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.
5 * Copyright (c) 2019, by the GROMACS development team, led by
6 * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
7 * and including many others, as listed in the AUTHORS file in the
8 * top-level source directory and at http://www.gromacs.org.
10 * GROMACS is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU Lesser General Public License
12 * as published by the Free Software Foundation; either version 2.1
13 * of the License, or (at your option) any later version.
15 * GROMACS is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * Lesser General Public License for more details.
20 * You should have received a copy of the GNU Lesser General Public
21 * License along with GROMACS; if not, see
22 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
23 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
25 * If you want to redistribute modifications to GROMACS, please
26 * consider that scientific software is very special. Version
27 * control is crucial - bugs must be traceable. We will be happy to
28 * consider code for inclusion in the official distribution, but
29 * derived work must not be called official GROMACS. Details are found
30 * in the README & COPYING files - if they are missing, get the
31 * official version at http://www.gromacs.org.
33 * To help us fund GROMACS development, we humbly ask that you cite
34 * the research papers on the package. Check out http://www.gromacs.org.
54 #include "gromacs/domdec/builder.h"
55 #include "gromacs/domdec/collect.h"
56 #include "gromacs/domdec/dlb.h"
57 #include "gromacs/domdec/dlbtiming.h"
58 #include "gromacs/domdec/domdec_network.h"
59 #include "gromacs/domdec/ga2la.h"
60 #include "gromacs/domdec/gpuhaloexchange.h"
61 #include "gromacs/domdec/options.h"
62 #include "gromacs/domdec/partition.h"
63 #include "gromacs/gmxlib/network.h"
64 #include "gromacs/gmxlib/nrnb.h"
65 #include "gromacs/gpu_utils/gpu_utils.h"
66 #include "gromacs/hardware/hw_info.h"
67 #include "gromacs/listed_forces/manage_threading.h"
68 #include "gromacs/math/vec.h"
69 #include "gromacs/math/vectypes.h"
70 #include "gromacs/mdlib/calc_verletbuf.h"
71 #include "gromacs/mdlib/constr.h"
72 #include "gromacs/mdlib/constraintrange.h"
73 #include "gromacs/mdlib/updategroups.h"
74 #include "gromacs/mdlib/vsite.h"
75 #include "gromacs/mdtypes/commrec.h"
76 #include "gromacs/mdtypes/forceoutput.h"
77 #include "gromacs/mdtypes/inputrec.h"
78 #include "gromacs/mdtypes/mdrunoptions.h"
79 #include "gromacs/mdtypes/state.h"
80 #include "gromacs/pbcutil/ishift.h"
81 #include "gromacs/pbcutil/pbc.h"
82 #include "gromacs/pulling/pull.h"
83 #include "gromacs/timing/wallcycle.h"
84 #include "gromacs/topology/block.h"
85 #include "gromacs/topology/idef.h"
86 #include "gromacs/topology/ifunc.h"
87 #include "gromacs/topology/mtop_lookup.h"
88 #include "gromacs/topology/mtop_util.h"
89 #include "gromacs/topology/topology.h"
90 #include "gromacs/utility/basedefinitions.h"
91 #include "gromacs/utility/basenetwork.h"
92 #include "gromacs/utility/cstringutil.h"
93 #include "gromacs/utility/exceptions.h"
94 #include "gromacs/utility/fatalerror.h"
95 #include "gromacs/utility/gmxmpi.h"
96 #include "gromacs/utility/logger.h"
97 #include "gromacs/utility/real.h"
98 #include "gromacs/utility/smalloc.h"
99 #include "gromacs/utility/strconvert.h"
100 #include "gromacs/utility/stringstream.h"
101 #include "gromacs/utility/stringutil.h"
102 #include "gromacs/utility/textwriter.h"
104 #include "atomdistribution.h"
106 #include "cellsizes.h"
107 #include "distribute.h"
108 #include "domdec_constraints.h"
109 #include "domdec_internal.h"
110 #include "domdec_setup.h"
111 #include "domdec_vsite.h"
112 #include "redistribute.h"
115 // TODO remove this when moving domdec into gmx namespace
116 using gmx::DdRankOrder;
117 using gmx::DlbOption;
118 using gmx::DomdecOptions;
120 static const char* edlbs_names[int(DlbState::nr)] = { "off", "auto", "locked", "on", "on" };
122 /* The size per atom group of the cggl_flag buffer in gmx_domdec_comm_t */
125 /* The flags for the cggl_flag buffer in gmx_domdec_comm_t */
126 #define DD_FLAG_NRCG 65535
127 #define DD_FLAG_FW(d) (1 << (16 + (d)*2))
128 #define DD_FLAG_BW(d) (1 << (16 + (d)*2 + 1))
130 /* The DD zone order */
131 static const ivec dd_zo[DD_MAXZONE] = { { 0, 0, 0 }, { 1, 0, 0 }, { 1, 1, 0 }, { 0, 1, 0 },
132 { 0, 1, 1 }, { 0, 0, 1 }, { 1, 0, 1 }, { 1, 1, 1 } };
134 /* The non-bonded zone-pair setup for domain decomposition
135 * The first number is the i-zone, the second number the first j-zone seen by
136 * this i-zone, the third number the last+1 j-zone seen by this i-zone.
137 * As is, this is for 3D decomposition, where there are 4 i-zones.
138 * With 2D decomposition use only the first 2 i-zones and a last+1 j-zone of 4.
139 * With 1D decomposition use only the first i-zone and a last+1 j-zone of 2.
141 static const int ddNonbondedZonePairRanges[DD_MAXIZONE][3] = { { 0, 0, 8 },
148 #define dd_index(n,i) ((((i)[ZZ]*(n)[YY] + (i)[YY])*(n)[XX]) + (i)[XX])
150 static void index2xyz(ivec nc,int ind,ivec xyz)
152 xyz[XX] = ind % nc[XX];
153 xyz[YY] = (ind / nc[XX]) % nc[YY];
154 xyz[ZZ] = ind / (nc[YY]*nc[XX]);
158 static void ddindex2xyz(const ivec nc, int ind, ivec xyz)
160 xyz[XX] = ind / (nc[YY] * nc[ZZ]);
161 xyz[YY] = (ind / nc[ZZ]) % nc[YY];
162 xyz[ZZ] = ind % nc[ZZ];
165 static int ddcoord2ddnodeid(gmx_domdec_t* dd, ivec c)
169 const CartesianRankSetup& cartSetup = dd->comm->cartesianRankSetup;
170 const int ddindex = dd_index(dd->nc, c);
171 if (cartSetup.bCartesianPP_PME)
173 ddnodeid = cartSetup.ddindex2ddnodeid[ddindex];
175 else if (cartSetup.bCartesianPP)
178 MPI_Cart_rank(dd->mpi_comm_all, c, &ddnodeid);
189 int ddglatnr(const gmx_domdec_t* dd, int i)
199 if (i >= dd->comm->atomRanges.numAtomsTotal())
202 "glatnr called with %d, which is larger than the local number of atoms (%d)",
203 i, dd->comm->atomRanges.numAtomsTotal());
205 atnr = dd->globalAtomIndices[i] + 1;
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 int dd_numAtomsZones(const gmx_domdec_t& dd)
242 return dd.comm->atomRanges.end(DDAtomRanges::Type::Zones);
245 int dd_numHomeAtoms(const gmx_domdec_t& dd)
247 return dd.comm->atomRanges.numHomeAtoms();
250 int dd_natoms_mdatoms(const gmx_domdec_t* dd)
252 /* We currently set mdatoms entries for all atoms:
253 * local + non-local + communicated for vsite + constraints
256 return dd->comm->atomRanges.numAtomsTotal();
259 int dd_natoms_vsite(const gmx_domdec_t* dd)
261 return dd->comm->atomRanges.end(DDAtomRanges::Type::Vsites);
264 void dd_get_constraint_range(const gmx_domdec_t* dd, int* at_start, int* at_end)
266 *at_start = dd->comm->atomRanges.start(DDAtomRanges::Type::Constraints);
267 *at_end = dd->comm->atomRanges.end(DDAtomRanges::Type::Constraints);
270 void dd_move_x(gmx_domdec_t* dd, const matrix box, gmx::ArrayRef<gmx::RVec> x, gmx_wallcycle* wcycle)
272 wallcycle_start(wcycle, ewcMOVEX);
275 gmx_domdec_comm_t* comm;
276 gmx_domdec_comm_dim_t* cd;
277 rvec shift = { 0, 0, 0 };
278 gmx_bool bPBC, bScrew;
283 nat_tot = comm->atomRanges.numHomeAtoms();
284 for (int d = 0; d < dd->ndim; d++)
286 bPBC = (dd->ci[dd->dim[d]] == 0);
287 bScrew = (bPBC && dd->unitCellInfo.haveScrewPBC && dd->dim[d] == XX);
290 copy_rvec(box[dd->dim[d]], shift);
293 for (const gmx_domdec_ind_t& ind : cd->ind)
295 DDBufferAccess<gmx::RVec> sendBufferAccess(comm->rvecBuffer, ind.nsend[nzone + 1]);
296 gmx::ArrayRef<gmx::RVec>& sendBuffer = sendBufferAccess.buffer;
300 for (int j : ind.index)
302 sendBuffer[n] = x[j];
308 for (int j : ind.index)
310 /* We need to shift the coordinates */
311 for (int d = 0; d < DIM; d++)
313 sendBuffer[n][d] = x[j][d] + shift[d];
320 for (int j : ind.index)
323 sendBuffer[n][XX] = x[j][XX] + shift[XX];
325 * This operation requires a special shift force
326 * treatment, which is performed in calc_vir.
328 sendBuffer[n][YY] = box[YY][YY] - x[j][YY];
329 sendBuffer[n][ZZ] = box[ZZ][ZZ] - x[j][ZZ];
334 DDBufferAccess<gmx::RVec> receiveBufferAccess(
335 comm->rvecBuffer2, cd->receiveInPlace ? 0 : ind.nrecv[nzone + 1]);
337 gmx::ArrayRef<gmx::RVec> receiveBuffer;
338 if (cd->receiveInPlace)
340 receiveBuffer = gmx::arrayRefFromArray(x.data() + nat_tot, ind.nrecv[nzone + 1]);
344 receiveBuffer = receiveBufferAccess.buffer;
346 /* Send and receive the coordinates */
347 ddSendrecv(dd, d, dddirBackward, sendBuffer, receiveBuffer);
349 if (!cd->receiveInPlace)
352 for (int zone = 0; zone < nzone; zone++)
354 for (int i = ind.cell2at0[zone]; i < ind.cell2at1[zone]; i++)
356 x[i] = receiveBuffer[j++];
360 nat_tot += ind.nrecv[nzone + 1];
365 wallcycle_stop(wcycle, ewcMOVEX);
368 void dd_move_f(gmx_domdec_t* dd, gmx::ForceWithShiftForces* forceWithShiftForces, gmx_wallcycle* wcycle)
370 wallcycle_start(wcycle, ewcMOVEF);
372 gmx::ArrayRef<gmx::RVec> f = forceWithShiftForces->force();
373 gmx::ArrayRef<gmx::RVec> fshift = forceWithShiftForces->shiftForces();
375 gmx_domdec_comm_t& comm = *dd->comm;
376 int nzone = comm.zones.n / 2;
377 int nat_tot = comm.atomRanges.end(DDAtomRanges::Type::Zones);
378 for (int d = dd->ndim - 1; d >= 0; d--)
380 /* Only forces in domains near the PBC boundaries need to
381 consider PBC in the treatment of fshift */
382 const bool shiftForcesNeedPbc =
383 (forceWithShiftForces->computeVirial() && dd->ci[dd->dim[d]] == 0);
384 const bool applyScrewPbc =
385 (shiftForcesNeedPbc && dd->unitCellInfo.haveScrewPBC && dd->dim[d] == XX);
386 /* Determine which shift vector we need */
387 ivec vis = { 0, 0, 0 };
389 const int is = IVEC2IS(vis);
391 /* Loop over the pulses */
392 const gmx_domdec_comm_dim_t& cd = comm.cd[d];
393 for (int p = cd.numPulses() - 1; p >= 0; p--)
395 const gmx_domdec_ind_t& ind = cd.ind[p];
396 DDBufferAccess<gmx::RVec> receiveBufferAccess(comm.rvecBuffer, ind.nsend[nzone + 1]);
397 gmx::ArrayRef<gmx::RVec>& receiveBuffer = receiveBufferAccess.buffer;
399 nat_tot -= ind.nrecv[nzone + 1];
401 DDBufferAccess<gmx::RVec> sendBufferAccess(
402 comm.rvecBuffer2, cd.receiveInPlace ? 0 : ind.nrecv[nzone + 1]);
404 gmx::ArrayRef<gmx::RVec> sendBuffer;
405 if (cd.receiveInPlace)
407 sendBuffer = gmx::arrayRefFromArray(f.data() + nat_tot, ind.nrecv[nzone + 1]);
411 sendBuffer = sendBufferAccess.buffer;
413 for (int zone = 0; zone < nzone; zone++)
415 for (int i = ind.cell2at0[zone]; i < ind.cell2at1[zone]; i++)
417 sendBuffer[j++] = f[i];
421 /* Communicate the forces */
422 ddSendrecv(dd, d, dddirForward, sendBuffer, receiveBuffer);
423 /* Add the received forces */
425 if (!shiftForcesNeedPbc)
427 for (int j : ind.index)
429 for (int d = 0; d < DIM; d++)
431 f[j][d] += receiveBuffer[n][d];
436 else if (!applyScrewPbc)
438 for (int j : ind.index)
440 for (int d = 0; d < DIM; d++)
442 f[j][d] += receiveBuffer[n][d];
444 /* Add this force to the shift force */
445 for (int d = 0; d < DIM; d++)
447 fshift[is][d] += receiveBuffer[n][d];
454 for (int j : ind.index)
456 /* Rotate the force */
457 f[j][XX] += receiveBuffer[n][XX];
458 f[j][YY] -= receiveBuffer[n][YY];
459 f[j][ZZ] -= receiveBuffer[n][ZZ];
460 if (shiftForcesNeedPbc)
462 /* Add this force to the shift force */
463 for (int d = 0; d < DIM; d++)
465 fshift[is][d] += receiveBuffer[n][d];
474 wallcycle_stop(wcycle, ewcMOVEF);
477 /* Convenience function for extracting a real buffer from an rvec buffer
479 * To reduce the number of temporary communication buffers and avoid
480 * cache polution, we reuse gmx::RVec buffers for storing reals.
