| 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/DarkSide/simParallel_interface.h" |
| 11 |
#include "math/randomSPRNG.hpp" |
| 12 |
|
| 13 |
mpiSimulation* mpiSim; |
| 14 |
|
| 15 |
mpiSimulation::mpiSimulation(SimInfo* the_entryPlug) |
| 16 |
{ |
| 17 |
parallelData = new mpiSimData; |
| 18 |
|
| 19 |
MPI_Comm_size(MPI_COMM_WORLD, &(parallelData->nProcessors) ); |
| 20 |
parallelData->myNode = worldRank; |
| 21 |
|
| 22 |
MolToProcMap = new int[entryPlug->n_mol]; |
| 23 |
|
| 24 |
} |
| 25 |
|
| 26 |
|
| 27 |
mpiSimulation::~mpiSimulation(){ |
| 28 |
|
| 29 |
delete[] MolToProcMap; |
| 30 |
|
| 31 |
delete parallelData; |
| 32 |
// perhaps we should let fortran know the party is over. |
| 33 |
|
| 34 |
} |
| 35 |
|
| 36 |
void mpiSimulation::divideLabor( ){ |
| 37 |
|
| 38 |
int nComponents; |
| 39 |
MoleculeStamp** compStamps; |
| 40 |
randomSPRNG *myRandom; |
| 41 |
int* componentsNmol; |
| 42 |
int* AtomsPerProc; |
| 43 |
int* GroupsPerProc; |
| 44 |
|
| 45 |
double numerator; |
| 46 |
double denominator; |
| 47 |
double precast; |
| 48 |
double x, y, a; |
| 49 |
int old_atoms, add_atoms, new_atoms; |
| 50 |
int old_groups, add_groups, new_groups; |
| 51 |
|
| 52 |
int nTarget; |
| 53 |
int molIndex, atomIndex, groupIndex; |
| 54 |
int done; |
| 55 |
int i, j, loops, which_proc; |
| 56 |
int nmol_global, nmol_local; |
| 57 |
int ngroups_global, ngroups_local; |
| 58 |
int natoms_global, natoms_local; |
| 59 |
int ncutoff_groups, nAtomsInGroups; |
| 60 |
int local_index; |
| 61 |
int baseSeed = entryPlug->getSeed(); |
| 62 |
CutoffGroupStamp* cg; |
| 63 |
|
| 64 |
nComponents = entryPlug->nComponents; |
| 65 |
compStamps = entryPlug->compStamps; |
| 66 |
componentsNmol = entryPlug->componentsNmol; |
| 67 |
AtomsPerProc = new int[parallelData->nProcessors]; |
| 68 |
GroupsPerProc = new int[parallelData->nProcessors]; |
| 69 |
|
| 70 |
parallelData->nAtomsGlobal = entryPlug->n_atoms; |
| 71 |
parallelData->nGroupsGlobal = entryPlug->ngroup; |
| 72 |
parallelData->nMolGlobal = entryPlug->n_mol; |
| 73 |
|
| 74 |
if (parallelData->nProcessors > parallelData->nMolGlobal) { |
| 75 |
sprintf( painCave.errMsg, |
| 76 |
"nProcessors (%d) > nMol (%d)\n" |
| 77 |
"\tThe number of processors is larger than\n" |
| 78 |
"\tthe number of molecules. This will not result in a \n" |
| 79 |
"\tusable division of atoms for force decomposition.\n" |
| 80 |
"\tEither try a smaller number of processors, or run the\n" |
| 81 |
"\tsingle-processor version of OOPSE.\n", |
| 82 |
parallelData->nProcessors, parallelData->nMolGlobal ); |
| 83 |
painCave.isFatal = 1; |
| 84 |
simError(); |
| 85 |
} |
| 86 |
|
| 87 |
myRandom = new randomSPRNG( baseSeed ); |
| 88 |
|
| 89 |
|
| 90 |
a = 3.0 * (double)parallelData->nMolGlobal / (double)parallelData->nAtomsGlobal; |
| 91 |
|
| 92 |
// Initialize things that we'll send out later: |
| 93 |
for (i = 0; i < parallelData->nProcessors; i++ ) { |
| 94 |
AtomsPerProc[i] = 0; |
| 95 |
GroupsPerProc[i] = 0; |
| 96 |
} |
| 97 |
for (i = 0; i < parallelData->nMolGlobal; i++ ) { |
| 98 |
// default to an error condition: |
| 99 |
MolToProcMap[i] = -1; |
| 100 |
} |
| 101 |
|
| 102 |
if (parallelData->myNode == 0) { |
