| 1 |
#ifdef IS_MPI |
| 2 |
#include <iostream> |
| 3 |
#include <stdlib.h> |
| 4 |
#include <string.h> |
| 5 |
#include <math.h> |
| 6 |
#include <mpi.h> |
| 7 |
|
| 8 |
#include "brains/mpiSimulation.hpp" |
| 9 |
#include "utils/simError.h" |
| 10 |
#include "UseTheForce/fortranWrappers.hpp" |
| 11 |
#include "math/randomSPRNG.hpp" |
| 12 |
|
| 13 |
mpiSimulation* mpiSim; |
| 14 |
|
| 15 |
mpiSimulation::mpiSimulation(SimInfo* the_entryPlug) |
| 16 |
{ |
| 17 |
entryPlug = the_entryPlug; |
| 18 |
parallelData = new mpiSimData; |
| 19 |
|
| 20 |
MPI_Comm_size(MPI_COMM_WORLD, &(parallelData->nProcessors) ); |
| 21 |
parallelData->myNode = worldRank; |
| 22 |
|
| 23 |
MolToProcMap = new int[entryPlug->n_mol]; |
| 24 |
MolComponentType = new int[entryPlug->n_mol]; |
| 25 |
AtomToProcMap = new int[entryPlug->n_atoms]; |
| 26 |
GroupToProcMap = new int[entryPlug->ngroup]; |
| 27 |
|
| 28 |
mpiSim = this; |
| 29 |
wrapMeSimParallel( this ); |
| 30 |
} |
| 31 |
|
| 32 |
|
| 33 |
mpiSimulation::~mpiSimulation(){ |
| 34 |
|
| 35 |
delete[] MolToProcMap; |
| 36 |
delete[] MolComponentType; |
| 37 |
delete[] AtomToProcMap; |
| 38 |
delete[] GroupToProcMap; |
| 39 |
|
| 40 |
delete parallelData; |
| 41 |
// perhaps we should let fortran know the party is over. |
| 42 |
|
| 43 |
} |
| 44 |
|
| 45 |
void mpiSimulation::divideLabor( ){ |
| 46 |
|
| 47 |
int nComponents; |
| 48 |
MoleculeStamp** compStamps; |
| 49 |
randomSPRNG *myRandom; |
| 50 |
int* componentsNmol; |
| 51 |
int* AtomsPerProc; |
| 52 |
int* GroupsPerProc; |
| 53 |
|
| 54 |
double numerator; |
| 55 |
double denominator; |
| 56 |
double precast; |
| 57 |
double x, y, a; |
| 58 |
int old_atoms, add_atoms, new_atoms; |
| 59 |
int old_groups, add_groups, new_groups; |
| 60 |
|
| 61 |
int nTarget; |
| 62 |
int molIndex, atomIndex, groupIndex; |
| 63 |
int done; |
| 64 |
int i, j, loops, which_proc; |
| 65 |
int nmol_global, nmol_local; |
| 66 |
int ngroups_global, ngroups_local; |
| 67 |
int natoms_global, natoms_local; |
| 68 |
int ncutoff_groups, nAtomsInGroups; |
| 69 |
int local_index; |
| 70 |
int baseSeed = entryPlug->getSeed(); |
| 71 |
CutoffGroupStamp* cg; |
| 72 |
|
| 73 |
nComponents = entryPlug->nComponents; |
| 74 |
compStamps = entryPlug->compStamps; |
| 75 |
componentsNmol = entryPlug->componentsNmol; |
| 76 |
AtomsPerProc = new int[parallelData->nProcessors]; |
| 77 |
GroupsPerProc = new int[parallelData->nProcessors]; |
| 78 |
|
| 79 |
parallelData->nAtomsGlobal = entryPlug->n_atoms; |
| 80 |
parallelData->nBondsGlobal = entryPlug->n_bonds; |
| 81 |
parallelData->nBendsGlobal = entryPlug->n_bends; |
| 82 |
parallelData->nTorsionsGlobal = entryPlug->n_torsions; |
| 83 |
parallelData->nSRIGlobal = entryPlug->n_SRI; |
| 84 |
parallelData->nGroupsGlobal = entryPlug->ngroup; |
| 85 |
parallelData->nMolGlobal = entryPlug->n_mol; |
| 86 |
|
| 87 |
if (parallelData->nProcessors > parallelData->nMolGlobal) { |
| 88 |
sprintf( painCave.errMsg, |
| 89 |
"nProcessors (%d) > nMol (%d)\n" |
| 90 |
"\tThe number of processors is larger than\n" |
| 91 |
