-only ranks (%d) is not a multiple of nx*ny (%d*%d) or "
1525 ddRankSetup.numRanksDoingPme, numDDCells[XX], numDDCells[YY],
1526 numDDCells[XX], numDDCells[ZZ]);
1528 .appendText("Will not use a Cartesian communicator for PP <-> PME\n");
1532 if (cartSetup.bCartesianPP_PME)
1539 .appendTextFormatted(
1540 "Will use a Cartesian communicator for PP <-> PME: %d x %d x %d",
1541 cartSetup.ntot[XX], cartSetup.ntot[YY], cartSetup.ntot[ZZ]);
1543 for (int i = 0; i < DIM; i++)
1548 MPI_Cart_create(cr->mpi_comm_mysim, DIM, cartSetup.ntot, periods, static_cast<int>(reorder),
1550 MPI_Comm_rank(comm_cart, &rank);
1551 if (MASTER(cr) && rank != 0)
1553 gmx_fatal(FARGS, "MPI rank 0 was renumbered by MPI_Cart_create, we do not allow this");
1556 /* With this assigment we loose the link to the original communicator
1557 * which will usually be MPI_COMM_WORLD, unless have multisim.
1559 cr->mpi_comm_mysim = comm_cart;
1560 cr->sim_nodeid = rank;
1562 MPI_Cart_coords(cr->mpi_comm_mysim, cr->sim_nodeid, DIM, ddCellIndex);
1565 .appendTextFormatted("Cartesian rank %d, coordinates %d %d %d\n", cr->sim_nodeid,
1566 ddCellIndex[XX], ddCellIndex[YY], ddCellIndex[ZZ]);
1568 if (ddCellIndex[cartSetup.cartpmedim] < numDDCells[cartSetup.cartpmedim])
1572 if (!ddRankSetup.usePmeOnlyRanks
1573 || ddCellIndex[cartSetup.cartpmedim] >= numDDCells[cartSetup.cartpmedim])
1575 cr->duty = DUTY_PME;
1578 /* Split the sim communicator into PP and PME only nodes */
1579 MPI_Comm_split(cr->mpi_comm_mysim, getThisRankDuties(cr),
1580 dd_index(cartSetup.ntot, ddCellIndex), &cr->mpi_comm_mygroup);
1582 GMX_UNUSED_VALUE(ddCellIndex);
1587 switch (ddRankOrder)
1589 case DdRankOrder::pp_pme:
1590 GMX_LOG(mdlog.info).appendText("Order of the ranks: PP first, PME last");
1592 case DdRankOrder::interleave:
1593 /* Interleave the PP-only and PME-only ranks */
1594 GMX_LOG(mdlog.info).appendText("Interleaving PP and PME ranks");
1595 *pmeRanks = dd_interleaved_pme_ranks(ddRankSetup);
1597 case DdRankOrder::cartesian: break;
1598 default: gmx_fatal(FARGS, "Invalid ddRankOrder=%d", static_cast<int>(ddRankOrder));
1601 if (dd_simnode2pmenode(ddRankSetup, cartSetup, *pmeRanks, cr, cr->sim_nodeid) == -1)
1603 cr->duty = DUTY_PME;
1610 /* Split the sim communicator into PP and PME only nodes */
1611 MPI_Comm_split(cr->mpi_comm_mysim, getThisRankDuties(cr), cr->nodeid, &cr->mpi_comm_mygroup);
1612 MPI_Comm_rank(cr->mpi_comm_mygroup, &cr->nodeid);
1617 .appendTextFormatted("This rank does only %s work.\n",
1618 thisRankHasDuty(cr, DUTY_PP) ? "particle-particle" : "PME-mesh");
1623 /*! \brief Makes the PP communicator and the PME communicator, when needed
1625 * Returns the Cartesian rank setup.
1626 * Sets \p cr->mpi_comm_mygroup
1627 * For PP ranks, sets the DD PP cell index in \p ddCellIndex.
1628 * With separate PME ranks in interleaved order, set the PME ranks in \p pmeRanks.
1630 static CartesianRankSetup makeGroupCommunicators(const gmx::MDLogger& mdlog,
1631 const DDSettings& ddSettings,
1632 const DdRankOrder ddRankOrder,
1633 const DDRankSetup& ddRankSetup,
1636 std::vector<int>* pmeRanks)
1638 CartesianRankSetup cartSetup;
1640 if (ddRankSetup.usePmeOnlyRanks)
1642 /* Split the communicator into a PP and PME part */
1643 cartSetup = split_communicator(mdlog, cr, ddRankOrder, ddSettings.useCartesianReorder,
1644 ddRankSetup, ddCellIndex, pmeRanks);
1648 /* All nodes do PP and PME */
1649 /* We do not require separate communicators */
1650 cr->mpi_comm_mygroup = cr->mpi_comm_mysim;
1652 cartSetup.bCartesianPP = false;
1653 cartSetup.bCartesianPP_PME = false;
1659 /*! \brief For PP ranks, sets or makes the communicator
1661 * For PME ranks get the rank id.
1662 * For PP only ranks, sets the PME-only rank.
1664 static void setupGroupCommunication(const gmx::MDLogger& mdlog,
1665 const DDSettings& ddSettings,
1666 gmx::ArrayRef<const int> pmeRanks,
1668 const int numAtomsInSystem,
1671 const DDRankSetup& ddRankSetup = dd->comm->ddRankSetup;
1672 const CartesianRankSetup& cartSetup = dd->comm->cartesianRankSetup;
1674 if (thisRankHasDuty(cr, DUTY_PP))
1676 /* Copy or make a new PP communicator */
1678 /* We (possibly) reordered the nodes in split_communicator,
1679 * so it is no longer required in make_pp_communicator.
1681 const bool useCartesianReorder = (ddSettings.useCartesianReorder && !cartSetup.bCartesianPP_PME);
1683 make_pp_communicator(mdlog, dd, cr, useCartesianReorder);
1687 receive_ddindex2simnodeid(dd, cr);
1690 if (!thisRankHasDuty(cr, DUTY_PME))
1692 /* Set up the commnuication to our PME node */
1693 dd->pme_nodeid = dd_simnode2pmenode(ddRankSetup, cartSetup, pmeRanks, cr, cr->sim_nodeid);
1694 dd->pme_receive_vir_ener = receive_vir_ener(dd, pmeRanks, cr);
1697 fprintf(debug, "My pme_nodeid %d receive ener %s\n", dd->pme_nodeid,
1698 gmx::boolToString(dd->pme_receive_vir_ener));
1703 dd->pme_nodeid = -1;
1706 /* We can not use DDMASTER(dd), because dd->masterrank is set later */
1709 dd->ma = std::make_unique<AtomDistribution>(dd->nc, numAtomsInSystem, numAtomsInSystem);
1713 static real* get_slb_frac(const gmx::MDLogger& mdlog, const char* dir, int nc, const char* size_string)
1715 real * slb_frac, tot;
1720 if (nc > 1 && size_string != nullptr)
1722 GMX_LOG(mdlog.info).appendTextFormatted("Using static load balancing for the %s direction", dir);
1725 for (i = 0; i < nc; i++)
1728 sscanf(size_string, "%20lf%n", &dbl, &n);
1732 "Incorrect or not enough DD cell size entries for direction %s: '%s'",
1740 std::string relativeCellSizes = "Relative cell sizes:";
1741 for (i = 0; i < nc; i++)
1744 relativeCellSizes += gmx::formatString(" %5.3f", slb_frac[i]);
1746 GMX_LOG(mdlog.info).appendText(relativeCellSizes);
1752 static int multi_body_bondeds_count(const gmx_mtop_t* mtop)
1755 gmx_mtop_ilistloop_t iloop = gmx_mtop_ilistloop_init(mtop);
1757 while (const InteractionLists* ilists = gmx_mtop_ilistloop_next(iloop, &nmol))
1759 for (auto& ilist : extractILists(*ilists, IF_BOND))
1761 if (NRAL(ilist.functionType) > 2)
1763 n += nmol * (ilist.iatoms.size() / ilistStride(ilist));
1771 static int dd_getenv(const gmx::MDLogger& mdlog, const char* env_var, int def)
1777 val = getenv(env_var);
1780 if (sscanf(val, "%20d", &nst) <= 0)
1784 GMX_LOG(mdlog.info).appendTextFormatted("Found env.var. %s = %s, using value %d", env_var, val, nst);
1790 static void check_dd_restrictions(const gmx_domdec_t* dd, const t_inputrec* ir, const gmx::MDLogger& mdlog)
1792 if (ir->ePBC == epbcSCREW && (dd->nc[XX] == 1 || dd->nc[YY] > 1 || dd->nc[ZZ] > 1))
1794 gmx_fatal(FARGS, "With pbc=%s can only do domain decomposition in the x-direction",
1795 epbc_names[ir->ePBC]);
1798 if (ir->nstlist == 0)
1800 gmx_fatal(FARGS, "Domain decomposition does not work with nstlist=0");
1803 if (ir->comm_mode == ecmANGULAR && ir->ePBC != epbcNONE)
1805 GMX_LOG(mdlog.warning)
1807 "comm-mode angular will give incorrect results when the comm group "
1808 "partially crosses a periodic boundary");
1812 static real average_cellsize_min(const gmx_ddbox_t& ddbox, const ivec numDomains)
1814 real r = ddbox.box_size[XX];
1815 for (int d = 0; d < DIM; d++)
1817 if (numDomains[d] > 1)
1819 /* Check using the initial average cell size */
1820 r = std::min(r, ddbox.box_size[d] * ddbox.skew_fac[d] / numDomains[d]);
1827 /*! \brief Depending on the DLB initial value return the DLB switched off state or issue an error.