481 * This functions return a real buffer reference with the same number
482 * of elements as the gmx::RVec buffer (so 1/3 of the size in bytes).
484 static gmx::ArrayRef<real> realArrayRefFromRvecArrayRef(gmx::ArrayRef<gmx::RVec> arrayRef)
486 return gmx::arrayRefFromArray(as_rvec_array(arrayRef.data())[0], arrayRef.size());
489 void dd_atom_spread_real(gmx_domdec_t* dd, real v[])
492 gmx_domdec_comm_t* comm;
493 gmx_domdec_comm_dim_t* cd;
498 nat_tot = comm->atomRanges.numHomeAtoms();
499 for (int d = 0; d < dd->ndim; d++)
502 for (const gmx_domdec_ind_t& ind : cd->ind)
504 /* Note: We provision for RVec instead of real, so a factor of 3
505 * more than needed. The buffer actually already has this size
506 * and we pass a plain pointer below, so this does not matter.
508 DDBufferAccess<gmx::RVec> sendBufferAccess(comm->rvecBuffer, ind.nsend[nzone + 1]);
509 gmx::ArrayRef<real> sendBuffer = realArrayRefFromRvecArrayRef(sendBufferAccess.buffer);
511 for (int j : ind.index)
513 sendBuffer[n++] = v[j];
516 DDBufferAccess<gmx::RVec> receiveBufferAccess(
517 comm->rvecBuffer2, cd->receiveInPlace ? 0 : ind.nrecv[nzone + 1]);
519 gmx::ArrayRef<real> receiveBuffer;
520 if (cd->receiveInPlace)
522 receiveBuffer = gmx::arrayRefFromArray(v + nat_tot, ind.nrecv[nzone + 1]);
526 receiveBuffer = realArrayRefFromRvecArrayRef(receiveBufferAccess.buffer);
528 /* Send and receive the data */
529 ddSendrecv(dd, d, dddirBackward, sendBuffer, receiveBuffer);
530 if (!cd->receiveInPlace)
533 for (int zone = 0; zone < nzone; zone++)
535 for (int i = ind.cell2at0[zone]; i < ind.cell2at1[zone]; i++)
537 v[i] = receiveBuffer[j++];
541 nat_tot += ind.nrecv[nzone + 1];
547 void dd_atom_sum_real(gmx_domdec_t* dd, real v[])
550 gmx_domdec_comm_t* comm;
551 gmx_domdec_comm_dim_t* cd;
555 nzone = comm->zones.n / 2;
556 nat_tot = comm->atomRanges.end(DDAtomRanges::Type::Zones);
557 for (int d = dd->ndim - 1; d >= 0; d--)
560 for (int p = cd->numPulses() - 1; p >= 0; p--)
562 const gmx_domdec_ind_t& ind = cd->ind[p];
564 /* Note: We provision for RVec instead of real, so a factor of 3
565 * more than needed. The buffer actually already has this size
566 * and we typecast, so this works as intended.
568 DDBufferAccess<gmx::RVec> receiveBufferAccess(comm->rvecBuffer, ind.nsend[nzone + 1]);
569 gmx::ArrayRef<real> receiveBuffer = realArrayRefFromRvecArrayRef(receiveBufferAccess.buffer);
570 nat_tot -= ind.nrecv[nzone + 1];
572 DDBufferAccess<gmx::RVec> sendBufferAccess(
573 comm->rvecBuffer2, cd->receiveInPlace ? 0 : ind.nrecv[nzone + 1]);
575 gmx::ArrayRef<real> sendBuffer;
576 if (cd->receiveInPlace)
578 sendBuffer = gmx::arrayRefFromArray(v + nat_tot, ind.nrecv[nzone + 1]);
582 sendBuffer = realArrayRefFromRvecArrayRef(sendBufferAccess.buffer);
584 for (int zone = 0; zone < nzone; zone++)
586 for (int i = ind.cell2at0[zone]; i < ind.cell2at1[zone]; i++)
588 sendBuffer[j++] = v[i];
592 /* Communicate the forces */
593 ddSendrecv(dd, d, dddirForward, sendBuffer, receiveBuffer);
594 /* Add the received forces */
596 for (int j : ind.index)
598 v[j] += receiveBuffer[n];
606 real dd_cutoff_multibody(const gmx_domdec_t* dd)
608 const gmx_domdec_comm_t& comm = *dd->comm;
609 const DDSystemInfo& systemInfo = comm.systemInfo;
612 if (systemInfo.haveInterDomainMultiBodyBondeds)
614 if (comm.cutoff_mbody > 0)
616 r = comm.cutoff_mbody;
620 /* cutoff_mbody=0 means we do not have DLB */
621 r = comm.cellsize_min[dd->dim[0]];
622 for (int di = 1; di < dd->ndim; di++)
624 r = std::min(r, comm.cellsize_min[dd->dim[di]]);
626 if (comm.systemInfo.filterBondedCommunication)
628 r = std::max(r, comm.cutoff_mbody);
632 r = std::min(r, systemInfo.cutoff);
640 real dd_cutoff_twobody(const gmx_domdec_t* dd)
644 r_mb = dd_cutoff_multibody(dd);
646 return std::max(dd->comm->systemInfo.cutoff, r_mb);
650 static void dd_cart_coord2pmecoord(const DDRankSetup& ddRankSetup,
651 const CartesianRankSetup& cartSetup,
655 const int nc = ddRankSetup.numPPCells[cartSetup.cartpmedim];
656 const int ntot = cartSetup.ntot[cartSetup.cartpmedim];
657 copy_ivec(coord, coord_pme);
658 coord_pme[cartSetup.cartpmedim] =
659 nc + (coord[cartSetup.cartpmedim] * (ntot - nc) + (ntot - nc) / 2) / nc;
662 /* Returns the PME rank index in 0...npmenodes-1 for the PP cell with index ddCellIndex */
663 static int ddindex2pmeindex(const DDRankSetup& ddRankSetup, const int ddCellIndex)
665 const int npp = ddRankSetup.numPPRanks;
666 const int npme = ddRankSetup.numRanksDoingPme;
668 /* Here we assign a PME node to communicate with this DD node
669 * by assuming that the major index of both is x.
670 * We add npme/2 to obtain an even distribution.
672 return (ddCellIndex * npme + npme / 2) / npp;
675 static std::vector<int> dd_interleaved_pme_ranks(const DDRankSetup& ddRankSetup)
677 std::vector<int> pmeRanks(ddRankSetup.numRanksDoingPme);
680 for (int i = 0; i < ddRankSetup.numPPRanks; i++)
682 const int p0 = ddindex2pmeindex(ddRankSetup, i);
683 const int p1 = ddindex2pmeindex(ddRankSetup, i + 1);
684 if (i + 1 == ddRankSetup.numPPRanks || p1 > p0)
688 fprintf(debug, "pme_rank[%d] = %d\n", n, i + 1 + n);
690 pmeRanks[n] = i + 1 + n;
698 static int gmx_ddcoord2pmeindex(const t_commrec* cr, int x, int y, int z)
706 if (dd->comm->bCartesian) {
707 gmx_ddindex2xyz(dd->nc,ddindex,coords);
708 dd_coords2pmecoords(dd,coords,coords_pme);
709 copy_ivec(dd->ntot,nc);
710 nc[dd->cartpmedim] -= dd->nc[dd->cartpmedim];
711 coords_pme[dd->cartpmedim] -= dd->nc[dd->cartpmedim];
713 slab = (coords_pme[XX]*nc[YY] + coords_pme[YY])*nc[ZZ] + coords_pme[ZZ];
715 slab = (ddindex*cr->npmenodes + cr->npmenodes/2)/dd->nnodes;
721 slab = ddindex2pmeindex(dd->comm->ddRankSetup, dd_index(dd->nc, coords));
726 static int ddcoord2simnodeid(const t_commrec* cr, int x, int y, int z)
728 const CartesianRankSetup& cartSetup = cr->dd->comm->cartesianRankSetup;
729 ivec coords = { x, y, z };
732 if (cartSetup.bCartesianPP_PME)
735 MPI_Cart_rank(cr->mpi_comm_mysim, coords, &nodeid);
740 int ddindex = dd_index(cr->dd->nc, coords);
741 if (cartSetup.bCartesianPP)
743 nodeid = cartSetup.ddindex2simnodeid[ddindex];
747 if (cr->dd->comm->ddRankSetup.usePmeOnlyRanks)
749 nodeid = ddindex + gmx_ddcoord2pmeindex(cr, x, y, z);
761 static int dd_simnode2pmenode(const DDRankSetup& ddRankSetup,
762 const CartesianRankSetup& cartSetup,
763 gmx::ArrayRef<const int> pmeRanks,
764 const t_commrec gmx_unused* cr,
765 const int sim_nodeid)
769 /* This assumes a uniform x domain decomposition grid cell size */
770 if (cartSetup.bCartesianPP_PME)
773 ivec coord, coord_pme;
774 MPI_Cart_coords(cr->mpi_comm_mysim, sim_nodeid, DIM, coord);
775 if (coord[cartSetup.cartpmedim] < ddRankSetup.numPPCells[cartSetup.cartpmedim])
777 /* This is a PP rank */
778 dd_cart_coord2pmecoord(ddRankSetup, cartSetup, coord, coord_pme);
779 MPI_Cart_rank(cr->mpi_comm_mysim, coord_pme, &pmenode);
783 else if (cartSetup.bCartesianPP)
785 if (sim_nodeid < ddRankSetup.numPPRanks)
787 pmenode = ddRankSetup.numPPRanks + ddindex2pmeindex(ddRankSetup, sim_nodeid);
792 /* This assumes DD cells with identical x coordinates
793 * are numbered sequentially.
795 if (pmeRanks.empty())
797 if (sim_nodeid < ddRankSetup.numPPRanks)
799 /* The DD index equals the nodeid */
800 pmenode = ddRankSetup.numPPRanks + ddindex2pmeindex(ddRankSetup, sim_nodeid);
806 while (sim_nodeid > pmeRanks[i])
810 if (sim_nodeid < pmeRanks[i])
812 pmenode = pmeRanks[i];
820 NumPmeDomains getNumPmeDomains(const gmx_domdec_t* dd)
824 return { dd->comm->ddRankSetup.npmenodes_x, dd->comm->ddRankSetup.npmenodes_y };
832 std::vector<int> get_pme_ddranks(const t_commrec* cr, const int pmenodeid)
834 const DDRankSetup& ddRankSetup = cr->dd->comm->ddRankSetup;
835 const CartesianRankSetup& cartSetup = cr->dd->comm->cartesianRankSetup;
836 GMX_RELEASE_ASSERT(ddRankSetup.usePmeOnlyRanks,
837 "This function should only be called when PME-only ranks are in use");
838 std::vector<int> ddranks;
839 ddranks.reserve((ddRankSetup.numPPRanks + ddRankSetup.numRanksDoingPme - 1) / ddRankSetup.numRanksDoingPme);
841 for (int x = 0; x < ddRankSetup.numPPCells[XX]; x++)
843 for (int y = 0; y < ddRankSetup.numPPCells[YY]; y++)
845 for (int z = 0; z < ddRankSetup.numPPCells[ZZ]; z++)
847 if (cartSetup.bCartesianPP_PME)
849 ivec coord = { x, y, z };
851 dd_cart_coord2pmecoord(ddRankSetup, cartSetup, coord, coord_pme);
852 if (cr->dd->ci[XX] == coord_pme[XX] && cr->dd->ci[YY] == coord_pme[YY]
853 && cr->dd->ci[ZZ] == coord_pme[ZZ])
855 ddranks.push_back(ddcoord2simnodeid(cr, x, y, z));
860 /* The slab corresponds to the nodeid in the PME group */
861 if (gmx_ddcoord2pmeindex(cr, x, y, z) == pmenodeid)
863 ddranks.push_back(ddcoord2simnodeid(cr, x, y, z));
872 static gmx_bool receive_vir_ener(const gmx_domdec_t* dd, gmx::ArrayRef<const int> pmeRanks, const t_commrec* cr)
874 gmx_bool bReceive = TRUE;
876 const DDRankSetup& ddRankSetup = dd->comm->ddRankSetup;
877 if (ddRankSetup.usePmeOnlyRanks)
879 const CartesianRankSetup& cartSetup = dd->comm->cartesianRankSetup;
880 if (cartSetup.bCartesianPP_PME)
883 int pmenode = dd_simnode2pmenode(ddRankSetup, cartSetup, pmeRanks, cr, cr->sim_nodeid);
885 MPI_Cart_coords(cr->mpi_comm_mysim, cr->sim_nodeid, DIM, coords);
886 coords[cartSetup.cartpmedim]++;
887 if (coords[cartSetup.cartpmedim] < dd->nc[cartSetup.cartpmedim])
890 MPI_Cart_rank(cr->mpi_comm_mysim, coords, &rank);
891 if (dd_simnode2pmenode(ddRankSetup, cartSetup, pmeRanks, cr, rank) == pmenode)
893 /* This is not the last PP node for pmenode */
900 "Without MPI we should not have Cartesian PP-PME with #PMEnodes < #DDnodes");
905 int pmenode = dd_simnode2pmenode(ddRankSetup, cartSetup, pmeRanks, cr, cr->sim_nodeid);
906 if (cr->sim_nodeid + 1 < cr->nnodes
907 && dd_simnode2pmenode(ddRankSetup, cartSetup, pmeRanks, cr, cr->sim_nodeid + 1) == pmenode)
909 /* This is not the last PP node for pmenode */
918 static void set_slb_pme_dim_f(gmx_domdec_t* dd, int dim, real** dim_f)
920 gmx_domdec_comm_t* comm;
925 snew(*dim_f, dd->nc[dim] + 1);
927 for (i = 1; i < dd->nc[dim]; i++)
929 if (comm->slb_frac[dim])
931 (*dim_f)[i] = (*dim_f)[i - 1] + comm->slb_frac[dim][i - 1];
935 (*dim_f)[i] = static_cast<real>(i) / static_cast<real>(dd->nc[dim]);
938 (*dim_f)[dd->nc[dim]] = 1;
941 static void init_ddpme(gmx_domdec_t* dd, gmx_ddpme_t* ddpme, int dimind)
943 const DDRankSetup& ddRankSetup = dd->comm->ddRankSetup;
945 if (dimind == 0 && dd->dim[0] == YY && ddRankSetup.npmenodes_x == 1)
953 ddpme->dim_match = (ddpme->dim == dd->dim[dimind]);
955 ddpme->nslab = (ddpme->dim == 0 ? ddRankSetup.npmenodes_x : ddRankSetup.npmenodes_y);
957 if (ddpme->nslab <= 1)
962 const int nso = ddRankSetup.numRanksDoingPme / ddpme->nslab;
963 /* Determine for each PME slab the PP location range for dimension dim */
964 snew(ddpme->pp_min, ddpme->nslab);
965 snew(ddpme->pp_max, ddpme->nslab);
966 for (int slab = 0; slab < ddpme->nslab; slab++)
968 ddpme->pp_min[slab] = dd->nc[dd->dim[dimind]] - 1;
969 ddpme->pp_max[slab] = 0;
971 for (int i = 0; i < dd->nnodes; i++)
974 ddindex2xyz(dd->nc, i, xyz);
975 /* For y only use our y/z slab.