"\tthe number of molecules. This will not result in a \n" |
| 92 |
"\tusable division of atoms for force decomposition.\n" |
| 93 |
"\tEither try a smaller number of processors, or run the\n" |
| 94 |
"\tsingle-processor version of OOPSE.\n", |
| 95 |
parallelData->nProcessors, parallelData->nMolGlobal ); |
| 96 |
painCave.isFatal = 1; |
| 97 |
simError(); |
| 98 |
} |
| 99 |
|
| 100 |
myRandom = new randomSPRNG( baseSeed ); |
| 101 |
|
| 102 |
|
| 103 |
a = 3.0 * (double)parallelData->nMolGlobal / (double)parallelData->nAtomsGlobal; |
| 104 |
|
| 105 |
// Initialize things that we'll send out later: |
| 106 |
for (i = 0; i < parallelData->nProcessors; i++ ) { |
| 107 |
AtomsPerProc[i] = 0; |
| 108 |
GroupsPerProc[i] = 0; |
| 109 |
} |
| 110 |
for (i = 0; i < parallelData->nMolGlobal; i++ ) { |
| 111 |
// default to an error condition: |
| 112 |
MolToProcMap[i] = -1; |
| 113 |
MolComponentType[i] = -1; |
| 114 |
} |
| 115 |
for (i = 0; i < parallelData->nAtomsGlobal; i++ ) { |
| 116 |
// default to an error condition: |
| 117 |
AtomToProcMap[i] = -1; |
| 118 |
} |
| 119 |
for (i = 0; i < parallelData->nGroupsGlobal; i++ ) { |
| 120 |
// default to an error condition: |
| 121 |
GroupToProcMap[i] = -1; |
| 122 |
} |
| 123 |
|
| 124 |
if (parallelData->myNode == 0) { |
| 125 |
numerator = (double) entryPlug->n_atoms; |
| 126 |
denominator = (double) parallelData->nProcessors; |
| 127 |
precast = numerator / denominator; |
| 128 |
nTarget = (int)( precast + 0.5 ); |
| 129 |
|
| 130 |
// Build the array of molecule component types first |
| 131 |
molIndex = 0; |
| 132 |
for (i=0; i < nComponents; i++) { |
| 133 |
for (j=0; j < componentsNmol[i]; j++) { |
| 134 |
MolComponentType[molIndex] = i; |
| 135 |
molIndex++; |
| 136 |
} |
| 137 |
} |
| 138 |
|
| 139 |
atomIndex = 0; |
| 140 |
groupIndex = 0; |
| 141 |
|
| 142 |
for (i = 0; i < molIndex; i++ ) { |
| 143 |
|
| 144 |
done = 0; |
| 145 |
loops = 0; |
| 146 |
|
| 147 |
while( !done ){ |
| 148 |
loops++; |
| 149 |
|
| 150 |
// Pick a processor at random |
| 151 |
|
| 152 |
which_proc = (int) (myRandom->getRandom() * parallelData->nProcessors); |
| 153 |
|
| 154 |
// How many atoms does this processor have? |
| 155 |
|
| 156 |
old_atoms = AtomsPerProc[which_proc]; |
| 157 |
add_atoms = compStamps[MolComponentType[i]]->getNAtoms(); |
| 158 |
new_atoms = old_atoms + add_atoms; |
| 159 |
|
| 160 |
old_groups = GroupsPerProc[which_proc]; |
| 161 |
ncutoff_groups = compStamps[MolComponentType[i]]->getNCutoffGroups(); |
| 162 |
nAtomsInGroups = 0; |
| 163 |
for (j = 0; j < ncutoff_groups; j++) { |
| 164 |
cg = compStamps[MolComponentType[i]]->getCutoffGroup(j); |
| 165 |
nAtomsInGroups += cg->getNMembers(); |
| 166 |
} |
| 167 |
add_groups = add_atoms - nAtomsInGroups + ncutoff_groups; |
| 168 |
new_groups = old_groups + add_groups; |
| 169 |
|
| 170 |
// If we've been through this loop too many times, we need |
| 171 |
// to just give up and assign the molecule to this processor |
| 172 |
// and be done with it. |
| 173 |
|
| 174 |
if (loops > 100) { |
| 175 |
sprintf( painCave.errMsg, |