1829 static DlbState forceDlbOffOrBail(DlbState cmdlineDlbState,
1830 const std::string& reasonStr,
1831 const gmx::MDLogger& mdlog)
1833 std::string dlbNotSupportedErr = "Dynamic load balancing requested, but ";
1834 std::string dlbDisableNote = "NOTE: disabling dynamic load balancing as ";
1836 if (cmdlineDlbState == DlbState::onUser)
1838 gmx_fatal(FARGS, "%s", (dlbNotSupportedErr + reasonStr).c_str());
1840 else if (cmdlineDlbState == DlbState::offCanTurnOn)
1842 GMX_LOG(mdlog.info).appendText(dlbDisableNote + reasonStr);
1844 return DlbState::offForever;
1847 /*! \brief Return the dynamic load balancer's initial state based on initial conditions and user inputs.
1849 * This function parses the parameters of "-dlb" command line option setting
1850 * corresponding state values. Then it checks the consistency of the determined
1851 * state with other run parameters and settings. As a result, the initial state
1852 * may be altered or an error may be thrown if incompatibility of options is detected.
1854 * \param [in] mdlog Logger.
1855 * \param [in] dlbOption Enum value for the DLB option.
1856 * \param [in] bRecordLoad True if the load balancer is recording load information.
1857 * \param [in] mdrunOptions Options for mdrun.
1858 * \param [in] ir Pointer mdrun to input parameters.
1859 * \returns DLB initial/startup state.
1861 static DlbState determineInitialDlbState(const gmx::MDLogger& mdlog,
1862 DlbOption dlbOption,
1863 gmx_bool bRecordLoad,
1864 const gmx::MdrunOptions& mdrunOptions,
1865 const t_inputrec* ir)
1867 DlbState dlbState = DlbState::offCanTurnOn;
1871 case DlbOption::turnOnWhenUseful: dlbState = DlbState::offCanTurnOn; break;
1872 case DlbOption::no: dlbState = DlbState::offUser; break;
1873 case DlbOption::yes: dlbState = DlbState::onUser; break;
1874 default: gmx_incons("Invalid dlbOption enum value");
1877 /* Reruns don't support DLB: bail or override auto mode */
1878 if (mdrunOptions.rerun)
1880 std::string reasonStr = "it is not supported in reruns.";
1881 return forceDlbOffOrBail(dlbState, reasonStr, mdlog);
1884 /* Unsupported integrators */
1885 if (!EI_DYNAMICS(ir->eI))
1887 auto reasonStr = gmx::formatString(
1888 "it is only supported with dynamics, not with integrator '%s'.", EI(ir->eI));
1889 return forceDlbOffOrBail(dlbState, reasonStr, mdlog);
1892 /* Without cycle counters we can't time work to balance on */
1895 std::string reasonStr =
1896 "cycle counters unsupported or not enabled in the operating system kernel.";
1897 return forceDlbOffOrBail(dlbState, reasonStr, mdlog);
1900 if (mdrunOptions.reproducible)
1902 std::string reasonStr = "you started a reproducible run.";
1905 case DlbState::offUser: break;
1906 case DlbState::offForever:
1907 GMX_RELEASE_ASSERT(false, "DlbState::offForever is not a valid initial state");
1909 case DlbState::offCanTurnOn: return forceDlbOffOrBail(dlbState, reasonStr, mdlog);
1910 case DlbState::onCanTurnOff:
1911 GMX_RELEASE_ASSERT(false, "DlbState::offCanTurnOff is not a valid initial state");
1913 case DlbState::onUser:
1914 return forceDlbOffOrBail(
1917 + " In load balanced runs binary reproducibility cannot be "
1921 gmx_fatal(FARGS, "Death horror: undefined case (%d) for load balancing choice",
1922 static_cast<int>(dlbState));
1929 static gmx_domdec_comm_t* init_dd_comm()
1931 gmx_domdec_comm_t* comm = new gmx_domdec_comm_t;
1933 comm->n_load_have = 0;
1934 comm->n_load_collect = 0;
1936 comm->haveTurnedOffDlb = false;
1938 for (int i = 0; i < static_cast<int>(DDAtomRanges::Type::Number); i++)
1940 comm->sum_nat[i] = 0;
1944 comm->load_step = 0;
1947 clear_ivec(comm->load_lim);
1951 /* This should be replaced by a unique pointer */
1952 comm->balanceRegion = ddBalanceRegionAllocate();
1957 /* Returns whether mtop contains constraints and/or vsites */
1958 static bool systemHasConstraintsOrVsites(const gmx_mtop_t& mtop)
1960 auto ilistLoop = gmx_mtop_ilistloop_init(mtop);
1962 while (const InteractionLists* ilists = gmx_mtop_ilistloop_next(ilistLoop, &nmol))
1964 if (!extractILists(*ilists, IF_CONSTRAINT | IF_VSITE).empty())
1973 static void setupUpdateGroups(const gmx::MDLogger& mdlog,
1974 const gmx_mtop_t& mtop,
1975 const t_inputrec& inputrec,
1976 const real cutoffMargin,
1977 DDSystemInfo* systemInfo)
1979 /* When we have constraints and/or vsites, it is beneficial to use
1980 * update groups (when possible) to allow independent update of groups.
1982 if (!systemHasConstraintsOrVsites(mtop))
1984 /* No constraints or vsites, atoms can be updated independently */
1988 systemInfo->updateGroupingPerMoleculetype = gmx::makeUpdateGroups(mtop);
1989 systemInfo->useUpdateGroups = (!systemInfo->updateGroupingPerMoleculetype.empty()
1990 && getenv("GMX_NO_UPDATEGROUPS") == nullptr);
1992 if (systemInfo->useUpdateGroups)
1994 int numUpdateGroups = 0;
1995 for (const auto& molblock : mtop.molblock)
1997 numUpdateGroups += molblock.nmol
1998 * systemInfo->updateGroupingPerMoleculetype[molblock.type].numBlocks();
2001 systemInfo->maxUpdateGroupRadius = computeMaxUpdateGroupRadius(
2002 mtop, systemInfo->updateGroupingPerMoleculetype, maxReferenceTemperature(inputrec));
2004 /* To use update groups, the large domain-to-domain cutoff distance
2005 * should be compatible with the box size.