976 * This assumes that the PME x grid size matches the DD grid size.
978 if (dimind == 0 || xyz[XX] == dd->ci[XX])
980 const int pmeindex = ddindex2pmeindex(ddRankSetup, i);
984 slab = pmeindex / nso;
988 slab = pmeindex % ddpme->nslab;
990 ddpme->pp_min[slab] = std::min(ddpme->pp_min[slab], xyz[dimind]);
991 ddpme->pp_max[slab] = std::max(ddpme->pp_max[slab], xyz[dimind]);
995 set_slb_pme_dim_f(dd, ddpme->dim, &ddpme->slb_dim_f);
998 int dd_pme_maxshift_x(const gmx_domdec_t* dd)
1000 const DDRankSetup& ddRankSetup = dd->comm->ddRankSetup;
1002 if (ddRankSetup.ddpme[0].dim == XX)
1004 return ddRankSetup.ddpme[0].maxshift;
1012 int dd_pme_maxshift_y(const gmx_domdec_t* dd)
1014 const DDRankSetup& ddRankSetup = dd->comm->ddRankSetup;
1016 if (ddRankSetup.ddpme[0].dim == YY)
1018 return ddRankSetup.ddpme[0].maxshift;
1020 else if (ddRankSetup.npmedecompdim >= 2 && ddRankSetup.ddpme[1].dim == YY)
1022 return ddRankSetup.ddpme[1].maxshift;
1030 bool ddHaveSplitConstraints(const gmx_domdec_t& dd)
1032 return dd.comm->systemInfo.haveSplitConstraints;
1035 void dd_cycles_add(const gmx_domdec_t* dd, float cycles, int ddCycl)
1037 /* Note that the cycles value can be incorrect, either 0 or some
1038 * extremely large value, when our thread migrated to another core
1039 * with an unsynchronized cycle counter. If this happens less often
1040 * that once per nstlist steps, this will not cause issues, since
1041 * we later subtract the maximum value from the sum over nstlist steps.
1042 * A zero count will slightly lower the total, but that's a small effect.
1043 * Note that the main purpose of the subtraction of the maximum value
1044 * is to avoid throwing off the load balancing when stalls occur due
1045 * e.g. system activity or network congestion.
1047 dd->comm->cycl[ddCycl] += cycles;
1048 dd->comm->cycl_n[ddCycl]++;
1049 if (cycles > dd->comm->cycl_max[ddCycl])
1051 dd->comm->cycl_max[ddCycl] = cycles;
1056 static void make_load_communicator(gmx_domdec_t* dd, int dim_ind, ivec loc)
1061 gmx_bool bPartOfGroup = FALSE;
1063 dim = dd->dim[dim_ind];
1064 copy_ivec(loc, loc_c);
1065 for (i = 0; i < dd->nc[dim]; i++)
1068 rank = dd_index(dd->nc, loc_c);
1069 if (rank == dd->rank)
1071 /* This process is part of the group */
1072 bPartOfGroup = TRUE;
1075 MPI_Comm_split(dd->mpi_comm_all, bPartOfGroup ? 0 : MPI_UNDEFINED, dd->rank, &c_row);
1078 dd->comm->mpi_comm_load[dim_ind] = c_row;
1079 if (!isDlbDisabled(dd->comm))
1081 DDCellsizesWithDlb& cellsizes = dd->comm->cellsizesWithDlb[dim_ind];
1083 if (dd->ci[dim] == dd->master_ci[dim])
1085 /* This is the root process of this row */
1086 cellsizes.rowMaster = std::make_unique<RowMaster>();
1088 RowMaster& rowMaster = *cellsizes.rowMaster;
1089 rowMaster.cellFrac.resize(ddCellFractionBufferSize(dd, dim_ind));
1090 rowMaster.oldCellFrac.resize(dd->nc[dim] + 1);
1091 rowMaster.isCellMin.resize(dd->nc[dim]);
1094 rowMaster.bounds.resize(dd->nc[dim]);
1096 rowMaster.buf_ncd.resize(dd->nc[dim]);
1100 /* This is not a root process, we only need to receive cell_f */
1101 cellsizes.fracRow.resize(ddCellFractionBufferSize(dd, dim_ind));
1104 if (dd->ci[dim] == dd->master_ci[dim])
1106 snew(dd->comm->load[dim_ind].load, dd->nc[dim] * DD_NLOAD_MAX);
1112 void dd_setup_dlb_resource_sharing(const t_commrec* cr, int gpu_id)
1115 int physicalnode_id_hash;
1117 MPI_Comm mpi_comm_pp_physicalnode;
1119 if (!thisRankHasDuty(cr, DUTY_PP) || gpu_id < 0)
1121 /* Only ranks with short-ranged tasks (currently) use GPUs.
1122 * If we don't have GPUs assigned, there are no resources to share.
1127 physicalnode_id_hash = gmx_physicalnode_id_hash();
1133 fprintf(debug, "dd_setup_dd_dlb_gpu_sharing:\n");
1134 fprintf(debug, "DD PP rank %d physical node hash %d gpu_id %d\n", dd->rank,
1135 physicalnode_id_hash, gpu_id);
1137 /* Split the PP communicator over the physical nodes */
1138 /* TODO: See if we should store this (before), as it's also used for
1139 * for the nodecomm summation.
1141 // TODO PhysicalNodeCommunicator could be extended/used to handle
1142 // the need for per-node per-group communicators.
1143 MPI_Comm_split(dd->mpi_comm_all, physicalnode_id_hash, dd->rank, &mpi_comm_pp_physicalnode);
1144 MPI_Comm_split(mpi_comm_pp_physicalnode, gpu_id, dd->rank, &dd->comm->mpi_comm_gpu_shared);
1145 MPI_Comm_free(&mpi_comm_pp_physicalnode);
1146 MPI_Comm_size(dd->comm->mpi_comm_gpu_shared, &dd->comm->nrank_gpu_shared);
1150 fprintf(debug, "nrank_gpu_shared %d\n", dd->comm->nrank_gpu_shared);
1153 /* Note that some ranks could share a GPU, while others don't */
1155 if (dd->comm->nrank_gpu_shared == 1)
1157 MPI_Comm_free(&dd->comm->mpi_comm_gpu_shared);
1160 GMX_UNUSED_VALUE(cr);
1161 GMX_UNUSED_VALUE(gpu_id);
1165 static void make_load_communicators(gmx_domdec_t gmx_unused* dd)
1168 int dim0, dim1, i, j;
1173 fprintf(debug, "Making load communicators\n");
1176 dd->comm->load = new domdec_load_t[std::max(dd->ndim, 1)];
1177 snew(dd->comm->mpi_comm_load, std::max(dd->ndim, 1));
1185 make_load_communicator(dd, 0, loc);
1189 for (i = 0; i < dd->nc[dim0]; i++)
1192 make_load_communicator(dd, 1, loc);
1198 for (i = 0; i < dd->nc[dim0]; i++)
1202 for (j = 0; j < dd->nc[dim1]; j++)
1205 make_load_communicator(dd, 2, loc);
1212 fprintf(debug, "Finished making load communicators\n");
1217 /*! \brief Sets up the relation between neighboring domains and zones */
1218 static void setup_neighbor_relations(gmx_domdec_t* dd)
1222 gmx_domdec_zones_t* zones;
1223 GMX_ASSERT((dd->ndim >= 0) && (dd->ndim <= DIM), "Must have valid number of dimensions for DD");
1225 for (d = 0; d < dd->ndim; d++)
1228 copy_ivec(dd->ci, tmp);
1229 tmp[dim] = (tmp[dim] + 1) % dd->nc[dim];
1230 dd->neighbor[d][0] = ddcoord2ddnodeid(dd, tmp);
1231 copy_ivec(dd->ci, tmp);
1232 tmp[dim] = (tmp[dim] - 1 + dd->nc[dim]) % dd->nc[dim];
1233 dd->neighbor[d][1] = ddcoord2ddnodeid(dd, tmp);
1236 fprintf(debug, "DD rank %d neighbor ranks in dir %d are + %d - %d\n", dd->rank, dim,
1237 dd->neighbor[d][0], dd->neighbor[d][1]);
1241 int nzone = (1 << dd->ndim);
1242 int nizone = (1 << std::max(dd->ndim - 1, 0));
1243 assert(nizone >= 1 && nizone <= DD_MAXIZONE);
1245 zones = &dd->comm->zones;
1247 for (int i = 0; i < nzone; i++)
1250 clear_ivec(zones->shift[i]);
1251 for (d = 0; d < dd->ndim; d++)
1253 zones->shift[i][dd->dim[d]] = dd_zo[i][m++];
1258 for (int i = 0; i < nzone; i++)
1260 for (d = 0; d < DIM; d++)
1262 s[d] = dd->ci[d] - zones->shift[i][d];
1267 else if (s[d] >= dd->nc[d])
1273 for (int iZoneIndex = 0; iZoneIndex < nizone; iZoneIndex++)
1276 ddNonbondedZonePairRanges[iZoneIndex][0] == iZoneIndex,
1277 "The first element for each ddNonbondedZonePairRanges should match its index");
1279 DDPairInteractionRanges iZone;
1280 iZone.iZoneIndex = iZoneIndex;
1281 /* dd_zp3 is for 3D decomposition, for fewer dimensions use only
1282 * j-zones up to nzone.
1284 iZone.jZoneRange = gmx::Range<int>(std::min(ddNonbondedZonePairRanges[iZoneIndex][1], nzone),
1285 std::min(ddNonbondedZonePairRanges[iZoneIndex][2], nzone));
1286 for (dim = 0; dim < DIM; dim++)
1288 if (dd->nc[dim] == 1)
1290 /* All shifts should be allowed */
1291 iZone.shift0[dim] = -1;
1292 iZone.shift1[dim] = 1;
1296 /* Determine the min/max j-zone shift wrt the i-zone */
1297 iZone.shift0[dim] = 1;
1298 iZone.shift1[dim] = -1;
1299 for (int jZone : iZone.jZoneRange)
1301 int shift_diff = zones->shift[jZone][dim] - zones->shift[iZoneIndex][dim];
1302 if (shift_diff < iZone.shift0[dim])
1304 iZone.shift0[dim] = shift_diff;
1306 if (shift_diff > iZone.shift1[dim])
1308 iZone.shift1[dim] = shift_diff;
1314 zones->iZones.push_back(iZone);
1317 if (!isDlbDisabled(dd->comm))
1319 dd->comm->cellsizesWithDlb.resize(dd->ndim);
1322 if (dd->comm->ddSettings.recordLoad)
1324 make_load_communicators(dd);
1328 static void make_pp_communicator(const gmx::MDLogger& mdlog,
1330 t_commrec gmx_unused* cr,
1331 bool gmx_unused reorder)
1334 gmx_domdec_comm_t* comm = dd->comm;
1335 CartesianRankSetup& cartSetup = comm->cartesianRankSetup;
1337 if (cartSetup.bCartesianPP)
1339 /* Set up cartesian communication for the particle-particle part */
1341 .appendTextFormatted("Will use a Cartesian communicator: %d x %d x %d", dd->nc[XX],
1342 dd->nc[YY], dd->nc[ZZ]);
1345 for (int i = 0; i < DIM; i++)
1350 MPI_Cart_create(cr->mpi_comm_mygroup, DIM, dd->nc, periods, static_cast<int>(reorder), &comm_cart);
1351 /* We overwrite the old communicator with the new cartesian one */
1352 cr->mpi_comm_mygroup = comm_cart;
1355 dd->mpi_comm_all = cr->mpi_comm_mygroup;
1356 MPI_Comm_rank(dd->mpi_comm_all, &dd->rank);
1358 if (cartSetup.bCartesianPP_PME)
1360 /* Since we want to use the original cartesian setup for sim,
1361 * and not the one after split, we need to make an index.