| 176 |
"I've tried 100 times to assign molecule %d to a " |
| 177 |
" processor, but can't find a good spot.\n" |
| 178 |
"I'm assigning it at random to processor %d.\n", |
| 179 |
i, which_proc); |
| 180 |
painCave.isFatal = 0; |
| 181 |
simError(); |
| 182 |
|
| 183 |
MolToProcMap[i] = which_proc; |
| 184 |
AtomsPerProc[which_proc] += add_atoms; |
| 185 |
for (j = 0 ; j < add_atoms; j++ ) { |
| 186 |
AtomToProcMap[atomIndex] = which_proc; |
| 187 |
atomIndex++; |
| 188 |
} |
| 189 |
GroupsPerProc[which_proc] += add_groups; |
| 190 |
for (j=0; j < add_groups; j++) { |
| 191 |
GroupToProcMap[groupIndex] = which_proc; |
| 192 |
groupIndex++; |
| 193 |
} |
| 194 |
done = 1; |
| 195 |
continue; |
| 196 |
} |
| 197 |
|
| 198 |
// If we can add this molecule to this processor without sending |
| 199 |
// it above nTarget, then go ahead and do it: |
| 200 |
|
| 201 |
if (new_atoms <= nTarget) { |
| 202 |
MolToProcMap[i] = which_proc; |
| 203 |
AtomsPerProc[which_proc] += add_atoms; |
| 204 |
for (j = 0 ; j < add_atoms; j++ ) { |
| 205 |
AtomToProcMap[atomIndex] = which_proc; |
| 206 |
atomIndex++; |
| 207 |
} |
| 208 |
GroupsPerProc[which_proc] += add_groups; |
| 209 |
for (j=0; j < add_groups; j++) { |
| 210 |
GroupToProcMap[groupIndex] = which_proc; |
| 211 |
groupIndex++; |
| 212 |
} |
| 213 |
done = 1; |
| 214 |
continue; |
| 215 |
} |
| 216 |
|
| 217 |
|
| 218 |
// The only situation left is when new_atoms > nTarget. We |
| 219 |
// want to accept this with some probability that dies off the |
| 220 |
// farther we are from nTarget |
| 221 |
|
| 222 |
// roughly: x = new_atoms - nTarget |
| 223 |
// Pacc(x) = exp(- a * x) |
| 224 |
// where a = penalty / (average atoms per molecule) |
| 225 |
|
| 226 |
x = (double) (new_atoms - nTarget); |
| 227 |
y = myRandom->getRandom(); |
| 228 |
|
| 229 |
if (y < exp(- a * x)) { |
| 230 |
MolToProcMap[i] = which_proc; |
| 231 |
AtomsPerProc[which_proc] += add_atoms; |
| 232 |
for (j = 0 ; j < add_atoms; j++ ) { |
| 233 |
AtomToProcMap[atomIndex] = which_proc; |
| 234 |
atomIndex++; |
| 235 |
} |
| 236 |
GroupsPerProc[which_proc] += add_groups; |
| 237 |
for (j=0; j < add_groups; j++) { |
| 238 |
GroupToProcMap[groupIndex] = which_proc; |
| 239 |
groupIndex++; |
| 240 |
} |
| 241 |
done = 1; |
| 242 |
continue; |
| 243 |
} else { |
| 244 |
continue; |
| 245 |
} |
| 246 |
|
| 247 |
} |
| 248 |
} |
| 249 |
|
| 250 |
|
| 251 |
// Spray out this nonsense to all other processors: |
| 252 |
|
| 253 |
//std::cerr << "node 0 mol2proc = \n"; |
| 254 |
//for (i = 0; i < parallelData->nMolGlobal; i++) |
| 255 |
// std::cerr << i << "\t" << MolToProcMap[i] << "\n"; |
| 256 |
|
| 257 |
MPI_Bcast(MolToProcMap, parallelData->nMolGlobal, |
| 258 |
MPI_INT, 0, MPI_COMM_WORLD); |
| 259 |
|
| 260 |
MPI_Bcast(AtomToProcMap, parallelData->nAtomsGlobal, |
| 261 |
MPI_INT, 0, MPI_COMM_WORLD); |
| 262 |
|
| 263 |
MPI_Bcast(GroupToProcMap, parallelData->nGroupsGlobal, |
| 264 |
MPI_INT, 0, MPI_COMM_WORLD); |
| 265 |
|
| 266 |
MPI_Bcast(MolComponentType, parallelData->nMolGlobal, |
| 267 |
MPI_INT, 0, MPI_COMM_WORLD); |