2007 systemInfo->useUpdateGroups = (atomToAtomIntoDomainToDomainCutoff(*systemInfo, 0) < cutoffMargin);
2009 if (systemInfo->useUpdateGroups)
2012 .appendTextFormatted(
2013 "Using update groups, nr %d, average size %.1f atoms, max. radius %.3f "
2015 numUpdateGroups, mtop.natoms / static_cast<double>(numUpdateGroups),
2016 systemInfo->maxUpdateGroupRadius);
2021 .appendTextFormatted(
2022 "The combination of rlist and box size prohibits the use of update "
2024 systemInfo->updateGroupingPerMoleculetype.clear();
2029 UnitCellInfo::UnitCellInfo(const t_inputrec& ir) :
2030 npbcdim(ePBC2npbcdim(ir.ePBC)),
2031 numBoundedDimensions(inputrec2nboundeddim(&ir)),
2032 ddBoxIsDynamic(numBoundedDimensions < DIM || inputrecDynamicBox(&ir)),
2033 haveScrewPBC(ir.ePBC == epbcSCREW)
2037 /* Returns whether molecules are always whole, i.e. not broken by PBC */
2038 static bool moleculesAreAlwaysWhole(const gmx_mtop_t& mtop,
2039 const bool useUpdateGroups,
2040 gmx::ArrayRef<const gmx::RangePartitioning> updateGroupingPerMoleculetype)
2042 if (useUpdateGroups)
2044 GMX_RELEASE_ASSERT(updateGroupingPerMoleculetype.size() == mtop.moltype.size(),
2045 "Need one grouping per moltype");
2046 for (size_t mol = 0; mol < mtop.moltype.size(); mol++)
2048 if (updateGroupingPerMoleculetype[mol].numBlocks() > 1)
2056 for (const auto& moltype : mtop.moltype)
2058 if (moltype.atoms.nr > 1)
2068 /*! \brief Generate the simulation system information */
2069 static DDSystemInfo getSystemInfo(const gmx::MDLogger& mdlog,
2070 const t_commrec* cr,
2071 const DomdecOptions& options,
2072 const gmx_mtop_t& mtop,
2073 const t_inputrec& ir,
2075 gmx::ArrayRef<const gmx::RVec> xGlobal)
2077 const real tenPercentMargin = 1.1;
2079 DDSystemInfo systemInfo;
2081 /* We need to decide on update groups early, as this affects communication distances */
2082 systemInfo.useUpdateGroups = false;
2083 if (ir.cutoff_scheme == ecutsVERLET)
2085 real cutoffMargin = std::sqrt(max_cutoff2(ir.ePBC, box)) - ir.rlist;
2086 setupUpdateGroups(mdlog, mtop, ir, cutoffMargin, &systemInfo);
2089 systemInfo.moleculesAreAlwaysWhole = moleculesAreAlwaysWhole(
2090 mtop, systemInfo.useUpdateGroups, systemInfo.updateGroupingPerMoleculetype);
2091 systemInfo.haveInterDomainBondeds =
2092 (!systemInfo.moleculesAreAlwaysWhole || mtop.bIntermolecularInteractions);
2093 systemInfo.haveInterDomainMultiBodyBondeds =
2094 (systemInfo.haveInterDomainBondeds && multi_body_bondeds_count(&mtop) > 0);
2096 if (systemInfo.useUpdateGroups)
2098 systemInfo.haveSplitConstraints = false;
2099 systemInfo.haveSplitSettles = false;
2103 systemInfo.haveSplitConstraints = (gmx_mtop_ftype_count(mtop, F_CONSTR) > 0
2104 || gmx_mtop_ftype_count(mtop, F_CONSTRNC) > 0);
2105 systemInfo.haveSplitSettles = (gmx_mtop_ftype_count(mtop, F_SETTLE) > 0);
2110 /* Set the cut-off to some very large value,
2111 * so we don't need if statements everywhere in the code.
2112 * We use sqrt, since the cut-off is squared in some places.
2114 systemInfo.cutoff = GMX_CUTOFF_INF;
2118 systemInfo.cutoff = atomToAtomIntoDomainToDomainCutoff(systemInfo, ir.rlist);
2120 systemInfo.minCutoffForMultiBody = 0;
2122 /* Determine the minimum cell size limit, affected by many factors */
2123 systemInfo.cellsizeLimit = 0;
2124 systemInfo.filterBondedCommunication = false;
2126 /* We do not allow home atoms to move beyond the neighboring domain
2127 * between domain decomposition steps, which limits the cell size.
2128 * Get an estimate of cell size limit due to atom displacement.
2129 * In most cases this is a large overestimate, because it assumes
2130 * non-interaction atoms.
2131 * We set the chance to 1 in a trillion steps.
2133 constexpr real c_chanceThatAtomMovesBeyondDomain = 1e-12;
2134 const real limitForAtomDisplacement = minCellSizeForAtomDisplacement(
2135 mtop, ir, systemInfo.updateGroupingPerMoleculetype, c_chanceThatAtomMovesBeyondDomain);
2136 GMX_LOG(mdlog.info).appendTextFormatted("Minimum cell size due to atom displacement: %.3f nm", limitForAtomDisplacement);
2138 systemInfo.cellsizeLimit = std::max(systemInfo.cellsizeLimit, limitForAtomDisplacement);
2140 /* TODO: PME decomposition currently requires atoms not to be more than
2141 * 2/3 of comm->cutoff, which is >=rlist, outside of their domain.
2142 * In nearly all cases, limitForAtomDisplacement will be smaller
2143 * than 2/3*rlist, so the PME requirement is satisfied.
2144 * But it would be better for both correctness and performance
2145 * to use limitForAtomDisplacement instead of 2/3*comm->cutoff.
2146 * Note that we would need to improve the pairlist buffer case.
2149 if (systemInfo.haveInterDomainBondeds)
2151 if (options.minimumCommunicationRange > 0)
2153 systemInfo.minCutoffForMultiBody =
2154 atomToAtomIntoDomainToDomainCutoff(systemInfo, options.minimumCommunicationRange);
2155 if (options.useBondedCommunication)
2157 systemInfo.filterBondedCommunication =
2158 (systemInfo.minCutoffForMultiBody > systemInfo.cutoff);
2162 systemInfo.cutoff = std::max(systemInfo.cutoff, systemInfo.minCutoffForMultiBody);
2165 else if (ir.bPeriodicMols)
2167 /* Can not easily determine the required cut-off */
2168 GMX_LOG(mdlog.warning)
2170 "NOTE: Periodic molecules are present in this system. Because of this, "
2171 "the domain decomposition algorithm cannot easily determine the "
2172 "minimum cell size that it requires for treating bonded interactions. "
2173 "Instead, domain decomposition will assume that half the non-bonded "
2174 "cut-off will be a suitable lower bound.");
2175 systemInfo.minCutoffForMultiBody = systemInfo.cutoff / 2;
2183 dd_bonded_cg_distance(mdlog, &mtop, &ir, as_rvec_array(xGlobal.data()), box,
2184 options.checkBondedInteractions, &r_2b, &r_mb);
2186 gmx_bcast(sizeof(r_2b), &r_2b, cr);
2187 gmx_bcast(sizeof(r_mb), &r_mb, cr);
2189 /* We use an initial margin of 10% for the minimum cell size,
2190 * except when we are just below the non-bonded cut-off.
2192 if (options.useBondedCommunication)
2194 if (std::max(r_2b, r_mb) > systemInfo.cutoff)
2196 const real r_bonded = std::max(r_2b, r_mb);
2197 systemInfo.minCutoffForMultiBody = tenPercentMargin * r_bonded;
2198 /* This is the (only) place where we turn on the filtering */
2199 systemInfo.filterBondedCommunication = true;
2203 const real r_bonded = r_mb;
2204 systemInfo.minCutoffForMultiBody =
2205 std::min(tenPercentMargin * r_bonded, systemInfo.cutoff);
2207 /* We determine cutoff_mbody later */
2208 systemInfo.increaseMultiBodyCutoff = true;
2212 /* No special bonded communication,
2213 * simply increase the DD cut-off.