1363 cartSetup.ddindex2ddnodeid.resize(dd->nnodes);
1364 cartSetup.ddindex2ddnodeid[dd_index(dd->nc, dd->ci)] = dd->rank;
1365 gmx_sumi(dd->nnodes, cartSetup.ddindex2ddnodeid.data(), cr);
1366 /* Get the rank of the DD master,
1367 * above we made sure that the master node is a PP node.
1378 MPI_Allreduce(&rank, &dd->masterrank, 1, MPI_INT, MPI_SUM, dd->mpi_comm_all);
1380 else if (cartSetup.bCartesianPP)
1382 if (!comm->ddRankSetup.usePmeOnlyRanks)
1384 /* The PP communicator is also
1385 * the communicator for this simulation
1387 cr->mpi_comm_mysim = cr->mpi_comm_mygroup;
1389 cr->nodeid = dd->rank;
1391 MPI_Cart_coords(dd->mpi_comm_all, dd->rank, DIM, dd->ci);
1393 /* We need to make an index to go from the coordinates
1394 * to the nodeid of this simulation.
1396 cartSetup.ddindex2simnodeid.resize(dd->nnodes);
1397 std::vector<int> buf(dd->nnodes);
1398 if (thisRankHasDuty(cr, DUTY_PP))
1400 buf[dd_index(dd->nc, dd->ci)] = cr->sim_nodeid;
1402 /* Communicate the ddindex to simulation nodeid index */
1403 MPI_Allreduce(buf.data(), cartSetup.ddindex2simnodeid.data(), dd->nnodes, MPI_INT, MPI_SUM,
1404 cr->mpi_comm_mysim);
1406 /* Determine the master coordinates and rank.
1407 * The DD master should be the same node as the master of this sim.
1409 for (int i = 0; i < dd->nnodes; i++)
1411 if (cartSetup.ddindex2simnodeid[i] == 0)
1413 ddindex2xyz(dd->nc, i, dd->master_ci);
1414 MPI_Cart_rank(dd->mpi_comm_all, dd->master_ci, &dd->masterrank);
1419 fprintf(debug, "The master rank is %d\n", dd->masterrank);
1424 /* No Cartesian communicators */
1425 /* We use the rank in dd->comm->all as DD index */
1426 ddindex2xyz(dd->nc, dd->rank, dd->ci);
1427 /* The simulation master nodeid is 0, so the DD master rank is also 0 */
1429 clear_ivec(dd->master_ci);
1434 .appendTextFormatted("Domain decomposition rank %d, coordinates %d %d %d\n", dd->rank,
1435 dd->ci[XX], dd->ci[YY], dd->ci[ZZ]);
1438 fprintf(debug, "Domain decomposition rank %d, coordinates %d %d %d\n\n", dd->rank,
1439 dd->ci[XX], dd->ci[YY], dd->ci[ZZ]);
1443 static void receive_ddindex2simnodeid(gmx_domdec_t* dd, t_commrec* cr)
1446 CartesianRankSetup& cartSetup = dd->comm->cartesianRankSetup;
1448 if (!cartSetup.bCartesianPP_PME && cartSetup.bCartesianPP)
1450 cartSetup.ddindex2simnodeid.resize(dd->nnodes);
1451 std::vector<int> buf(dd->nnodes);
1452 if (thisRankHasDuty(cr, DUTY_PP))
1454 buf[dd_index(dd->nc, dd->ci)] = cr->sim_nodeid;
1456 /* Communicate the ddindex to simulation nodeid index */
1457 MPI_Allreduce(buf.data(), cartSetup.ddindex2simnodeid.data(), dd->nnodes, MPI_INT, MPI_SUM,
1458 cr->mpi_comm_mysim);
1461 GMX_UNUSED_VALUE(dd);
1462 GMX_UNUSED_VALUE(cr);
1466 static CartesianRankSetup split_communicator(const gmx::MDLogger& mdlog,
1468 const DdRankOrder ddRankOrder,
1469 bool gmx_unused reorder,
1470 const DDRankSetup& ddRankSetup,
1472 std::vector<int>* pmeRanks)
1474 CartesianRankSetup cartSetup;
1476 cartSetup.bCartesianPP = (ddRankOrder == DdRankOrder::cartesian);
1477 cartSetup.bCartesianPP_PME = false;
1479 const ivec& numDDCells = ddRankSetup.numPPCells;
1480 /* Initially we set ntot to the number of PP cells */
1481 copy_ivec(numDDCells, cartSetup.ntot);
1483 if (cartSetup.bCartesianPP)
1485 const int numDDCellsTot = ddRankSetup.numPPRanks;
1487 for (int i = 1; i < DIM; i++)
1489 bDiv[i] = ((ddRankSetup.numRanksDoingPme * numDDCells[i]) % numDDCellsTot == 0);
1491 if (bDiv[YY] || bDiv[ZZ])
1493 cartSetup.bCartesianPP_PME = TRUE;
1494 /* If we have 2D PME decomposition, which is always in x+y,
1495 * we stack the PME only nodes in z.
1496 * Otherwise we choose the direction that provides the thinnest slab
1497 * of PME only nodes as this will have the least effect
1498 * on the PP communication.
1499 * But for the PME communication the opposite might be better.
1501 if (bDiv[ZZ] && (ddRankSetup.npmenodes_y > 1 || !bDiv[YY] || numDDCells[YY] > numDDCells[ZZ]))
1503 cartSetup.cartpmedim = ZZ;
1507 cartSetup.cartpmedim = YY;
1509 cartSetup.ntot[cartSetup.cartpmedim] +=
1510 (ddRankSetup.numRanksDoingPme * numDDCells[cartSetup.cartpmedim]) / numDDCellsTot;
1515 .appendTextFormatted(
1516 "Number of PME-only ranks (%d) is not a multiple of nx*ny (%d*%d) or "
1518 ddRankSetup.numRanksDoingPme, numDDCells[XX], numDDCells[YY],
1519 numDDCells[XX], numDDCells[ZZ]);
1521 .appendText("Will not use a Cartesian communicator for PP <-> PME\n");
1525 if (cartSetup.bCartesianPP_PME)
1532 .appendTextFormatted(
1533 "Will use a Cartesian communicator for PP <-> PME: %d x %d x %d",
1534 cartSetup.ntot[XX], cartSetup.ntot[YY], cartSetup.ntot[ZZ]);
1536 for (int i = 0; i < DIM; i++)
1541 MPI_Cart_create(cr->mpi_comm_mysim, DIM, cartSetup.ntot, periods, static_cast<int>(reorder),
1543 MPI_Comm_rank(comm_cart, &rank);
1544 if (MASTER(cr) && rank != 0)
1546 gmx_fatal(FARGS, "MPI rank 0 was renumbered by MPI_Cart_create, we do not allow this");
1549 /* With this assigment we loose the link to the original communicator
1550 * which will usually be MPI_COMM_WORLD, unless have multisim.
1552 cr->mpi_comm_mysim = comm_cart;
1553 cr->sim_nodeid = rank;
1555 MPI_Cart_coords(cr->mpi_comm_mysim, cr->sim_nodeid, DIM, ddCellIndex);
1558 .appendTextFormatted("Cartesian rank %d, coordinates %d %d %d\n", cr->sim_nodeid,
1559 ddCellIndex[XX], ddCellIndex[YY], ddCellIndex[ZZ]);
1561 if (ddCellIndex[cartSetup.cartpmedim] < numDDCells[cartSetup.cartpmedim])
1565 if (!ddRankSetup.usePmeOnlyRanks
1566 || ddCellIndex[cartSetup.cartpmedim] >= numDDCells[cartSetup.cartpmedim])
1568 cr->duty = DUTY_PME;
1571 /* Split the sim communicator into PP and PME only nodes */
1572 MPI_Comm_split(cr->mpi_comm_mysim, getThisRankDuties(cr),
1573 dd_index(cartSetup.ntot, ddCellIndex), &cr->mpi_comm_mygroup);
1575 GMX_UNUSED_VALUE(ddCellIndex);
1580 switch (ddRankOrder)
1582 case DdRankOrder::pp_pme:
1583 GMX_LOG(mdlog.info).appendText("Order of the ranks: PP first, PME last");
1585 case DdRankOrder::interleave:
1586 /* Interleave the PP-only and PME-only ranks */
1587 GMX_LOG(mdlog.info).appendText("Interleaving PP and PME ranks");
1588 *pmeRanks = dd_interleaved_pme_ranks(ddRankSetup);
1590 case DdRankOrder::cartesian: break;
1591 default: gmx_fatal(FARGS, "Invalid ddRankOrder=%d", static_cast<int>(ddRankOrder));
1594 if (dd_simnode2pmenode(ddRankSetup, cartSetup, *pmeRanks, cr, cr->sim_nodeid) == -1)
1596 cr->duty = DUTY_PME;
1603 /* Split the sim communicator into PP and PME only nodes */
1604 MPI_Comm_split(cr->mpi_comm_mysim, getThisRankDuties(cr), cr->nodeid, &cr->mpi_comm_mygroup);
1605 MPI_Comm_rank(cr->mpi_comm_mygroup, &cr->nodeid);
1610 .appendTextFormatted("This rank does only %s work.\n",
1611 thisRankHasDuty(cr, DUTY_PP) ? "particle-particle" : "PME-mesh");
1616 /*! \brief Makes the PP communicator and the PME communicator, when needed
1618 * Returns the Cartesian rank setup.
1619 * Sets \p cr->mpi_comm_mygroup
1620 * For PP ranks, sets the DD PP cell index in \p ddCellIndex.
1621 * With separate PME ranks in interleaved order, set the PME ranks in \p pmeRanks.
1623 static CartesianRankSetup makeGroupCommunicators(const gmx::MDLogger& mdlog,
1624 const DDSettings& ddSettings,
1625 const DdRankOrder ddRankOrder,
1626 const DDRankSetup& ddRankSetup,
1629 std::vector<int>* pmeRanks)
1631 CartesianRankSetup cartSetup;
1633 if (ddRankSetup.usePmeOnlyRanks)
1635 /* Split the communicator into a PP and PME part */
1636 cartSetup = split_communicator(mdlog, cr, ddRankOrder, ddSettings.useCartesianReorder,
1637 ddRankSetup, ddCellIndex, pmeRanks);
1641 /* All nodes do PP and PME */
1642 /* We do not require separate communicators */
1643 cr->mpi_comm_mygroup = cr->mpi_comm_mysim;
1645 cartSetup.bCartesianPP = false;
1646 cartSetup.bCartesianPP_PME = false;
1652 /*! \brief For PP ranks, sets or makes the communicator
1654 * For PME ranks get the rank id.
1655 * For PP only ranks, sets the PME-only rank.
1657 static void setupGroupCommunication(const gmx::MDLogger& mdlog,
1658 const DDSettings& ddSettings,
1659 gmx::ArrayRef<const int> pmeRanks,
1661 const int numAtomsInSystem,
1664 const DDRankSetup& ddRankSetup = dd->comm->ddRankSetup;
1665 const CartesianRankSetup& cartSetup = dd->comm->cartesianRankSetup;
1667 if (thisRankHasDuty(cr, DUTY_PP))
1669 /* Copy or make a new PP communicator */
1671 /* We (possibly) reordered the nodes in split_communicator,
1672 * so it is no longer required in make_pp_communicator.
1674 const bool useCartesianReorder = (ddSettings.useCartesianReorder && !cartSetup.bCartesianPP_PME);
1676 make_pp_communicator(mdlog, dd, cr, useCartesianReorder);
1680 receive_ddindex2simnodeid(dd, cr);
1683 if (!thisRankHasDuty(cr, DUTY_PME))
1685 /* Set up the commnuication to our PME node */
1686 dd->pme_nodeid = dd_simnode2pmenode(ddRankSetup, cartSetup, pmeRanks, cr, cr->sim_nodeid);
1687 dd->pme_receive_vir_ener = receive_vir_ener(dd, pmeRanks, cr);
1690 fprintf(debug, "My pme_nodeid %d receive ener %s\n", dd->pme_nodeid,
1691 gmx::boolToString(dd->pme_receive_vir_ener));
1696 dd->pme_nodeid = -1;
1699 /* We can not use DDMASTER(dd), because dd->masterrank is set later */
1702 dd->ma = std::make_unique<AtomDistribution>(dd->nc, numAtomsInSystem, numAtomsInSystem);
1706 static real* get_slb_frac(const gmx::MDLogger& mdlog, const char* dir, int nc, const char* size_string)
1708 real * slb_frac, tot;
1713 if (nc > 1 && size_string != nullptr)
1715 GMX_LOG(mdlog.info).appendTextFormatted("Using static load balancing for the %s direction", dir);
1718 for (i = 0; i < nc; i++)
1721 sscanf(size_string, "%20lf%n", &dbl, &n);
1725 "Incorrect or not enough DD cell size entries for direction %s: '%s'",
1733 std::string relativeCellSizes = "Relative cell sizes:";
1734 for (i = 0; i < nc; i++)
1737 relativeCellSizes += gmx::formatString(" %5.3f", slb_frac[i]);
1739 GMX_LOG(mdlog.info).appendText(relativeCellSizes);
1745 static int multi_body_bondeds_count(const gmx_mtop_t* mtop)
1748 gmx_mtop_ilistloop_t iloop = gmx_mtop_ilistloop_init(mtop);
1750 while (const InteractionLists* ilists = gmx_mtop_ilistloop_next(iloop, &nmol))
1752 for (auto& ilist : extractILists(*ilists, IF_BOND))
1754 if (NRAL(ilist.functionType) > 2)
1756 n += nmol * (ilist.iatoms.size() / ilistStride(ilist));
1764 static int dd_getenv(const gmx::MDLogger& mdlog, const char* env_var, int def)
1770 val = getenv(env_var);
1773 if (sscanf(val, "%20d", &nst) <= 0)
1777 GMX_LOG(mdlog.info).appendTextFormatted("Found env.var. %s = %s, using value %d", env_var, val, nst);
1783 static void check_dd_restrictions(const gmx_domdec_t* dd, const t_inputrec* ir, const gmx::MDLogger& mdlog)
1785 if (ir->ePBC == epbcSCREW && (dd->nc[XX] == 1 || dd->nc[YY] > 1 || dd->nc[ZZ] > 1))
1787 gmx_fatal(FARGS, "With pbc=%s can only do domain decomposition in the x-direction",
1788 epbc_names[ir->ePBC]);
1791 if (ir->nstlist == 0)
1793 gmx_fatal(FARGS, "Domain decomposition does not work with nstlist=0");
1796 if (ir->comm_mode == ecmANGULAR && ir->ePBC != epbcNONE)
1798 GMX_LOG(mdlog.warning)
1800 "comm-mode angular will give incorrect results when the comm group "
1801 "partially crosses a periodic boundary");
1805 static real average_cellsize_min(const gmx_ddbox_t& ddbox, const ivec numDomains)
1807 real r = ddbox.box_size[XX];
1808 for (int d = 0; d < DIM; d++)
1810 if (numDomains[d] > 1)
1812 /* Check using the initial average cell size */
1813 r = std::min(r, ddbox.box_size[d] * ddbox.skew_fac[d] / numDomains[d]);
1820 /*! \brief Depending on the DLB initial value return the DLB switched off state or issue an error.