| 268 |
|
| 269 |
MPI_Bcast(AtomsPerProc, parallelData->nProcessors, |
| 270 |
MPI_INT, 0, MPI_COMM_WORLD); |
| 271 |
|
| 272 |
MPI_Bcast(GroupsPerProc, parallelData->nProcessors, |
| 273 |
MPI_INT, 0, MPI_COMM_WORLD); |
| 274 |
} else { |
| 275 |
|
| 276 |
// Listen to your marching orders from processor 0: |
| 277 |
|
| 278 |
MPI_Bcast(MolToProcMap, parallelData->nMolGlobal, |
| 279 |
MPI_INT, 0, MPI_COMM_WORLD); |
| 280 |
|
| 281 |
MPI_Bcast(AtomToProcMap, parallelData->nAtomsGlobal, |
| 282 |
MPI_INT, 0, MPI_COMM_WORLD); |
| 283 |
|
| 284 |
MPI_Bcast(GroupToProcMap, parallelData->nGroupsGlobal, |
| 285 |
MPI_INT, 0, MPI_COMM_WORLD); |
| 286 |
|
| 287 |
MPI_Bcast(MolComponentType, parallelData->nMolGlobal, |
| 288 |
MPI_INT, 0, MPI_COMM_WORLD); |
| 289 |
|
| 290 |
MPI_Bcast(AtomsPerProc, parallelData->nProcessors, |
| 291 |
MPI_INT, 0, MPI_COMM_WORLD); |
| 292 |
|
| 293 |
MPI_Bcast(GroupsPerProc, parallelData->nProcessors, |
| 294 |
MPI_INT, 0, MPI_COMM_WORLD); |
| 295 |
|
| 296 |
|
| 297 |
} |
| 298 |
|
| 299 |
// Let's all check for sanity: |
| 300 |
|
| 301 |
nmol_local = 0; |
| 302 |
for (i = 0 ; i < parallelData->nMolGlobal; i++ ) { |
| 303 |
if (MolToProcMap[i] == parallelData->myNode) { |
| 304 |
nmol_local++; |
| 305 |
} |
| 306 |
} |
| 307 |
|
| 308 |
natoms_local = 0; |
| 309 |
for (i = 0; i < parallelData->nAtomsGlobal; i++) { |
| 310 |
if (AtomToProcMap[i] == parallelData->myNode) { |
| 311 |
natoms_local++; |
| 312 |
} |
| 313 |
} |
| 314 |
|
| 315 |
ngroups_local = 0; |
| 316 |
for (i = 0; i < parallelData->nGroupsGlobal; i++) { |
| 317 |
if (GroupToProcMap[i] == parallelData->myNode) { |
| 318 |
ngroups_local++; |
| 319 |
} |
| 320 |
} |
| 321 |
|
| 322 |
MPI_Allreduce(&nmol_local,&nmol_global,1,MPI_INT,MPI_SUM, |
| 323 |
MPI_COMM_WORLD); |
| 324 |
|
| 325 |
MPI_Allreduce(&natoms_local,&natoms_global,1,MPI_INT, |
| 326 |
MPI_SUM, MPI_COMM_WORLD); |
| 327 |
|
| 328 |
MPI_Allreduce(&ngroups_local,&ngroups_global,1,MPI_INT, |
| 329 |
MPI_SUM, MPI_COMM_WORLD); |
| 330 |
|
| 331 |
if( nmol_global != entryPlug->n_mol ){ |
| 332 |
sprintf( painCave.errMsg, |
| 333 |
"The sum of all nmol_local, %d, did not equal the " |
| 334 |
"total number of molecules, %d.\n", |
| 335 |
nmol_global, entryPlug->n_mol ); |
| 336 |
painCave.isFatal = 1; |
| 337 |
simError(); |
| 338 |
} |
| 339 |
|
| 340 |
if( natoms_global != entryPlug->n_atoms ){ |
| 341 |
sprintf( painCave.errMsg, |
| 342 |
"The sum of all natoms_local, %d, did not equal the " |
| 343 |
"total number of atoms, %d.\n", |
| 344 |
natoms_global, entryPlug->n_atoms ); |
| 345 |
painCave.isFatal = 1; |
| 346 |
simError(); |
| 347 |
} |
| 348 |
|
| 349 |
if( ngroups_global != entryPlug->ngroup ){ |
| 350 |
sprintf( painCave.errMsg, |
| 351 |
"The sum of all ngroups_local, %d, did not equal the " |
| 352 |
"total number of cutoffGroups, %d.\n", |
| 353 |
ngroups_global, entryPlug->ngroup ); |
| 354 |
painCave.isFatal = 1; |
| 355 |
simError(); |
| 356 |
} |
| 357 |
|
| 358 |
sprintf( checkPointMsg, |
| 359 |
"Successfully divided the molecules among the processors.\n" ); |