2215 systemInfo.minCutoffForMultiBody = tenPercentMargin * std::max(r_2b, r_mb);
2216 systemInfo.cutoff = std::max(systemInfo.cutoff, systemInfo.minCutoffForMultiBody);
2220 .appendTextFormatted("Minimum cell size due to bonded interactions: %.3f nm",
2221 systemInfo.minCutoffForMultiBody);
2223 systemInfo.cellsizeLimit = std::max(systemInfo.cellsizeLimit, systemInfo.minCutoffForMultiBody);
2226 systemInfo.constraintCommunicationRange = 0;
2227 if (systemInfo.haveSplitConstraints && options.constraintCommunicationRange <= 0)
2229 /* There is a cell size limit due to the constraints (P-LINCS) */
2230 systemInfo.constraintCommunicationRange = gmx::constr_r_max(mdlog, &mtop, &ir);
2232 .appendTextFormatted("Estimated maximum distance required for P-LINCS: %.3f nm",
2233 systemInfo.constraintCommunicationRange);
2234 if (systemInfo.constraintCommunicationRange > systemInfo.cellsizeLimit)
2238 "This distance will limit the DD cell size, you can override this with "
2242 else if (options.constraintCommunicationRange > 0)
2244 /* Here we do not check for dd->splitConstraints.
2245 * because one can also set a cell size limit for virtual sites only
2246 * and at this point we don't know yet if there are intercg v-sites.
2249 .appendTextFormatted("User supplied maximum distance required for P-LINCS: %.3f nm",
2250 options.constraintCommunicationRange);
2251 systemInfo.constraintCommunicationRange = options.constraintCommunicationRange;
2253 systemInfo.cellsizeLimit = std::max(systemInfo.cellsizeLimit, systemInfo.constraintCommunicationRange);
2258 /*! \brief Exit with a fatal error if the DDGridSetup cannot be
2260 static void checkDDGridSetup(const DDGridSetup& ddGridSetup,
2261 const t_commrec* cr,
2262 const DomdecOptions& options,
2263 const DDSettings& ddSettings,
2264 const DDSystemInfo& systemInfo,
2265 const real cellsizeLimit,
2266 const gmx_ddbox_t& ddbox)
2268 if (options.numCells[XX] <= 0 && (ddGridSetup.numDomains[XX] == 0))
2271 gmx_bool bC = (systemInfo.haveSplitConstraints
2272 && systemInfo.constraintCommunicationRange > systemInfo.minCutoffForMultiBody);
2273 sprintf(buf, "Change the number of ranks or mdrun option %s%s%s", !bC ? "-rdd" : "-rcon",
2274 ddSettings.initialDlbState != DlbState::offUser ? " or -dds" : "",
2275 bC ? " or your LINCS settings" : "");
2277 gmx_fatal_collective(FARGS, cr->mpi_comm_mysim, MASTER(cr),
2278 "There is no domain decomposition for %d ranks that is compatible "
2279 "with the given box and a minimum cell size of %g nm\n"
2281 "Look in the log file for details on the domain decomposition",
2282 cr->nnodes - ddGridSetup.numPmeOnlyRanks, cellsizeLimit, buf);
2285 const real acs = average_cellsize_min(ddbox, ddGridSetup.numDomains);
2286 if (acs < cellsizeLimit)
2288 if (options.numCells[XX] <= 0)
2292 "dd_choose_grid() should return a grid that satisfies the cell size limits");
2296 gmx_fatal_collective(
2297 FARGS, cr->mpi_comm_mysim, MASTER(cr),
2298 "The initial cell size (%f) is smaller than the cell size limit (%f), change "
2299 "options -dd, -rdd or -rcon, see the log file for details",
2300 acs, cellsizeLimit);
2304 const int numPPRanks =
2305 ddGridSetup.numDomains[XX] * ddGridSetup.numDomains[YY] * ddGridSetup.numDomains[ZZ];
2306 if (cr->nnodes - numPPRanks != ddGridSetup.numPmeOnlyRanks)
2308 gmx_fatal_collective(FARGS, cr->mpi_comm_mysim, MASTER(cr),
2309 "The size of the domain decomposition grid (%d) does not match the "
2310 "number of PP ranks (%d). The total number of ranks is %d",
2311 numPPRanks, cr->nnodes - ddGridSetup.numPmeOnlyRanks, cr->nnodes);
2313 if (ddGridSetup.numPmeOnlyRanks > numPPRanks)
2315 gmx_fatal_collective(FARGS, cr->mpi_comm_mysim, MASTER(cr),
2316 "The number of separate PME ranks (%d) is larger than the number of "
2317 "PP ranks (%d), this is not supported.",
2318 ddGridSetup.numPmeOnlyRanks, numPPRanks);
2322 /*! \brief Set the cell size and interaction limits, as well as the DD grid */
2323 static DDRankSetup getDDRankSetup(const gmx::MDLogger& mdlog,
2325 const DDGridSetup& ddGridSetup,
2326 const t_inputrec& ir)
2329 .appendTextFormatted("Domain decomposition grid %d x %d x %d, separate PME ranks %d",
2330 ddGridSetup.numDomains[XX], ddGridSetup.numDomains[YY],
2331 ddGridSetup.numDomains[ZZ], ddGridSetup.numPmeOnlyRanks);
2333 DDRankSetup ddRankSetup;
2335 ddRankSetup.numPPRanks = cr->nnodes - ddGridSetup.numPmeOnlyRanks;
2336 copy_ivec(ddGridSetup.numDomains, ddRankSetup.numPPCells);
2338 ddRankSetup.usePmeOnlyRanks = (ddGridSetup.numPmeOnlyRanks > 0);
2339 if (ddRankSetup.usePmeOnlyRanks)
2341 ddRankSetup.numRanksDoingPme = ddGridSetup.numPmeOnlyRanks;
2345 ddRankSetup.numRanksDoingPme =
2346 ddGridSetup.numDomains[XX] * ddGridSetup.numDomains[YY] * ddGridSetup.numDomains[ZZ];
2349 if (EEL_PME(ir.coulombtype) || EVDW_PME(ir.vdwtype))
2351 /* The following choices should match those
2352 * in comm_cost_est in domdec_setup.c.
2353 * Note that here the checks have to take into account
2354 * that the decomposition might occur in a different order than xyz
2355 * (for instance through the env.var. GMX_DD_ORDER_ZYX),
2356 * in which case they will not match those in comm_cost_est,
2357 * but since that is mainly for testing purposes that's fine.
2359 if (ddGridSetup.numDDDimensions >= 2 && ddGridSetup.ddDimensions[0] == XX
2360 && ddGridSetup.ddDimensions[1] == YY
2361 && ddRankSetup.numRanksDoingPme > ddGridSetup.numDomains[XX]
2362 && ddRankSetup.numRanksDoingPme % ddGridSetup.numDomains[XX] == 0
2363 && getenv("GMX_PMEONEDD") == nullptr)
2365 ddRankSetup.npmedecompdim = 2;
2366 ddRankSetup.npmenodes_x = ddGridSetup.numDomains[XX];
2367 ddRankSetup.npmenodes_y = ddRankSetup.numRanksDoingPme / ddRankSetup.npmenodes_x;
2371 /* In case nc is 1 in both x and y we could still choose to
2372 * decompose pme in y instead of x, but we use x for simplicity.