1822 static DlbState forceDlbOffOrBail(DlbState cmdlineDlbState,
1823 const std::string& reasonStr,
1824 const gmx::MDLogger& mdlog)
1826 std::string dlbNotSupportedErr = "Dynamic load balancing requested, but ";
1827 std::string dlbDisableNote = "NOTE: disabling dynamic load balancing as ";
1829 if (cmdlineDlbState == DlbState::onUser)
1831 gmx_fatal(FARGS, "%s", (dlbNotSupportedErr + reasonStr).c_str());
1833 else if (cmdlineDlbState == DlbState::offCanTurnOn)
1835 GMX_LOG(mdlog.info).appendText(dlbDisableNote + reasonStr);
1837 return DlbState::offForever;
1840 /*! \brief Return the dynamic load balancer's initial state based on initial conditions and user inputs.
1842 * This function parses the parameters of "-dlb" command line option setting
1843 * corresponding state values. Then it checks the consistency of the determined
1844 * state with other run parameters and settings. As a result, the initial state
1845 * may be altered or an error may be thrown if incompatibility of options is detected.
1847 * \param [in] mdlog Logger.
1848 * \param [in] dlbOption Enum value for the DLB option.
1849 * \param [in] bRecordLoad True if the load balancer is recording load information.
1850 * \param [in] mdrunOptions Options for mdrun.
1851 * \param [in] ir Pointer mdrun to input parameters.
1852 * \returns DLB initial/startup state.
1854 static DlbState determineInitialDlbState(const gmx::MDLogger& mdlog,
1855 DlbOption dlbOption,
1856 gmx_bool bRecordLoad,
1857 const gmx::MdrunOptions& mdrunOptions,
1858 const t_inputrec* ir)
1860 DlbState dlbState = DlbState::offCanTurnOn;
1864 case DlbOption::turnOnWhenUseful: dlbState = DlbState::offCanTurnOn; break;
1865 case DlbOption::no: dlbState = DlbState::offUser; break;
1866 case DlbOption::yes: dlbState = DlbState::onUser; break;
1867 default: gmx_incons("Invalid dlbOption enum value");
1870 /* Reruns don't support DLB: bail or override auto mode */
1871 if (mdrunOptions.rerun)
1873 std::string reasonStr = "it is not supported in reruns.";
1874 return forceDlbOffOrBail(dlbState, reasonStr, mdlog);
1877 /* Unsupported integrators */
1878 if (!EI_DYNAMICS(ir->eI))
1880 auto reasonStr = gmx::formatString(
1881 "it is only supported with dynamics, not with integrator '%s'.", EI(ir->eI));
1882 return forceDlbOffOrBail(dlbState, reasonStr, mdlog);
1885 /* Without cycle counters we can't time work to balance on */
1888 std::string reasonStr =
1889 "cycle counters unsupported or not enabled in the operating system kernel.";
1890 return forceDlbOffOrBail(dlbState, reasonStr, mdlog);
1893 if (mdrunOptions.reproducible)
1895 std::string reasonStr = "you started a reproducible run.";
1898 case DlbState::offUser: break;
1899 case DlbState::offForever:
1900 GMX_RELEASE_ASSERT(false, "DlbState::offForever is not a valid initial state");
1902 case DlbState::offCanTurnOn: return forceDlbOffOrBail(dlbState, reasonStr, mdlog);
1903 case DlbState::onCanTurnOff:
1904 GMX_RELEASE_ASSERT(false, "DlbState::offCanTurnOff is not a valid initial state");
1906 case DlbState::onUser:
1907 return forceDlbOffOrBail(
1910 + " In load balanced runs binary reproducibility cannot be "
1914 gmx_fatal(FARGS, "Death horror: undefined case (%d) for load balancing choice",
1915 static_cast<int>(dlbState));
1922 static gmx_domdec_comm_t* init_dd_comm()
1924 gmx_domdec_comm_t* comm = new gmx_domdec_comm_t;
1926 comm->n_load_have = 0;
1927 comm->n_load_collect = 0;
1929 comm->haveTurnedOffDlb = false;
1931 for (int i = 0; i < static_cast<int>(DDAtomRanges::Type::Number); i++)
1933 comm->sum_nat[i] = 0;
1937 comm->load_step = 0;
1940 clear_ivec(comm->load_lim);
1944 /* This should be replaced by a unique pointer */
1945 comm->balanceRegion = ddBalanceRegionAllocate();
1950 /* Returns whether mtop contains constraints and/or vsites */
1951 static bool systemHasConstraintsOrVsites(const gmx_mtop_t& mtop)
1953 auto ilistLoop = gmx_mtop_ilistloop_init(mtop);
1955 while (const InteractionLists* ilists = gmx_mtop_ilistloop_next(ilistLoop, &nmol))
1957 if (!extractILists(*ilists, IF_CONSTRAINT | IF_VSITE).empty())
1966 static void setupUpdateGroups(const gmx::MDLogger& mdlog,
1967 const gmx_mtop_t& mtop,
1968 const t_inputrec& inputrec,
1969 const real cutoffMargin,
1970 DDSystemInfo* systemInfo)
1972 /* When we have constraints and/or vsites, it is beneficial to use
1973 * update groups (when possible) to allow independent update of groups.
1975 if (!systemHasConstraintsOrVsites(mtop))
1977 /* No constraints or vsites, atoms can be updated independently */
1981 systemInfo->updateGroupingPerMoleculetype = gmx::makeUpdateGroups(mtop);
1982 systemInfo->useUpdateGroups = (!systemInfo->updateGroupingPerMoleculetype.empty()
1983 && getenv("GMX_NO_UPDATEGROUPS") == nullptr);
1985 if (systemInfo->useUpdateGroups)
1987 int numUpdateGroups = 0;
1988 for (const auto& molblock : mtop.molblock)
1990 numUpdateGroups += molblock.nmol
1991 * systemInfo->updateGroupingPerMoleculetype[molblock.type].numBlocks();
1994 systemInfo->maxUpdateGroupRadius = computeMaxUpdateGroupRadius(
1995 mtop, systemInfo->updateGroupingPerMoleculetype, maxReferenceTemperature(inputrec));
1997 /* To use update groups, the large domain-to-domain cutoff distance
1998 * should be compatible with the box size.
2000 systemInfo->useUpdateGroups = (atomToAtomIntoDomainToDomainCutoff(*systemInfo, 0) < cutoffMargin);
2002 if (systemInfo->useUpdateGroups)
2005 .appendTextFormatted(
2006 "Using update groups, nr %d, average size %.1f atoms, max. radius %.3f "
2008 numUpdateGroups, mtop.natoms / static_cast<double>(numUpdateGroups),
2009 systemInfo->maxUpdateGroupRadius);
2014 .appendTextFormatted(
2015 "The combination of rlist and box size prohibits the use of update "
2017 systemInfo->updateGroupingPerMoleculetype.clear();
2022 UnitCellInfo::UnitCellInfo(const t_inputrec& ir) :
2023 npbcdim(ePBC2npbcdim(ir.ePBC)),
2024 numBoundedDimensions(inputrec2nboundeddim(&ir)),
2025 ddBoxIsDynamic(numBoundedDimensions < DIM || inputrecDynamicBox(&ir)),
2026 haveScrewPBC(ir.ePBC == epbcSCREW)
2030 /* Returns whether molecules are always whole, i.e. not broken by PBC */
2031 static bool moleculesAreAlwaysWhole(const gmx_mtop_t& mtop,
2032 const bool useUpdateGroups,
2033 gmx::ArrayRef<const gmx::RangePartitioning> updateGroupingPerMoleculetype)
2035 if (useUpdateGroups)
2037 GMX_RELEASE_ASSERT(updateGroupingPerMoleculetype.size() == mtop.moltype.size(),
2038 "Need one grouping per moltype");
2039 for (size_t mol = 0; mol < mtop.moltype.size(); mol++)
2041 if (updateGroupingPerMoleculetype[mol].numBlocks() > 1)
2049 for (const auto& moltype : mtop.moltype)
2051 if (moltype.atoms.nr > 1)
2061 /*! \brief Generate the simulation system information */
2062 static DDSystemInfo getSystemInfo(const gmx::MDLogger& mdlog,
2063 const t_commrec* cr,
2064 const DomdecOptions& options,
2065 const gmx_mtop_t& mtop,
2066 const t_inputrec& ir,
2068 gmx::ArrayRef<const gmx::RVec> xGlobal)
2070 const real tenPercentMargin = 1.1;
2072 DDSystemInfo systemInfo;
2074 /* We need to decide on update groups early, as this affects communication distances */
2075 systemInfo.useUpdateGroups = false;
2076 if (ir.cutoff_scheme == ecutsVERLET)
2078 real cutoffMargin = std::sqrt(max_cutoff2(ir.ePBC, box)) - ir.rlist;
2079 setupUpdateGroups(mdlog, mtop, ir, cutoffMargin, &systemInfo);
2082 systemInfo.moleculesAreAlwaysWhole = moleculesAreAlwaysWhole(
2083 mtop, systemInfo.useUpdateGroups, systemInfo.updateGroupingPerMoleculetype);
2084 systemInfo.haveInterDomainBondeds =
2085 (!systemInfo.moleculesAreAlwaysWhole || mtop.bIntermolecularInteractions);
2086 systemInfo.haveInterDomainMultiBodyBondeds =
2087 (systemInfo.haveInterDomainBondeds && multi_body_bondeds_count(&mtop) > 0);
2089 if (systemInfo.useUpdateGroups)
2091 systemInfo.haveSplitConstraints = false;
2092 systemInfo.haveSplitSettles = false;
2096 systemInfo.haveSplitConstraints = (gmx_mtop_ftype_count(mtop, F_CONSTR) > 0
2097 || gmx_mtop_ftype_count(mtop, F_CONSTRNC) > 0);
2098 systemInfo.haveSplitSettles = (gmx_mtop_ftype_count(mtop, F_SETTLE) > 0);
2103 /* Set the cut-off to some very large value,
2104 * so we don't need if statements everywhere in the code.
2105 * We use sqrt, since the cut-off is squared in some places.
2107 systemInfo.cutoff = GMX_CUTOFF_INF;
2111 systemInfo.cutoff = atomToAtomIntoDomainToDomainCutoff(systemInfo, ir.rlist);
2113 systemInfo.minCutoffForMultiBody = 0;
2115 /* Determine the minimum cell size limit, affected by many factors */
2116 systemInfo.cellsizeLimit = 0;
2117 systemInfo.filterBondedCommunication = false;
2119 /* We do not allow home atoms to move beyond the neighboring domain
2120 * between domain decomposition steps, which limits the cell size.
2121 * Get an estimate of cell size limit due to atom displacement.
2122 * In most cases this is a large overestimate, because it assumes
2123 * non-interaction atoms.
2124 * We set the chance to 1 in a trillion steps.
2126 constexpr real c_chanceThatAtomMovesBeyondDomain = 1e-12;
2127 const real limitForAtomDisplacement = minCellSizeForAtomDisplacement(
2128 mtop, ir, systemInfo.updateGroupingPerMoleculetype, c_chanceThatAtomMovesBeyondDomain);
2129 GMX_LOG(mdlog.info).appendTextFormatted("Minimum cell size due to atom displacement: %.3f nm", limitForAtomDisplacement);
2131 systemInfo.cellsizeLimit = std::max(systemInfo.cellsizeLimit, limitForAtomDisplacement);
2133 /* TODO: PME decomposition currently requires atoms not to be more than
2134 * 2/3 of comm->cutoff, which is >=rlist, outside of their domain.
2135 * In nearly all cases, limitForAtomDisplacement will be smaller
2136 * than 2/3*rlist, so the PME requirement is satisfied.
2137 * But it would be better for both correctness and performance
2138 * to use limitForAtomDisplacement instead of 2/3*comm->cutoff.
2139 * Note that we would need to improve the pairlist buffer case.