| 360 |
MPIcheckPoint(); |
| 361 |
|
| 362 |
parallelData->nMolLocal = nmol_local; |
| 363 |
parallelData->nAtomsLocal = natoms_local; |
| 364 |
parallelData->nGroupsLocal = ngroups_local; |
| 365 |
|
| 366 |
globalAtomIndex.resize(parallelData->nAtomsLocal); |
| 367 |
globalToLocalAtom.resize(parallelData->nAtomsGlobal); |
| 368 |
local_index = 0; |
| 369 |
for (i = 0; i < parallelData->nAtomsGlobal; i++) { |
| 370 |
if (AtomToProcMap[i] == parallelData->myNode) { |
| 371 |
globalAtomIndex[local_index] = i; |
| 372 |
|
| 373 |
globalToLocalAtom[i] = local_index; |
| 374 |
local_index++; |
| 375 |
|
| 376 |
} |
| 377 |
else |
| 378 |
globalToLocalAtom[i] = -1; |
| 379 |
} |
| 380 |
|
| 381 |
globalGroupIndex.resize(parallelData->nGroupsLocal); |
| 382 |
globalToLocalGroup.resize(parallelData->nGroupsGlobal); |
| 383 |
local_index = 0; |
| 384 |
for (i = 0; i < parallelData->nGroupsGlobal; i++) { |
| 385 |
if (GroupToProcMap[i] == parallelData->myNode) { |
| 386 |
globalGroupIndex[local_index] = i; |
| 387 |
|
| 388 |
globalToLocalGroup[i] = local_index; |
| 389 |
local_index++; |
| 390 |
|
| 391 |
} |
| 392 |
else |
| 393 |
globalToLocalGroup[i] = -1; |
| 394 |
} |
| 395 |
|
| 396 |
globalMolIndex.resize(parallelData->nMolLocal); |
| 397 |
globalToLocalMol.resize(parallelData->nMolGlobal); |
| 398 |
local_index = 0; |
| 399 |
for (i = 0; i < parallelData->nMolGlobal; i++) { |
| 400 |
if (MolToProcMap[i] == parallelData->myNode) { |
| 401 |
globalMolIndex[local_index] = i; |
| 402 |
globalToLocalMol[i] = local_index; |
| 403 |
local_index++; |
| 404 |
} |
| 405 |
else |
| 406 |
globalToLocalMol[i] = -1; |
| 407 |
} |
| 408 |
|
| 409 |
} |
| 410 |
|
| 411 |
|
| 412 |
void mpiSimulation::mpiRefresh( void ){ |
| 413 |
|
| 414 |
int isError, i; |
| 415 |
int *localToGlobalAtomIndex = new int[parallelData->nAtomsLocal]; |
| 416 |
int *localToGlobalGroupIndex = new int[parallelData->nGroupsLocal]; |
| 417 |
|
| 418 |
// Fortran indexing needs to be increased by 1 in order to get the 2 |
| 419 |
// languages to not barf |
| 420 |
|
| 421 |
for(i = 0; i < parallelData->nAtomsLocal; i++) |
| 422 |
localToGlobalAtomIndex[i] = globalAtomIndex[i] + 1; |
| 423 |
|
| 424 |
for(i = 0; i < parallelData->nGroupsLocal; i++) |
| 425 |
localToGlobalGroupIndex[i] = globalGroupIndex[i] + 1; |
| 426 |
|
| 427 |
isError = 0; |
| 428 |
|
| 429 |
setFsimParallel( parallelData, |
| 430 |
&(parallelData->nAtomsLocal), localToGlobalAtomIndex, |
| 431 |
&(parallelData->nGroupsLocal), localToGlobalGroupIndex, |
| 432 |
&isError ); |
| 433 |
|
| 434 |
if( isError ){ |
| 435 |
|
| 436 |
sprintf( painCave.errMsg, |
| 437 |
"mpiRefresh errror: fortran didn't like something we gave it.\n" ); |
| 438 |
painCave.isFatal = 1; |
| 439 |
simError(); |
| 440 |
} |
| 441 |
|
| 442 |
delete[] localToGlobalGroupIndex; |
| 443 |
delete[] localToGlobalAtomIndex; |
| 444 |
|
| 445 |
|
| 446 |
sprintf( checkPointMsg, |
| 447 |
" mpiRefresh successful.\n" ); |
| 448 |
MPIcheckPoint(); |
| 449 |
} |
| 450 |
|
| 451 |
|
| 452 |
#endif // is_mpi |