2374 ddRankSetup.npmedecompdim = 1;
2375 if (ddGridSetup.ddDimensions[0] == YY)
2377 ddRankSetup.npmenodes_x = 1;
2378 ddRankSetup.npmenodes_y = ddRankSetup.numRanksDoingPme;
2382 ddRankSetup.npmenodes_x = ddRankSetup.numRanksDoingPme;
2383 ddRankSetup.npmenodes_y = 1;
2387 .appendTextFormatted("PME domain decomposition: %d x %d x %d",
2388 ddRankSetup.npmenodes_x, ddRankSetup.npmenodes_y, 1);
2392 ddRankSetup.npmedecompdim = 0;
2393 ddRankSetup.npmenodes_x = 0;
2394 ddRankSetup.npmenodes_y = 0;
2400 /*! \brief Set the cell size and interaction limits */
2401 static void set_dd_limits(const gmx::MDLogger& mdlog,
2404 const DomdecOptions& options,
2405 const DDSettings& ddSettings,
2406 const DDSystemInfo& systemInfo,
2407 const DDGridSetup& ddGridSetup,
2408 const int numPPRanks,
2409 const gmx_mtop_t* mtop,
2410 const t_inputrec* ir,
2411 const gmx_ddbox_t& ddbox)
2413 gmx_domdec_comm_t* comm = dd->comm;
2414 comm->ddSettings = ddSettings;
2416 /* Initialize to GPU share count to 0, might change later */
2417 comm->nrank_gpu_shared = 0;
2419 comm->dlbState = comm->ddSettings.initialDlbState;
2420 dd_dlb_set_should_check_whether_to_turn_dlb_on(dd, TRUE);
2421 /* To consider turning DLB on after 2*nstlist steps we need to check
2422 * at partitioning count 3. Thus we need to increase the first count by 2.
2424 comm->ddPartioningCountFirstDlbOff += 2;
2426 comm->bPMELoadBalDLBLimits = FALSE;
2428 /* Allocate the charge group/atom sorting struct */
2429 comm->sort = std::make_unique<gmx_domdec_sort_t>();
2431 comm->systemInfo = systemInfo;
2433 if (systemInfo.useUpdateGroups)
2435 /* Note: We would like to use dd->nnodes for the atom count estimate,
2436 * but that is not yet available here. But this anyhow only
2437 * affect performance up to the second dd_partition_system call.
2439 const int homeAtomCountEstimate = mtop->natoms / numPPRanks;
2440 comm->updateGroupsCog = std::make_unique<gmx::UpdateGroupsCog>(
2441 *mtop, systemInfo.updateGroupingPerMoleculetype, maxReferenceTemperature(*ir),
2442 homeAtomCountEstimate);
2445 /* Set the DD setup given by ddGridSetup */
2446 copy_ivec(ddGridSetup.numDomains, dd->nc);
2447 dd->ndim = ddGridSetup.numDDDimensions;
2448 copy_ivec(ddGridSetup.ddDimensions, dd->dim);
2450 dd->nnodes = dd->nc[XX] * dd->nc[YY] * dd->nc[ZZ];
2452 snew(comm->slb_frac, DIM);
2453 if (isDlbDisabled(comm))
2455 comm->slb_frac[XX] = get_slb_frac(mdlog, "x", dd->nc[XX], options.cellSizeX);
2456 comm->slb_frac[YY] = get_slb_frac(mdlog, "y", dd->nc[YY], options.cellSizeY);
2457 comm->slb_frac[ZZ] = get_slb_frac(mdlog, "z", dd->nc[ZZ], options.cellSizeZ);
2460 /* Set the multi-body cut-off and cellsize limit for DLB */
2461 comm->cutoff_mbody = systemInfo.minCutoffForMultiBody;
2462 comm->cellsize_limit = systemInfo.cellsizeLimit;
2463 if (systemInfo.haveInterDomainBondeds && systemInfo.increaseMultiBodyCutoff)
2465 if (systemInfo.filterBondedCommunication || !isDlbDisabled(comm))
2467 /* Set the bonded communication distance to halfway
2468 * the minimum and the maximum,
2469 * since the extra communication cost is nearly zero.
2471 real acs = average_cellsize_min(ddbox, dd->nc);
2472 comm->cutoff_mbody = 0.5 * (systemInfo.minCutoffForMultiBody + acs);
2473 if (!isDlbDisabled(comm))
2475 /* Check if this does not limit the scaling */
2476 comm->cutoff_mbody = std::min(comm->cutoff_mbody, options.dlbScaling * acs);
2478 if (!systemInfo.filterBondedCommunication)
2480 /* Without bBondComm do not go beyond the n.b. cut-off */
2481 comm->cutoff_mbody = std::min(comm->cutoff_mbody, systemInfo.cutoff);
2482 if (comm->cellsize_limit >= systemInfo.cutoff)
2484 /* We don't loose a lot of efficieny
2485 * when increasing it to the n.b. cut-off.
2486 * It can even be slightly faster, because we need
2487 * less checks for the communication setup.
2489 comm->cutoff_mbody = systemInfo.cutoff;
2492 /* Check if we did not end up below our original limit */
2493 comm->cutoff_mbody = std::max(comm->cutoff_mbody, systemInfo.minCutoffForMultiBody);
2495 if (comm->cutoff_mbody > comm->cellsize_limit)
2497 comm->cellsize_limit = comm->cutoff_mbody;
2500 /* Without DLB and cutoff_mbody<cutoff, cutoff_mbody is dynamic */
2506 "Bonded atom communication beyond the cut-off: %s\n"
2507 "cellsize limit %f\n",
2508 gmx::boolToString(systemInfo.filterBondedCommunication), comm->cellsize_limit);
2513 check_dd_restrictions(dd, ir, mdlog);
2517 void dd_init_bondeds(FILE* fplog,
2519 const gmx_mtop_t* mtop,
2520 const gmx_vsite_t* vsite,
2521 const t_inputrec* ir,
2523 cginfo_mb_t* cginfo_mb)
2525 gmx_domdec_comm_t* comm;
2527 dd_make_reverse_top(fplog, dd, mtop, vsite, ir, bBCheck);
2531 if (comm->systemInfo.filterBondedCommunication)
2533 /* Communicate atoms beyond the cut-off for bonded interactions */
2534 comm->bondedLinks = makeBondedLinks(mtop, cginfo_mb);
2538 /* Only communicate atoms based on cut-off */
2539 comm->bondedLinks = nullptr;
2543 static void writeSettings(gmx::TextWriter* log,
2545 const gmx_mtop_t* mtop,
2546 const t_inputrec* ir,
2547 gmx_bool bDynLoadBal,
2549 const gmx_ddbox_t* ddbox)
2551 gmx_domdec_comm_t* comm;
2560 log->writeString("The maximum number of communication pulses is:");
2561 for (d = 0; d < dd->ndim; d++)
2563 log->writeStringFormatted(" %c %d", dim2char(dd->dim[d]), comm->cd[d].np_dlb);
2565 log->ensureLineBreak();
2566 log->writeLineFormatted("The minimum size for domain decomposition cells is %.3f nm",
2567 comm->cellsize_limit);
2568 log->writeLineFormatted("The requested allowed shrink of DD cells (option -dds) is: %.2f", dlb_scale);
2569 log->writeString("The allowed shrink of domain decomposition cells is:");
2570 for (d = 0; d < DIM; d++)
2574 if (d >= ddbox->npbcdim && dd->nc[d] == 2)
2580 shrink = comm->cellsize_min_dlb[d]
2581 / (ddbox->box_size[d] * ddbox->skew_fac[d] / dd->nc[d]);
2583 log->writeStringFormatted(" %c %.2f", dim2char(d), shrink);
2586 log->ensureLineBreak();
2590 set_dd_cell_sizes_slb(dd, ddbox, setcellsizeslbPULSE_ONLY, np);