2142 if (systemInfo.haveInterDomainBondeds)
2144 if (options.minimumCommunicationRange > 0)
2146 systemInfo.minCutoffForMultiBody =
2147 atomToAtomIntoDomainToDomainCutoff(systemInfo, options.minimumCommunicationRange);
2148 if (options.useBondedCommunication)
2150 systemInfo.filterBondedCommunication =
2151 (systemInfo.minCutoffForMultiBody > systemInfo.cutoff);
2155 systemInfo.cutoff = std::max(systemInfo.cutoff, systemInfo.minCutoffForMultiBody);
2158 else if (ir.bPeriodicMols)
2160 /* Can not easily determine the required cut-off */
2161 GMX_LOG(mdlog.warning)
2163 "NOTE: Periodic molecules are present in this system. Because of this, "
2164 "the domain decomposition algorithm cannot easily determine the "
2165 "minimum cell size that it requires for treating bonded interactions. "
2166 "Instead, domain decomposition will assume that half the non-bonded "
2167 "cut-off will be a suitable lower bound.");
2168 systemInfo.minCutoffForMultiBody = systemInfo.cutoff / 2;
2176 dd_bonded_cg_distance(mdlog, &mtop, &ir, as_rvec_array(xGlobal.data()), box,
2177 options.checkBondedInteractions, &r_2b, &r_mb);
2179 gmx_bcast(sizeof(r_2b), &r_2b, cr);
2180 gmx_bcast(sizeof(r_mb), &r_mb, cr);
2182 /* We use an initial margin of 10% for the minimum cell size,
2183 * except when we are just below the non-bonded cut-off.
2185 if (options.useBondedCommunication)
2187 if (std::max(r_2b, r_mb) > systemInfo.cutoff)
2189 const real r_bonded = std::max(r_2b, r_mb);
2190 systemInfo.minCutoffForMultiBody = tenPercentMargin * r_bonded;
2191 /* This is the (only) place where we turn on the filtering */
2192 systemInfo.filterBondedCommunication = true;
2196 const real r_bonded = r_mb;
2197 systemInfo.minCutoffForMultiBody =
2198 std::min(tenPercentMargin * r_bonded, systemInfo.cutoff);
2200 /* We determine cutoff_mbody later */
2201 systemInfo.increaseMultiBodyCutoff = true;
2205 /* No special bonded communication,
2206 * simply increase the DD cut-off.
2208 systemInfo.minCutoffForMultiBody = tenPercentMargin * std::max(r_2b, r_mb);
2209 systemInfo.cutoff = std::max(systemInfo.cutoff, systemInfo.minCutoffForMultiBody);
2213 .appendTextFormatted("Minimum cell size due to bonded interactions: %.3f nm",
2214 systemInfo.minCutoffForMultiBody);
2216 systemInfo.cellsizeLimit = std::max(systemInfo.cellsizeLimit, systemInfo.minCutoffForMultiBody);
2219 systemInfo.constraintCommunicationRange = 0;
2220 if (systemInfo.haveSplitConstraints && options.constraintCommunicationRange <= 0)
2222 /* There is a cell size limit due to the constraints (P-LINCS) */
2223 systemInfo.constraintCommunicationRange = gmx::constr_r_max(mdlog, &mtop, &ir);
2225 .appendTextFormatted("Estimated maximum distance required for P-LINCS: %.3f nm",
2226 systemInfo.constraintCommunicationRange);
2227 if (systemInfo.constraintCommunicationRange > systemInfo.cellsizeLimit)
2231 "This distance will limit the DD cell size, you can override this with "
2235 else if (options.constraintCommunicationRange > 0)
2237 /* Here we do not check for dd->splitConstraints.
2238 * because one can also set a cell size limit for virtual sites only
2239 * and at this point we don't know yet if there are intercg v-sites.
2242 .appendTextFormatted("User supplied maximum distance required for P-LINCS: %.3f nm",
2243 options.constraintCommunicationRange);
2244 systemInfo.constraintCommunicationRange = options.constraintCommunicationRange;
2246 systemInfo.cellsizeLimit = std::max(systemInfo.cellsizeLimit, systemInfo.constraintCommunicationRange);
2251 /*! \brief Exit with a fatal error if the DDGridSetup cannot be
2253 static void checkDDGridSetup(const DDGridSetup& ddGridSetup,
2254 const t_commrec* cr,
2255 const DomdecOptions& options,
2256 const DDSettings& ddSettings,
2257 const DDSystemInfo& systemInfo,
2258 const real cellsizeLimit,
2259 const gmx_ddbox_t& ddbox)
2261 if (options.numCells[XX] <= 0 && (ddGridSetup.numDomains[XX] == 0))
2264 gmx_bool bC = (systemInfo.haveSplitConstraints
2265 && systemInfo.constraintCommunicationRange > systemInfo.minCutoffForMultiBody);
2266 sprintf(buf, "Change the number of ranks or mdrun option %s%s%s", !bC ? "-rdd" : "-rcon",
2267 ddSettings.initialDlbState != DlbState::offUser ? " or -dds" : "",
2268 bC ? " or your LINCS settings" : "");
2270 gmx_fatal_collective(FARGS, cr->mpi_comm_mysim, MASTER(cr),
2271 "There is no domain decomposition for %d ranks that is compatible "
2272 "with the given box and a minimum cell size of %g nm\n"
2274 "Look in the log file for details on the domain decomposition",
2275 cr->nnodes - ddGridSetup.numPmeOnlyRanks, cellsizeLimit, buf);
2278 const real acs = average_cellsize_min(ddbox, ddGridSetup.numDomains);
2279 if (acs < cellsizeLimit)
2281 if (options.numCells[XX] <= 0)
2285 "dd_choose_grid() should return a grid that satisfies the cell size limits");
2289 gmx_fatal_collective(
2290 FARGS, cr->mpi_comm_mysim, MASTER(cr),
2291 "The initial cell size (%f) is smaller than the cell size limit (%f), change "
2292 "options -dd, -rdd or -rcon, see the log file for details",
2293 acs, cellsizeLimit);
2297 const int numPPRanks =
2298 ddGridSetup.numDomains[XX] * ddGridSetup.numDomains[YY] * ddGridSetup.numDomains[ZZ];
2299 if (cr->nnodes - numPPRanks != ddGridSetup.numPmeOnlyRanks)
2301 gmx_fatal_collective(FARGS, cr->mpi_comm_mysim, MASTER(cr),
2302 "The size of the domain decomposition grid (%d) does not match the "
2303 "number of PP ranks (%d). The total number of ranks is %d",
2304 numPPRanks, cr->nnodes - ddGridSetup.numPmeOnlyRanks, cr->nnodes);
2306 if (ddGridSetup.numPmeOnlyRanks > numPPRanks)
2308 gmx_fatal_collective(FARGS, cr->mpi_comm_mysim, MASTER(cr),
2309 "The number of separate PME ranks (%d) is larger than the number of "
2310 "PP ranks (%d), this is not supported.",
2311 ddGridSetup.numPmeOnlyRanks, numPPRanks);
2315 /*! \brief Set the cell size and interaction limits, as well as the DD grid */
2316 static DDRankSetup getDDRankSetup(const gmx::MDLogger& mdlog,
2318 const DDGridSetup& ddGridSetup,
2319 const t_inputrec& ir)
2322 .appendTextFormatted("Domain decomposition grid %d x %d x %d, separate PME ranks %d",
2323 ddGridSetup.numDomains[XX], ddGridSetup.numDomains[YY],
2324 ddGridSetup.numDomains[ZZ], ddGridSetup.numPmeOnlyRanks);
2326 DDRankSetup ddRankSetup;
2328 ddRankSetup.numPPRanks = cr->nnodes - ddGridSetup.numPmeOnlyRanks;
2329 copy_ivec(ddGridSetup.numDomains, ddRankSetup.numPPCells);
2331 ddRankSetup.usePmeOnlyRanks = (ddGridSetup.numPmeOnlyRanks > 0);
2332 if (ddRankSetup.usePmeOnlyRanks)
2334 ddRankSetup.numRanksDoingPme = ddGridSetup.numPmeOnlyRanks;
2338 ddRankSetup.numRanksDoingPme =
2339 ddGridSetup.numDomains[XX] * ddGridSetup.numDomains[YY] * ddGridSetup.numDomains[ZZ];
2342 if (EEL_PME(ir.coulombtype) || EVDW_PME(ir.vdwtype))
2344 /* The following choices should match those
2345 * in comm_cost_est in domdec_setup.c.
2346 * Note that here the checks have to take into account
2347 * that the decomposition might occur in a different order than xyz
2348 * (for instance through the env.var. GMX_DD_ORDER_ZYX),
2349 * in which case they will not match those in comm_cost_est,
2350 * but since that is mainly for testing purposes that's fine.
2352 if (ddGridSetup.numDDDimensions >= 2 && ddGridSetup.ddDimensions[0] == XX
2353 && ddGridSetup.ddDimensions[1] == YY
2354 && ddRankSetup.numRanksDoingPme > ddGridSetup.numDomains[XX]
2355 && ddRankSetup.numRanksDoingPme % ddGridSetup.numDomains[XX] == 0
2356 && getenv("GMX_PMEONEDD") == nullptr)
2358 ddRankSetup.npmedecompdim = 2;
2359 ddRankSetup.npmenodes_x = ddGridSetup.numDomains[XX];
2360 ddRankSetup.npmenodes_y = ddRankSetup.numRanksDoingPme / ddRankSetup.npmenodes_x;
2364 /* In case nc is 1 in both x and y we could still choose to
2365 * decompose pme in y instead of x, but we use x for simplicity.
2367 ddRankSetup.npmedecompdim = 1;
2368 if (ddGridSetup.ddDimensions[0] == YY)
2370 ddRankSetup.npmenodes_x = 1;
2371 ddRankSetup.npmenodes_y = ddRankSetup.numRanksDoingPme;
2375 ddRankSetup.npmenodes_x = ddRankSetup.numRanksDoingPme;
2376 ddRankSetup.npmenodes_y = 1;
2380 .appendTextFormatted("PME domain decomposition: %d x %d x %d",
2381 ddRankSetup.npmenodes_x, ddRankSetup.npmenodes_y, 1);
2385 ddRankSetup.npmedecompdim = 0;
2386 ddRankSetup.npmenodes_x = 0;
2387 ddRankSetup.npmenodes_y = 0;
2393 /*! \brief Set the cell size and interaction limits */
2394 static void set_dd_limits(const gmx::MDLogger& mdlog,
2397 const DomdecOptions& options,
2398 const DDSettings& ddSettings,
2399 const DDSystemInfo& systemInfo,
2400 const DDGridSetup& ddGridSetup,
2401 const int numPPRanks,
2402 const gmx_mtop_t* mtop,
2403 const t_inputrec* ir,
2404 const gmx_ddbox_t& ddbox)
2406 gmx_domdec_comm_t* comm = dd->comm;
2407 comm->ddSettings = ddSettings;
2409 /* Initialize to GPU share count to 0, might change later */
2410 comm->nrank_gpu_shared = 0;
2412 comm->dlbState = comm->ddSettings.initialDlbState;
2413 dd_dlb_set_should_check_whether_to_turn_dlb_on(dd, TRUE);
2414 /* To consider turning DLB on after 2*nstlist steps we need to check
2415 * at partitioning count 3. Thus we need to increase the first count by 2.
2417 comm->ddPartioningCountFirstDlbOff += 2;
2419 comm->bPMELoadBalDLBLimits = FALSE;
2421 /* Allocate the charge group/atom sorting struct */
2422 comm->sort = std::make_unique<gmx_domdec_sort_t>();
2424 comm->systemInfo = systemInfo;
2426 if (systemInfo.useUpdateGroups)
2428 /* Note: We would like to use dd->nnodes for the atom count estimate,
2429 * but that is not yet available here. But this anyhow only
2430 * affect performance up to the second dd_partition_system call.
2432 const int homeAtomCountEstimate = mtop->natoms / numPPRanks;
2433 comm->updateGroupsCog = std::make_unique<gmx::UpdateGroupsCog>(
2434 *mtop, systemInfo.updateGroupingPerMoleculetype, maxReferenceTemperature(*ir),
2435 homeAtomCountEstimate);
2438 /* Set the DD setup given by ddGridSetup */
2439 copy_ivec(ddGridSetup.numDomains, dd->nc);
2440 dd->ndim = ddGridSetup.numDDDimensions;
2441 copy_ivec(ddGridSetup.ddDimensions, dd->dim);
2443 dd->nnodes = dd->nc[XX] * dd->nc[YY] * dd->nc[ZZ];
2445 snew(comm->slb_frac, DIM);
2446 if (isDlbDisabled(comm))
2448 comm->slb_frac[XX] = get_slb_frac(mdlog, "x", dd->nc[XX], options.cellSizeX);
2449 comm->slb_frac[YY] = get_slb_frac(mdlog, "y", dd->nc[YY], options.cellSizeY);
2450 comm->slb_frac[ZZ] = get_slb_frac(mdlog, "z", dd->nc[ZZ], options.cellSizeZ);
2453 /* Set the multi-body cut-off and cellsize limit for DLB */
2454 comm->cutoff_mbody = systemInfo.minCutoffForMultiBody;
2455 comm->cellsize_limit = systemInfo.cellsizeLimit;
2456 if (systemInfo.haveInterDomainBondeds && systemInfo.increaseMultiBodyCutoff)
2458 if (systemInfo.filterBondedCommunication || !isDlbDisabled(comm))
2460 /* Set the bonded communication distance to halfway
2461 * the minimum and the maximum,
2462 * since the extra communication cost is nearly zero.