2591 log->writeString("The initial number of communication pulses is:");
2592 for (d = 0; d < dd->ndim; d++)
2594 log->writeStringFormatted(" %c %d", dim2char(dd->dim[d]), np[dd->dim[d]]);
2596 log->ensureLineBreak();
2597 log->writeString("The initial domain decomposition cell size is:");
2598 for (d = 0; d < DIM; d++)
2602 log->writeStringFormatted(" %c %.2f nm", dim2char(d), dd->comm->cellsize_min[d]);
2605 log->ensureLineBreak();
2609 const bool haveInterDomainVsites =
2610 (countInterUpdategroupVsites(*mtop, comm->systemInfo.updateGroupingPerMoleculetype) != 0);
2612 if (comm->systemInfo.haveInterDomainBondeds || haveInterDomainVsites
2613 || comm->systemInfo.haveSplitConstraints || comm->systemInfo.haveSplitSettles)
2615 std::string decompUnits;
2616 if (comm->systemInfo.useUpdateGroups)
2618 decompUnits = "atom groups";
2622 decompUnits = "atoms";
2625 log->writeLineFormatted("The maximum allowed distance for %s involved in interactions is:",
2626 decompUnits.c_str());
2627 log->writeLineFormatted("%40s %-7s %6.3f nm", "non-bonded interactions", "",
2628 comm->systemInfo.cutoff);
2632 limit = dd->comm->cellsize_limit;
2636 if (dd->unitCellInfo.ddBoxIsDynamic)
2639 "(the following are initial values, they could change due to box "
2642 limit = dd->comm->cellsize_min[XX];
2643 for (d = 1; d < DIM; d++)
2645 limit = std::min(limit, dd->comm->cellsize_min[d]);
2649 if (comm->systemInfo.haveInterDomainBondeds)
2651 log->writeLineFormatted("%40s %-7s %6.3f nm", "two-body bonded interactions", "(-rdd)",
2652 std::max(comm->systemInfo.cutoff, comm->cutoff_mbody));
2653 log->writeLineFormatted("%40s %-7s %6.3f nm", "multi-body bonded interactions",
2655 (comm->systemInfo.filterBondedCommunication || isDlbOn(dd->comm))
2656 ? comm->cutoff_mbody
2657 : std::min(comm->systemInfo.cutoff, limit));
2659 if (haveInterDomainVsites)
2661 log->writeLineFormatted("%40s %-7s %6.3f nm", "virtual site constructions", "(-rcon)", limit);
2663 if (comm->systemInfo.haveSplitConstraints || comm->systemInfo.haveSplitSettles)
2665 std::string separation =
2666 gmx::formatString("atoms separated by up to %d constraints", 1 + ir->nProjOrder);
2667 log->writeLineFormatted("%40s %-7s %6.3f nm\n", separation.c_str(), "(-rcon)", limit);
2669 log->ensureLineBreak();
2673 static void logSettings(const gmx::MDLogger& mdlog,
2675 const gmx_mtop_t* mtop,
2676 const t_inputrec* ir,
2678 const gmx_ddbox_t* ddbox)
2680 gmx::StringOutputStream stream;
2681 gmx::TextWriter log(&stream);
2682 writeSettings(&log, dd, mtop, ir, isDlbOn(dd->comm), dlb_scale, ddbox);
2683 if (dd->comm->dlbState == DlbState::offCanTurnOn)
2686 log.ensureEmptyLine();
2688 "When dynamic load balancing gets turned on, these settings will change to:");
2690 writeSettings(&log, dd, mtop, ir, true, dlb_scale, ddbox);
2692 GMX_LOG(mdlog.info).asParagraph().appendText(stream.toString());
2695 static void set_cell_limits_dlb(const gmx::MDLogger& mdlog,
2698 const t_inputrec* ir,
2699 const gmx_ddbox_t* ddbox)
2701 gmx_domdec_comm_t* comm;
2702 int d, dim, npulse, npulse_d_max, npulse_d;
2707 bNoCutOff = (ir->rvdw == 0 || ir->rcoulomb == 0);
2709 /* Determine the maximum number of comm. pulses in one dimension */
2711 comm->cellsize_limit = std::max(comm->cellsize_limit, comm->cutoff_mbody);
2713 /* Determine the maximum required number of grid pulses */
2714 if (comm->cellsize_limit >= comm->systemInfo.cutoff)
2716 /* Only a single pulse is required */
2719 else if (!bNoCutOff && comm->cellsize_limit > 0)
2721 /* We round down slightly here to avoid overhead due to the latency
2722 * of extra communication calls when the cut-off
2723 * would be only slightly longer than the cell size.
2724 * Later cellsize_limit is redetermined,
2725 * so we can not miss interactions due to this rounding.
2727 npulse = static_cast<int>(0.96 + comm->systemInfo.cutoff / comm->cellsize_limit);
2731 /* There is no cell size limit */
2732 npulse = std::max(dd->nc[XX] - 1, std::max(dd->nc[YY] - 1, dd->nc[ZZ] - 1));
2735 if (!bNoCutOff && npulse > 1)
2737 /* See if we can do with less pulses, based on dlb_scale */
2739 for (d = 0; d < dd->ndim; d++)
2742 npulse_d = static_cast<int>(
2744 + dd->nc[dim] * comm->systemInfo.cutoff
2745 / (ddbox->box_size[dim] * ddbox->skew_fac[dim] * dlb_scale));
2746 npulse_d_max = std::max(npulse_d_max, npulse_d);
2748 npulse = std::min(npulse, npulse_d_max);
2751 /* This env var can override npulse */
2752 d = dd_getenv(mdlog, "GMX_DD_NPULSE", 0);
2759 comm->bVacDLBNoLimit = (ir->ePBC == epbcNONE);
2760 for (d = 0; d < dd->ndim; d++)
2762 if (comm->ddSettings.request1DAnd1Pulse)
2764 comm->cd[d].np_dlb = 1;
2768 comm->cd[d].np_dlb = std::min(npulse, dd->nc[dd->dim[d]] - 1);
2769 comm->maxpulse = std::max(comm->maxpulse, comm->cd[d].np_dlb);
2771 if (comm->cd[d].np_dlb < dd->nc[dd->dim[d]] - 1)
2773 comm->bVacDLBNoLimit = FALSE;
2777 /* cellsize_limit is set for LINCS in init_domain_decomposition */
2778 if (!comm->bVacDLBNoLimit)
2780 comm->cellsize_limit = std::max(comm->cellsize_limit, comm->systemInfo.cutoff / comm->maxpulse);
2782 comm->cellsize_limit = std::max(comm->cellsize_limit, comm->cutoff_mbody);
2783 /* Set the minimum cell size for each DD dimension */
2784 for (d = 0; d < dd->ndim; d++)
2786 if (comm->bVacDLBNoLimit || comm->cd[d].np_dlb * comm->cellsize_limit >= comm->systemInfo.cutoff)
2788 comm->cellsize_min_dlb[dd->dim[d]] = comm->cellsize_limit;
2792 comm->cellsize_min_dlb[dd->dim[d]] = comm->systemInfo.cutoff / comm->cd[d].np_dlb;
2795 if (comm->cutoff_mbody <= 0)
2797 comm->cutoff_mbody = std::min(comm->systemInfo.cutoff, comm->cellsize_limit);
2805 bool dd_moleculesAreAlwaysWhole(const gmx_domdec_t& dd)
2807 return dd.comm->systemInfo.moleculesAreAlwaysWhole;
2810 gmx_bool dd_bonded_molpbc(const gmx_domdec_t* dd, int ePBC)
2812 /* If each molecule is a single charge group
2813 * or we use domain decomposition for each periodic dimension,
2814 * we do not need to take pbc into account for the bonded interactions.