2464 real acs = average_cellsize_min(ddbox, dd->nc);
2465 comm->cutoff_mbody = 0.5 * (systemInfo.minCutoffForMultiBody + acs);
2466 if (!isDlbDisabled(comm))
2468 /* Check if this does not limit the scaling */
2469 comm->cutoff_mbody = std::min(comm->cutoff_mbody, options.dlbScaling * acs);
2471 if (!systemInfo.filterBondedCommunication)
2473 /* Without bBondComm do not go beyond the n.b. cut-off */
2474 comm->cutoff_mbody = std::min(comm->cutoff_mbody, systemInfo.cutoff);
2475 if (comm->cellsize_limit >= systemInfo.cutoff)
2477 /* We don't loose a lot of efficieny
2478 * when increasing it to the n.b. cut-off.
2479 * It can even be slightly faster, because we need
2480 * less checks for the communication setup.
2482 comm->cutoff_mbody = systemInfo.cutoff;
2485 /* Check if we did not end up below our original limit */
2486 comm->cutoff_mbody = std::max(comm->cutoff_mbody, systemInfo.minCutoffForMultiBody);
2488 if (comm->cutoff_mbody > comm->cellsize_limit)
2490 comm->cellsize_limit = comm->cutoff_mbody;
2493 /* Without DLB and cutoff_mbody<cutoff, cutoff_mbody is dynamic */
2499 "Bonded atom communication beyond the cut-off: %s\n"
2500 "cellsize limit %f\n",
2501 gmx::boolToString(systemInfo.filterBondedCommunication), comm->cellsize_limit);
2506 check_dd_restrictions(dd, ir, mdlog);
2510 void dd_init_bondeds(FILE* fplog,
2512 const gmx_mtop_t* mtop,
2513 const gmx_vsite_t* vsite,
2514 const t_inputrec* ir,
2516 cginfo_mb_t* cginfo_mb)
2518 gmx_domdec_comm_t* comm;
2520 dd_make_reverse_top(fplog, dd, mtop, vsite, ir, bBCheck);
2524 if (comm->systemInfo.filterBondedCommunication)
2526 /* Communicate atoms beyond the cut-off for bonded interactions */
2527 comm->bondedLinks = makeBondedLinks(mtop, cginfo_mb);
2531 /* Only communicate atoms based on cut-off */
2532 comm->bondedLinks = nullptr;
2536 static void writeSettings(gmx::TextWriter* log,
2538 const gmx_mtop_t* mtop,
2539 const t_inputrec* ir,
2540 gmx_bool bDynLoadBal,
2542 const gmx_ddbox_t* ddbox)
2544 gmx_domdec_comm_t* comm;
2553 log->writeString("The maximum number of communication pulses is:");
2554 for (d = 0; d < dd->ndim; d++)
2556 log->writeStringFormatted(" %c %d", dim2char(dd->dim[d]), comm->cd[d].np_dlb);
2558 log->ensureLineBreak();
2559 log->writeLineFormatted("The minimum size for domain decomposition cells is %.3f nm",
2560 comm->cellsize_limit);
2561 log->writeLineFormatted("The requested allowed shrink of DD cells (option -dds) is: %.2f", dlb_scale);
2562 log->writeString("The allowed shrink of domain decomposition cells is:");
2563 for (d = 0; d < DIM; d++)
2567 if (d >= ddbox->npbcdim && dd->nc[d] == 2)
2573 shrink = comm->cellsize_min_dlb[d]
2574 / (ddbox->box_size[d] * ddbox->skew_fac[d] / dd->nc[d]);
2576 log->writeStringFormatted(" %c %.2f", dim2char(d), shrink);
2579 log->ensureLineBreak();
2583 set_dd_cell_sizes_slb(dd, ddbox, setcellsizeslbPULSE_ONLY, np);
2584 log->writeString("The initial number of communication pulses is:");
2585 for (d = 0; d < dd->ndim; d++)
2587 log->writeStringFormatted(" %c %d", dim2char(dd->dim[d]), np[dd->dim[d]]);
2589 log->ensureLineBreak();
2590 log->writeString("The initial domain decomposition cell size is:");
2591 for (d = 0; d < DIM; d++)
2595 log->writeStringFormatted(" %c %.2f nm", dim2char(d), dd->comm->cellsize_min[d]);
2598 log->ensureLineBreak();
2602 const bool haveInterDomainVsites =
2603 (countInterUpdategroupVsites(*mtop, comm->systemInfo.updateGroupingPerMoleculetype) != 0);
2605 if (comm->systemInfo.haveInterDomainBondeds || haveInterDomainVsites
2606 || comm->systemInfo.haveSplitConstraints || comm->systemInfo.haveSplitSettles)
2608 std::string decompUnits;
2609 if (comm->systemInfo.useUpdateGroups)
2611 decompUnits = "atom groups";
2615 decompUnits = "atoms";
2618 log->writeLineFormatted("The maximum allowed distance for %s involved in interactions is:",
2619 decompUnits.c_str());
2620 log->writeLineFormatted("%40s %-7s %6.3f nm", "non-bonded interactions", "",
2621 comm->systemInfo.cutoff);
2625 limit = dd->comm->cellsize_limit;
2629 if (dd->unitCellInfo.ddBoxIsDynamic)
2632 "(the following are initial values, they could change due to box "
2635 limit = dd->comm->cellsize_min[XX];
2636 for (d = 1; d < DIM; d++)
2638 limit = std::min(limit, dd->comm->cellsize_min[d]);
2642 if (comm->systemInfo.haveInterDomainBondeds)
2644 log->writeLineFormatted("%40s %-7s %6.3f nm", "two-body bonded interactions", "(-rdd)",
2645 std::max(comm->systemInfo.cutoff, comm->cutoff_mbody));
2646 log->writeLineFormatted("%40s %-7s %6.3f nm", "multi-body bonded interactions",
2648 (comm->systemInfo.filterBondedCommunication || isDlbOn(dd->comm))
2649 ? comm->cutoff_mbody
2650 : std::min(comm->systemInfo.cutoff, limit));
2652 if (haveInterDomainVsites)
2654 log->writeLineFormatted("%40s %-7s %6.3f nm", "virtual site constructions", "(-rcon)", limit);
2656 if (comm->systemInfo.haveSplitConstraints || comm->systemInfo.haveSplitSettles)
2658 std::string separation =
2659 gmx::formatString("atoms separated by up to %d constraints", 1 + ir->nProjOrder);
2660 log->writeLineFormatted("%40s %-7s %6.3f nm\n", separation.c_str(), "(-rcon)", limit);
2662 log->ensureLineBreak();
2666 static void logSettings(const gmx::MDLogger& mdlog,
2668 const gmx_mtop_t* mtop,
2669 const t_inputrec* ir,
2671 const gmx_ddbox_t* ddbox)
2673 gmx::StringOutputStream stream;
2674 gmx::TextWriter log(&stream);
2675 writeSettings(&log, dd, mtop, ir, isDlbOn(dd->comm), dlb_scale, ddbox);
2676 if (dd->comm->dlbState == DlbState::offCanTurnOn)
2679 log.ensureEmptyLine();
2681 "When dynamic load balancing gets turned on, these settings will change to:");
2683 writeSettings(&log, dd, mtop, ir, true, dlb_scale, ddbox);
2685 GMX_LOG(mdlog.info).asParagraph().appendText(stream.toString());
2688 static void set_cell_limits_dlb(const gmx::MDLogger& mdlog,
2691 const t_inputrec* ir,
2692 const gmx_ddbox_t* ddbox)
2694 gmx_domdec_comm_t* comm;
2695 int d, dim, npulse, npulse_d_max, npulse_d;
2700 bNoCutOff = (ir->rvdw == 0 || ir->rcoulomb == 0);
2702 /* Determine the maximum number of comm. pulses in one dimension */
2704 comm->cellsize_limit = std::max(comm->cellsize_limit, comm->cutoff_mbody);
2706 /* Determine the maximum required number of grid pulses */
2707 if (comm->cellsize_limit >= comm->systemInfo.cutoff)
2709 /* Only a single pulse is required */
2712 else if (!bNoCutOff && comm->cellsize_limit > 0)
2714 /* We round down slightly here to avoid overhead due to the latency
2715 * of extra communication calls when the cut-off
2716 * would be only slightly longer than the cell size.
2717 * Later cellsize_limit is redetermined,
2718 * so we can not miss interactions due to this rounding.
2720 npulse = static_cast<int>(0.96 + comm->systemInfo.cutoff / comm->cellsize_limit);
2724 /* There is no cell size limit */
2725 npulse = std::max(dd->nc[XX] - 1, std::max(dd->nc[YY] - 1, dd->nc[ZZ] - 1));
2728 if (!bNoCutOff && npulse > 1)
2730 /* See if we can do with less pulses, based on dlb_scale */
2732 for (d = 0; d < dd->ndim; d++)
2735 npulse_d = static_cast<int>(
2737 + dd->nc[dim] * comm->systemInfo.cutoff
2738 / (ddbox->box_size[dim] * ddbox->skew_fac[dim] * dlb_scale));
2739 npulse_d_max = std::max(npulse_d_max, npulse_d);
2741 npulse = std::min(npulse, npulse_d_max);
2744 /* This env var can override npulse */
2745 d = dd_getenv(mdlog, "GMX_DD_NPULSE", 0);
2752 comm->bVacDLBNoLimit = (ir->ePBC == epbcNONE);
2753 for (d = 0; d < dd->ndim; d++)
2755 if (comm->ddSettings.request1DAnd1Pulse)
2757 comm->cd[d].np_dlb = 1;
2761 comm->cd[d].np_dlb = std::min(npulse, dd->nc[dd->dim[d]] - 1);
2762 comm->maxpulse = std::max(comm->maxpulse, comm->cd[d].np_dlb);
2764 if (comm->cd[d].np_dlb < dd->nc[dd->dim[d]] - 1)
2766 comm->bVacDLBNoLimit = FALSE;
2770 /* cellsize_limit is set for LINCS in init_domain_decomposition */
2771 if (!comm->bVacDLBNoLimit)
2773 comm->cellsize_limit = std::max(comm->cellsize_limit, comm->systemInfo.cutoff / comm->maxpulse);
2775 comm->cellsize_limit = std::max(comm->cellsize_limit, comm->cutoff_mbody);
2776 /* Set the minimum cell size for each DD dimension */
2777 for (d = 0; d < dd->ndim; d++)
2779 if (comm->bVacDLBNoLimit || comm->cd[d].np_dlb * comm->cellsize_limit >= comm->systemInfo.cutoff)
2781 comm->cellsize_min_dlb[dd->dim[d]] = comm->cellsize_limit;
2785 comm->cellsize_min_dlb[dd->dim[d]] = comm->systemInfo.cutoff / comm->cd[d].np_dlb;
2788 if (comm->cutoff_mbody <= 0)
2790 comm->cutoff_mbody = std::min(comm->systemInfo.cutoff, comm->cellsize_limit);
2798 bool dd_moleculesAreAlwaysWhole(const gmx_domdec_t& dd)
2800 return dd.comm->systemInfo.moleculesAreAlwaysWhole;
2803 gmx_bool dd_bonded_molpbc(const gmx_domdec_t* dd, int ePBC)
2805 /* If each molecule is a single charge group
2806 * or we use domain decomposition for each periodic dimension,
2807 * we do not need to take pbc into account for the bonded interactions.
2809 return (ePBC != epbcNONE && dd->comm->systemInfo.haveInterDomainBondeds
2810 && !(dd->nc[XX] > 1 && dd->nc[YY] > 1 && (dd->nc[ZZ] > 1 || ePBC == epbcXY)));
2813 /*! \brief Sets grid size limits and PP-PME setup, prints settings to log */
2814 static void set_ddgrid_parameters(const gmx::MDLogger& mdlog,
2817 const gmx_mtop_t* mtop,
2818 const t_inputrec* ir,
2819 const gmx_ddbox_t* ddbox)
2821 gmx_domdec_comm_t* comm = dd->comm;
2822 DDRankSetup& ddRankSetup = comm->ddRankSetup;
2824 if (EEL_PME(ir->coulombtype) || EVDW_PME(ir->vdwtype))
2826 init_ddpme(dd, &ddRankSetup.ddpme[0], 0);
2827 if (ddRankSetup.npmedecompdim >= 2)
2829 init_ddpme(dd, &ddRankSetup.ddpme[1], 1);
2834 ddRankSetup.numRanksDoingPme = 0;
2835 if (dd->pme_nodeid >= 0)
2837 gmx_fatal_collective(FARGS, dd->mpi_comm_all, DDMASTER(dd),
2838 "Can not have separate PME ranks without PME electrostatics");
2844 fprintf(debug, "The DD cut-off is %f\n", comm->systemInfo.cutoff);
2846 if (!isDlbDisabled(comm))
2848 set_cell_limits_dlb(mdlog, dd, dlb_scale, ir, ddbox);
2851 logSettings(mdlog, dd, mtop, ir, dlb_scale, ddbox);
2854 if (ir->ePBC == epbcNONE)
2856 vol_frac = 1 - 1 / static_cast<double>(dd->nnodes);
2860 vol_frac = (1 + comm_box_frac(dd->nc, comm->systemInfo.cutoff, *ddbox))
2861 / static_cast<double>(dd->nnodes);
2865 fprintf(debug, "Volume fraction for all DD zones: %f\n", vol_frac);
2867 int natoms_tot = mtop->natoms;
2869 dd->ga2la = new gmx_ga2la_t(natoms_tot, static_cast<int>(vol_frac * natoms_tot));
2872 /*! \brief Get some important DD parameters which can be modified by env.vars */
2873 static DDSettings getDDSettings(const gmx::MDLogger& mdlog,
2874 const DomdecOptions& options,
2875 const gmx::MdrunOptions& mdrunOptions,
2876 const t_inputrec& ir)
2878 DDSettings ddSettings;
2880 ddSettings.useSendRecv2 = (dd_getenv(mdlog, "GMX_DD_USE_SENDRECV2", 0) != 0);
2881 ddSettings.dlb_scale_lim = dd_getenv(mdlog, "GMX_DLB_MAX_BOX_SCALING", 10);
2882 ddSettings.request1DAnd1Pulse = bool(dd_getenv(mdlog, "GMX_DD_1D_1PULSE", 0));
2883 ddSettings.useDDOrderZYX = bool(dd_getenv(mdlog, "GMX_DD_ORDER_ZYX", 0));
2884 ddSettings.useCartesianReorder = bool(dd_getenv(mdlog, "GMX_NO_CART_REORDER", 1));
2885 ddSettings.eFlop = dd_getenv(mdlog, "GMX_DLB_BASED_ON_FLOPS", 0);
2886 const int recload = dd_getenv(mdlog, "GMX_DD_RECORD_LOAD", 1);
2887 ddSettings.nstDDDump = dd_getenv(mdlog, "GMX_DD_NST_DUMP", 0);
2888 ddSettings.nstDDDumpGrid = dd_getenv(mdlog, "GMX_DD_NST_DUMP_GRID", 0);
2889 ddSettings.DD_debug = dd_getenv(mdlog, "GMX_DD_DEBUG", 0);
2891 if (ddSettings.useSendRecv2)
2895 "Will use two sequential MPI_Sendrecv calls instead of two simultaneous "
2896 "non-blocking MPI_Irecv and MPI_Isend pairs for constraint and vsite "
2900 if (ddSettings.eFlop)
2902 GMX_LOG(mdlog.info).appendText("Will load balance based on FLOP count");
2903 ddSettings.recordLoad = true;
2907 ddSettings.recordLoad = (wallcycle_have_counter() && recload > 0);
2910 ddSettings.initialDlbState = determineInitialDlbState(mdlog, options.dlbOption,
2911 ddSettings.recordLoad, mdrunOptions, &ir);
2913 .appendTextFormatted("Dynamic load balancing: %s",
2914 edlbs_names[static_cast<int>(ddSettings.initialDlbState)]);
2919 gmx_domdec_t::gmx_domdec_t(const t_inputrec& ir) : unitCellInfo(ir) {}
2921 /*! \brief Return whether the simulation described can run a 1D single-pulse DD.