2816 return (ePBC != epbcNONE && dd->comm->systemInfo.haveInterDomainBondeds
2817 && !(dd->nc[XX] > 1 && dd->nc[YY] > 1 && (dd->nc[ZZ] > 1 || ePBC == epbcXY)));
2820 /*! \brief Sets grid size limits and PP-PME setup, prints settings to log */
2821 static void set_ddgrid_parameters(const gmx::MDLogger& mdlog,
2824 const gmx_mtop_t* mtop,
2825 const t_inputrec* ir,
2826 const gmx_ddbox_t* ddbox)
2828 gmx_domdec_comm_t* comm = dd->comm;
2829 DDRankSetup& ddRankSetup = comm->ddRankSetup;
2831 if (EEL_PME(ir->coulombtype) || EVDW_PME(ir->vdwtype))
2833 init_ddpme(dd, &ddRankSetup.ddpme[0], 0);
2834 if (ddRankSetup.npmedecompdim >= 2)
2836 init_ddpme(dd, &ddRankSetup.ddpme[1], 1);
2841 ddRankSetup.numRanksDoingPme = 0;
2842 if (dd->pme_nodeid >= 0)
2844 gmx_fatal_collective(FARGS, dd->mpi_comm_all, DDMASTER(dd),
2845 "Can not have separate PME ranks without PME electrostatics");
2851 fprintf(debug, "The DD cut-off is %f\n", comm->systemInfo.cutoff);
2853 if (!isDlbDisabled(comm))
2855 set_cell_limits_dlb(mdlog, dd, dlb_scale, ir, ddbox);
2858 logSettings(mdlog, dd, mtop, ir, dlb_scale, ddbox);
2861 if (ir->ePBC == epbcNONE)
2863 vol_frac = 1 - 1 / static_cast<double>(dd->nnodes);
2867 vol_frac = (1 + comm_box_frac(dd->nc, comm->systemInfo.cutoff, *ddbox))
2868 / static_cast<double>(dd->nnodes);
2872 fprintf(debug, "Volume fraction for all DD zones: %f\n", vol_frac);
2874 int natoms_tot = mtop->natoms;
2876 dd->ga2la = new gmx_ga2la_t(natoms_tot, static_cast<int>(vol_frac * natoms_tot));
2879 /*! \brief Get some important DD parameters which can be modified by env.vars */
2880 static DDSettings getDDSettings(const gmx::MDLogger& mdlog,
2881 const DomdecOptions& options,
2882 const gmx::MdrunOptions& mdrunOptions,
2883 const t_inputrec& ir)
2885 DDSettings ddSettings;
2887 ddSettings.useSendRecv2 = (dd_getenv(mdlog, "GMX_DD_USE_SENDRECV2", 0) != 0);
2888 ddSettings.dlb_scale_lim = dd_getenv(mdlog, "GMX_DLB_MAX_BOX_SCALING", 10);
2889 ddSettings.request1DAnd1Pulse = bool(dd_getenv(mdlog, "GMX_DD_1D_1PULSE", 0));
2890 ddSettings.useDDOrderZYX = bool(dd_getenv(mdlog, "GMX_DD_ORDER_ZYX", 0));
2891 ddSettings.useCartesianReorder = bool(dd_getenv(mdlog, "GMX_NO_CART_REORDER", 1));
2892 ddSettings.eFlop = dd_getenv(mdlog, "GMX_DLB_BASED_ON_FLOPS", 0);
2893 const int recload = dd_getenv(mdlog, "GMX_DD_RECORD_LOAD", 1);
2894 ddSettings.nstDDDump = dd_getenv(mdlog, "GMX_DD_NST_DUMP", 0);
2895 ddSettings.nstDDDumpGrid = dd_getenv(mdlog, "GMX_DD_NST_DUMP_GRID", 0);
2896 ddSettings.DD_debug = dd_getenv(mdlog, "GMX_DD_DEBUG", 0);
2898 if (ddSettings.useSendRecv2)
2902 "Will use two sequential MPI_Sendrecv calls instead of two simultaneous "
2903 "non-blocking MPI_Irecv and MPI_Isend pairs for constraint and vsite "
2907 if (ddSettings.eFlop)
2909 GMX_LOG(mdlog.info).appendText("Will load balance based on FLOP count");
2910 ddSettings.recordLoad = true;
2914 ddSettings.recordLoad = (wallcycle_have_counter() && recload > 0);
2917 ddSettings.initialDlbState = determineInitialDlbState(mdlog, options.dlbOption,
2918 ddSettings.recordLoad, mdrunOptions, &ir);
2920 .appendTextFormatted("Dynamic load balancing: %s",
2921 edlbs_names[static_cast<int>(ddSettings.initialDlbState)]);
2926 gmx_domdec_t::gmx_domdec_t(const t_inputrec& ir) : unitCellInfo(ir) {}
2928 /*! \brief Return whether the simulation described can run a 1D single-pulse DD.
2930 * The GPU halo exchange code requires a 1D single-pulse DD. Such a DD
2931 * generally requires a larger box than other possible decompositions
2932 * with the same rank count, so the calling code might need to decide
2933 * what is the most appropriate way to run the simulation based on
2934 * whether such a DD is possible.
2936 * This function works like init_domain_decomposition(), but will not
2937 * give a fatal error, and only uses \c cr for communicating between
2940 * It is safe to call before thread-MPI spawns ranks, so that
2941 * thread-MPI can decide whether and how to trigger the GPU halo
2942 * exchange code path. The number of PME ranks, if any, should be set
2943 * in \c options.numPmeRanks.
2945 static bool canMake1DAnd1PulseDomainDecomposition(const DDSettings& ddSettingsOriginal,
2946 const t_commrec* cr,
2947 const int numRanksRequested,
2948 const DomdecOptions& options,
2949 const gmx_mtop_t& mtop,
2950 const t_inputrec& ir,
2952 gmx::ArrayRef<const gmx::RVec> xGlobal)
2954 // Ensure we don't write any output from this checking routine
2955 gmx::MDLogger dummyLogger;
2957 DDSystemInfo systemInfo = getSystemInfo(dummyLogger, cr, options, mtop, ir, box, xGlobal);
2959 DDSettings ddSettings = ddSettingsOriginal;
2960 ddSettings.request1DAnd1Pulse = true;
2961 const real gridSetupCellsizeLimit = getDDGridSetupCellSizeLimit(
2962 dummyLogger, ddSettings.request1DAnd1Pulse, !isDlbDisabled(ddSettings.initialDlbState),
2963 options.dlbScaling, ir, systemInfo.cellsizeLimit);
2964 gmx_ddbox_t ddbox = { 0 };
2965 DDGridSetup ddGridSetup =
2966 getDDGridSetup(dummyLogger, cr, numRanksRequested, options, ddSettings, systemInfo,
2967 gridSetupCellsizeLimit, mtop, ir, box, xGlobal, &ddbox);
2969 const bool canMakeDDWith1DAnd1Pulse = (ddGridSetup.numDomains[XX] != 0);
2971 return canMakeDDWith1DAnd1Pulse;
2974 bool is1DAnd1PulseDD(const gmx_domdec_t& dd)
2976 const int maxDimensionSize = std::max(dd.nc[XX], std::max(dd.nc[YY], dd.nc[ZZ]));
2977 const int productOfDimensionSizes = dd.nc[XX] * dd.nc[YY] * dd.nc[ZZ];
2978 const bool decompositionHasOneDimension = (maxDimensionSize == productOfDimensionSizes);
2980 const bool hasMax1Pulse =
2981 ((isDlbDisabled(dd.comm) && dd.comm->cellsize_limit >= dd.comm->systemInfo.cutoff)
2982 || (!isDlbDisabled(dd.comm) && dd.comm->maxpulse == 1));
2984 return decompositionHasOneDimension && hasMax1Pulse;
2990 // TODO once the functionality stablizes, move this class and
2991 // supporting functionality into builder.cpp
2992 /*! \brief Impl class for DD builder */
2993 class DomainDecompositionBuilder::Impl
2997 Impl(const MDLogger& mdlog,
2999 const DomdecOptions& options,
3000 const MdrunOptions& mdrunOptions,
3001 bool prefer1DAnd1Pulse,
3002 const gmx_mtop_t& mtop,
3003 const t_inputrec& ir,
3005 ArrayRef<const RVec> xGlobal);
3007 //! Build the resulting DD manager
3008 gmx_domdec_t* build(LocalAtomSetManager* atomSets);
3010 //! Objects used in constructing and configuring DD
3013 const MDLogger& mdlog_;
3014 //! Communication object
3016 //! User-supplied options configuring DD behavior
3017 const DomdecOptions options_;
3018 //! Global system topology
3019 const gmx_mtop_t& mtop_;
3020 //! User input values from the tpr file
3021 const t_inputrec& ir_;
3024 //! Internal objects used in constructing DD
3026 //! Settings combined from the user input
3027 DDSettings ddSettings_;
3028 //! Information derived from the simulation system
3029 DDSystemInfo systemInfo_;
3031 gmx_ddbox_t ddbox_ = { 0 };
3032 //! Organization of the DD grids
3033 DDGridSetup ddGridSetup_;
3034 //! Organzation of the DD ranks
3035 DDRankSetup ddRankSetup_;
3036 //! Number of DD cells in each dimension
3037 ivec ddCellIndex_ = { 0, 0, 0 };
3038 //! IDs of PME-only ranks
3039 std::vector<int> pmeRanks_;
3040 //! Contains a valid Cartesian-communicator-based setup, or defaults.
3041 CartesianRankSetup cartSetup_;
3045 DomainDecompositionBuilder::Impl::Impl(const MDLogger& mdlog,
3047 const DomdecOptions& options,
3048 const MdrunOptions& mdrunOptions,
3049 const bool prefer1DAnd1Pulse,
3050 const gmx_mtop_t& mtop,
3051 const t_inputrec& ir,
3053 ArrayRef<const RVec> xGlobal) :
3060 GMX_LOG(mdlog_.info).appendTextFormatted("\nInitializing Domain Decomposition on %d ranks", cr_->nnodes);
3062 ddSettings_ = getDDSettings(mdlog_, options_, mdrunOptions, ir_);
3064 if (prefer1DAnd1Pulse
3065 && canMake1DAnd1PulseDomainDecomposition(ddSettings_, cr_, cr_->nnodes, options_, mtop_,
3068 ddSettings_.request1DAnd1Pulse = true;
3071 if (ddSettings_.eFlop > 1)
3073 /* Ensure that we have different random flop counts on different ranks */
3074 srand(1 + cr_->nodeid);
3077 systemInfo_ = getSystemInfo(mdlog_, cr_, options_, mtop_, ir_, box, xGlobal);
3079 const int numRanksRequested = cr_->nnodes;
3080 const bool checkForLargePrimeFactors = (options_.numCells[0] <= 0);
3081 checkForValidRankCountRequests(numRanksRequested, EEL_PME(ir_.coulombtype),
3082 options_.numPmeRanks, checkForLargePrimeFactors);
3084 // DD grid setup uses a more different cell size limit for
3085 // automated setup than the one in systemInfo_. The latter is used
3086 // in set_dd_limits() to configure DLB, for example.
3087 const real gridSetupCellsizeLimit = getDDGridSetupCellSizeLimit(
3088 mdlog_, ddSettings_.request1DAnd1Pulse, !isDlbDisabled(ddSettings_.initialDlbState),
3089 options_.dlbScaling, ir_, systemInfo_.cellsizeLimit);
3090 ddGridSetup_ = getDDGridSetup(mdlog_, cr_, numRanksRequested, options_, ddSettings_, systemInfo_,
3091 gridSetupCellsizeLimit, mtop_, ir_, box, xGlobal, &ddbox_);
3092 checkDDGridSetup(ddGridSetup_, cr_, options_, ddSettings_, systemInfo_, gridSetupCellsizeLimit, ddbox_);
3094 cr_->npmenodes = ddGridSetup_.numPmeOnlyRanks;
3096 ddRankSetup_ = getDDRankSetup(mdlog_, cr_, ddGridSetup_, ir_);
3098 /* Generate the group communicator, also decides the duty of each rank */
3099 cartSetup_ = makeGroupCommunicators(mdlog_, ddSettings_, options_.rankOrder, ddRankSetup_, cr_,
3100 ddCellIndex_, &pmeRanks_);
3103 gmx_domdec_t* DomainDecompositionBuilder::Impl::build(LocalAtomSetManager* atomSets)
3105 gmx_domdec_t* dd = new gmx_domdec_t(ir_);
3107 copy_ivec(ddCellIndex_, dd->ci);
3109 dd->comm = init_dd_comm();
3111 dd->comm->ddRankSetup = ddRankSetup_;
3112 dd->comm->cartesianRankSetup = cartSetup_;
3114 set_dd_limits(mdlog_, cr_, dd, options_, ddSettings_, systemInfo_, ddGridSetup_,
3115 ddRankSetup_.numPPRanks, &mtop_, &ir_, ddbox_);
3117 setupGroupCommunication(mdlog_, ddSettings_, pmeRanks_, cr_, mtop_.natoms, dd);
3119 if (thisRankHasDuty(cr_, DUTY_PP))
3121 set_ddgrid_parameters(mdlog_, dd, options_.dlbScaling, &mtop_, &ir_, &ddbox_);
3123 setup_neighbor_relations(dd);
3126 /* Set overallocation to avoid frequent reallocation of arrays */
3127 set_over_alloc_dd(TRUE);
3129 dd->atomSets = atomSets;
3134 DomainDecompositionBuilder::DomainDecompositionBuilder(const MDLogger& mdlog,
3136 const DomdecOptions& options,
3137 const MdrunOptions& mdrunOptions,
3138 const bool prefer1DAnd1Pulse,
3139 const gmx_mtop_t& mtop,
3140 const t_inputrec& ir,
3142 ArrayRef<const RVec> xGlobal) :
3143 impl_(new Impl(mdlog, cr, options, mdrunOptions, prefer1DAnd1Pulse, mtop, ir, box, xGlobal))
3147 gmx_domdec_t* DomainDecompositionBuilder::build(LocalAtomSetManager* atomSets)
3149 return impl_->build(atomSets);
3152 DomainDecompositionBuilder::~DomainDecompositionBuilder() = default;
3156 static gmx_bool test_dd_cutoff(t_commrec* cr, const matrix box, gmx::ArrayRef<const gmx::RVec> x, real cutoffRequested)
3166 set_ddbox(*dd, false, box, true, x, &ddbox);
3170 for (d = 0; d < dd->ndim; d++)
3174 inv_cell_size = DD_CELL_MARGIN * dd->nc[dim] / ddbox.box_size[dim];
3175 if (dd->unitCellInfo.ddBoxIsDynamic)
3177 inv_cell_size *= DD_PRES_SCALE_MARGIN;
3180 np = 1 + static_cast<int>(cutoffRequested * inv_cell_size * ddbox.skew_fac[dim]);
3182 if (dd->comm->ddSettings.request1DAnd1Pulse && np > 1)
3187 if (!isDlbDisabled(dd->comm) && (dim < ddbox.npbcdim) && (dd->comm->cd[d].np_dlb > 0))
3189 if (np > dd->comm->cd[d].np_dlb)
3194 /* If a current local cell size is smaller than the requested
3195 * cut-off, we could still fix it, but this gets very complicated.
3196 * Without fixing here, we might actually need more checks.
3198 real cellSizeAlongDim =
3199 (dd->comm->cell_x1[dim] - dd->comm->cell_x0[dim]) * ddbox.skew_fac[dim];
3200 if (cellSizeAlongDim * dd->comm->cd[d].np_dlb < cutoffRequested)
3207 if (!isDlbDisabled(dd->comm))
3209 /* If DLB is not active yet, we don't need to check the grid jumps.
3210 * Actually we shouldn't, because then the grid jump data is not set.
3212 if (isDlbOn(dd->comm) && check_grid_jump(0, dd, cutoffRequested, &ddbox, FALSE))
3217 gmx_sumi(1, &LocallyLimited, cr);
3219 if (LocallyLimited > 0)
3228 gmx_bool change_dd_cutoff(t_commrec* cr, const matrix box, gmx::ArrayRef<const gmx::RVec> x, real cutoffRequested)
3230 gmx_bool bCutoffAllowed;
3232 bCutoffAllowed = test_dd_cutoff(cr, box, x, cutoffRequested);
3236 cr->dd->comm->systemInfo.cutoff = cutoffRequested;
3239 return bCutoffAllowed;