2923 * The GPU halo exchange code requires a 1D single-pulse DD. Such a DD
2924 * generally requires a larger box than other possible decompositions
2925 * with the same rank count, so the calling code might need to decide
2926 * what is the most appropriate way to run the simulation based on
2927 * whether such a DD is possible.
2929 * This function works like init_domain_decomposition(), but will not
2930 * give a fatal error, and only uses \c cr for communicating between
2933 * It is safe to call before thread-MPI spawns ranks, so that
2934 * thread-MPI can decide whether and how to trigger the GPU halo
2935 * exchange code path. The number of PME ranks, if any, should be set
2936 * in \c options.numPmeRanks.
2938 static bool canMake1DAnd1PulseDomainDecomposition(const DDSettings& ddSettingsOriginal,
2939 const t_commrec* cr,
2940 const int numRanksRequested,
2941 const DomdecOptions& options,
2942 const gmx_mtop_t& mtop,
2943 const t_inputrec& ir,
2945 gmx::ArrayRef<const gmx::RVec> xGlobal)
2947 // Ensure we don't write any output from this checking routine
2948 gmx::MDLogger dummyLogger;
2950 DDSystemInfo systemInfo = getSystemInfo(dummyLogger, cr, options, mtop, ir, box, xGlobal);
2952 DDSettings ddSettings = ddSettingsOriginal;
2953 ddSettings.request1DAnd1Pulse = true;
2954 const real gridSetupCellsizeLimit = getDDGridSetupCellSizeLimit(
2955 dummyLogger, ddSettings.request1DAnd1Pulse, !isDlbDisabled(ddSettings.initialDlbState),
2956 options.dlbScaling, ir, systemInfo.cellsizeLimit);
2957 gmx_ddbox_t ddbox = { 0 };
2958 DDGridSetup ddGridSetup =
2959 getDDGridSetup(dummyLogger, cr, numRanksRequested, options, ddSettings, systemInfo,
2960 gridSetupCellsizeLimit, mtop, ir, box, xGlobal, &ddbox);
2962 const bool canMakeDDWith1DAnd1Pulse = (ddGridSetup.numDomains[XX] != 0);
2964 return canMakeDDWith1DAnd1Pulse;
2967 bool is1DAnd1PulseDD(const gmx_domdec_t& dd)
2969 const int maxDimensionSize = std::max(dd.nc[XX], std::max(dd.nc[YY], dd.nc[ZZ]));
2970 const int productOfDimensionSizes = dd.nc[XX] * dd.nc[YY] * dd.nc[ZZ];
2971 const bool decompositionHasOneDimension = (maxDimensionSize == productOfDimensionSizes);
2973 const bool hasMax1Pulse =
2974 ((isDlbDisabled(dd.comm) && dd.comm->cellsize_limit >= dd.comm->systemInfo.cutoff)
2975 || (!isDlbDisabled(dd.comm) && dd.comm->maxpulse == 1));
2977 return decompositionHasOneDimension && hasMax1Pulse;
2983 // TODO once the functionality stablizes, move this class and
2984 // supporting functionality into builder.cpp
2985 /*! \brief Impl class for DD builder */
2986 class DomainDecompositionBuilder::Impl
2990 Impl(const MDLogger& mdlog,
2992 const DomdecOptions& options,
2993 const MdrunOptions& mdrunOptions,
2994 bool prefer1DAnd1Pulse,
2995 const gmx_mtop_t& mtop,
2996 const t_inputrec& ir,
2998 ArrayRef<const RVec> xGlobal);
3000 //! Build the resulting DD manager
3001 gmx_domdec_t* build(LocalAtomSetManager* atomSets);
3003 //! Objects used in constructing and configuring DD
3006 const MDLogger& mdlog_;
3007 //! Communication object
3009 //! User-supplied options configuring DD behavior
3010 const DomdecOptions options_;
3011 //! Global system topology
3012 const gmx_mtop_t& mtop_;
3013 //! User input values from the tpr file
3014 const t_inputrec& ir_;
3017 //! Internal objects used in constructing DD
3019 //! Settings combined from the user input
3020 DDSettings ddSettings_;
3021 //! Information derived from the simulation system
3022 DDSystemInfo systemInfo_;
3024 gmx_ddbox_t ddbox_ = { 0 };
3025 //! Organization of the DD grids
3026 DDGridSetup ddGridSetup_;
3027 //! Organzation of the DD ranks
3028 DDRankSetup ddRankSetup_;
3029 //! Number of DD cells in each dimension
3030 ivec ddCellIndex_ = { 0, 0, 0 };
3031 //! IDs of PME-only ranks
3032 std::vector<int> pmeRanks_;
3033 //! Contains a valid Cartesian-communicator-based setup, or defaults.
3034 CartesianRankSetup cartSetup_;
3038 DomainDecompositionBuilder::Impl::Impl(const MDLogger& mdlog,
3040 const DomdecOptions& options,
3041 const MdrunOptions& mdrunOptions,
3042 const bool prefer1DAnd1Pulse,
3043 const gmx_mtop_t& mtop,
3044 const t_inputrec& ir,
3046 ArrayRef<const RVec> xGlobal) :
3053 GMX_LOG(mdlog_.info).appendTextFormatted("\nInitializing Domain Decomposition on %d ranks", cr_->nnodes);
3055 ddSettings_ = getDDSettings(mdlog_, options_, mdrunOptions, ir_);
3057 if (prefer1DAnd1Pulse
3058 && canMake1DAnd1PulseDomainDecomposition(ddSettings_, cr_, cr_->nnodes, options_, mtop_,
3061 ddSettings_.request1DAnd1Pulse = true;
3064 if (ddSettings_.eFlop > 1)
3066 /* Ensure that we have different random flop counts on different ranks */
3067 srand(1 + cr_->nodeid);
3070 systemInfo_ = getSystemInfo(mdlog_, cr_, options_, mtop_, ir_, box, xGlobal);
3072 const int numRanksRequested = cr_->nnodes;
3073 checkForValidRankCountRequests(numRanksRequested, EEL_PME(ir_.coulombtype), options_.numPmeRanks);
3075 // DD grid setup uses a more different cell size limit for
3076 // automated setup than the one in systemInfo_. The latter is used
3077 // in set_dd_limits() to configure DLB, for example.
3078 const real gridSetupCellsizeLimit = getDDGridSetupCellSizeLimit(
3079 mdlog_, ddSettings_.request1DAnd1Pulse, !isDlbDisabled(ddSettings_.initialDlbState),
3080 options_.dlbScaling, ir_, systemInfo_.cellsizeLimit);
3081 ddGridSetup_ = getDDGridSetup(mdlog_, cr_, numRanksRequested, options_, ddSettings_, systemInfo_,
3082 gridSetupCellsizeLimit, mtop_, ir_, box, xGlobal, &ddbox_);
3083 checkDDGridSetup(ddGridSetup_, cr_, options_, ddSettings_, systemInfo_, gridSetupCellsizeLimit, ddbox_);
3085 cr_->npmenodes = ddGridSetup_.numPmeOnlyRanks;
3087 ddRankSetup_ = getDDRankSetup(mdlog_, cr_, ddGridSetup_, ir_);
3089 /* Generate the group communicator, also decides the duty of each rank */
3090 cartSetup_ = makeGroupCommunicators(mdlog_, ddSettings_, options_.rankOrder, ddRankSetup_, cr_,
3091 ddCellIndex_, &pmeRanks_);
3094 gmx_domdec_t* DomainDecompositionBuilder::Impl::build(LocalAtomSetManager* atomSets)
3096 gmx_domdec_t* dd = new gmx_domdec_t(ir_);
3098 copy_ivec(ddCellIndex_, dd->ci);
3100 dd->comm = init_dd_comm();
3102 dd->comm->ddRankSetup = ddRankSetup_;
3103 dd->comm->cartesianRankSetup = cartSetup_;
3105 set_dd_limits(mdlog_, cr_, dd, options_, ddSettings_, systemInfo_, ddGridSetup_,
3106 ddRankSetup_.numPPRanks, &mtop_, &ir_, ddbox_);
3108 setupGroupCommunication(mdlog_, ddSettings_, pmeRanks_, cr_, mtop_.natoms, dd);
3110 if (thisRankHasDuty(cr_, DUTY_PP))
3112 set_ddgrid_parameters(mdlog_, dd, options_.dlbScaling, &mtop_, &ir_, &ddbox_);
3114 setup_neighbor_relations(dd);
3117 /* Set overallocation to avoid frequent reallocation of arrays */
3118 set_over_alloc_dd(TRUE);
3120 dd->atomSets = atomSets;
3125 DomainDecompositionBuilder::DomainDecompositionBuilder(const MDLogger& mdlog,
3127 const DomdecOptions& options,
3128 const MdrunOptions& mdrunOptions,
3129 const bool prefer1DAnd1Pulse,
3130 const gmx_mtop_t& mtop,
3131 const t_inputrec& ir,
3133 ArrayRef<const RVec> xGlobal) :
3134 impl_(new Impl(mdlog, cr, options, mdrunOptions, prefer1DAnd1Pulse, mtop, ir, box, xGlobal))
3138 gmx_domdec_t* DomainDecompositionBuilder::build(LocalAtomSetManager* atomSets)
3140 return impl_->build(atomSets);
3143 DomainDecompositionBuilder::~DomainDecompositionBuilder() = default;
3147 static gmx_bool test_dd_cutoff(t_commrec* cr, const matrix box, gmx::ArrayRef<const gmx::RVec> x, real cutoffRequested)
3157 set_ddbox(*dd, false, box, true, x, &ddbox);
3161 for (d = 0; d < dd->ndim; d++)
3165 inv_cell_size = DD_CELL_MARGIN * dd->nc[dim] / ddbox.box_size[dim];
3166 if (dd->unitCellInfo.ddBoxIsDynamic)
3168 inv_cell_size *= DD_PRES_SCALE_MARGIN;
3171 np = 1 + static_cast<int>(cutoffRequested * inv_cell_size * ddbox.skew_fac[dim]);
3173 if (dd->comm->ddSettings.request1DAnd1Pulse && np > 1)
3178 if (!isDlbDisabled(dd->comm) && (dim < ddbox.npbcdim) && (dd->comm->cd[d].np_dlb > 0))
3180 if (np > dd->comm->cd[d].np_dlb)
3185 /* If a current local cell size is smaller than the requested
3186 * cut-off, we could still fix it, but this gets very complicated.
3187 * Without fixing here, we might actually need more checks.
3189 real cellSizeAlongDim =
3190 (dd->comm->cell_x1[dim] - dd->comm->cell_x0[dim]) * ddbox.skew_fac[dim];
3191 if (cellSizeAlongDim * dd->comm->cd[d].np_dlb < cutoffRequested)
3198 if (!isDlbDisabled(dd->comm))
3200 /* If DLB is not active yet, we don't need to check the grid jumps.
3201 * Actually we shouldn't, because then the grid jump data is not set.
3203 if (isDlbOn(dd->comm) && check_grid_jump(0, dd, cutoffRequested, &ddbox, FALSE))
3208 gmx_sumi(1, &LocallyLimited, cr);
3210 if (LocallyLimited > 0)
3219 gmx_bool change_dd_cutoff(t_commrec* cr, const matrix box, gmx::ArrayRef<const gmx::RVec> x, real cutoffRequested)
3221 gmx_bool bCutoffAllowed;
3223 bCutoffAllowed = test_dd_cutoff(cr, box, x, cutoffRequested);
3227 cr->dd->comm->systemInfo.cutoff = cutoffRequested;
3230 return bCutoffAllowed;