| 103 |
numerator = (double) entryPlug->n_atoms; |
| 104 |
denominator = (double) parallelData->nProcessors; |
| 105 |
precast = numerator / denominator; |
| 106 |
nTarget = (int)( precast + 0.5 ); |
| 107 |
|
| 108 |
// Build the array of molecule component types first |
| 109 |
molIndex = 0; |
| 110 |
for (i=0; i < nComponents; i++) { |
| 111 |
for (j=0; j < componentsNmol[i]; j++) { |
| 112 |
molIndex++; |
| 113 |
} |
| 114 |
} |
| 115 |
|
| 116 |
atomIndex = 0; |
| 117 |
|
| 118 |
for (i = 0; i < molIndex; i++ ) { |
| 119 |
|
| 120 |
done = 0; |
| 121 |
loops = 0; |
| 122 |
|
| 123 |
while( !done ){ |
| 124 |
loops++; |
| 125 |
|
| 126 |
// Pick a processor at random |
| 127 |
|
| 128 |
which_proc = (int) (myRandom->getRandom() * parallelData->nProcessors); |
| 129 |
|
| 130 |
// How many atoms does this processor have? |
| 131 |
|
| 132 |
old_atoms = AtomsPerProc[which_proc]; |
| 133 |
add_atoms = compStamps[MolComponentType[i]]->getNAtoms(); |
| 134 |
new_atoms = old_atoms + add_atoms; |
| 135 |
|
| 136 |
// If we've been through this loop too many times, we need |
| 137 |
// to just give up and assign the molecule to this processor |
| 138 |
// and be done with it. |
| 139 |
|
| 140 |
if (loops > 100) { |
| 141 |
sprintf( painCave.errMsg, |
| 142 |
"I've tried 100 times to assign molecule %d to a " |
| 143 |
" processor, but can't find a good spot.\n" |
| 144 |
"I'm assigning it at random to processor %d.\n", |
| 145 |
i, which_proc); |
| 146 |
painCave.isFatal = 0; |
| 147 |
simError(); |
| 148 |
|
| 149 |
MolToProcMap[i] = which_proc; |
| 150 |
AtomsPerProc[which_proc] += add_atoms; |
| 151 |
|
| 152 |
done = 1; |
| 153 |
continue; |
| 154 |
} |
| 155 |
|
| 156 |
// If we can add this molecule to this processor without sending |
| 157 |
// it above nTarget, then go ahead and do it: |
| 158 |
|
| 159 |
if (new_atoms <= nTarget) { |
| 160 |
MolToProcMap[i] = which_proc; |
| 161 |
AtomsPerProc[which_proc] += add_atoms; |
| 162 |
|
| 163 |
done = 1; |
| 164 |
continue; |
| 165 |
} |
| 166 |
|
| 167 |
|
| 168 |
// The only situation left is when new_atoms > nTarget. We |
| 169 |
// want to accept this with some probability that dies off the |
| 170 |
// farther we are from nTarget |
| 171 |
|
| 172 |
// roughly: x = new_atoms - nTarget |
| 173 |
// Pacc(x) = exp(- a * x) |
| 174 |
// where a = penalty / (average atoms per molecule) |
| 175 |
|
| 176 |
x = (double) (new_atoms - nTarget); |
| 177 |
y = myRandom->getRandom(); |
| 178 |
|
| 179 |
if (y < exp(- a * x)) { |
| 180 |
MolToProcMap[i] = which_proc; |
| 181 |
AtomsPerProc[which_proc] += add_atoms; |
| 182 |
|
| 183 |
done = 1; |
| 184 |
continue; |
| 185 |
} else { |
| 186 |
continue; |
| 187 |
} |
| 188 |
|
| 189 |
} |
| 190 |
} |
| 191 |
|
| 192 |
|
| 193 |
// Spray out this nonsense to all other processors: |
| 194 |
|
| 195 |
MPI_Bcast(MolToProcMap, parallelData->nMolGlobal, |
| 196 |
MPI_INT, 0, MPI_COMM_WORLD); |
| 197 |
|
| 198 |
} else { |
| 199 |
|
| 200 |
// Listen to your marching orders from processor 0: |
| 201 |
|
| 202 |
MPI_Bcast(MolToProcMap, parallelData->nMolGlobal, |
| 203 |
MPI_INT, 0, MPI_COMM_WORLD); |
| 204 |
|
| 205 |
} |
| 206 |
|
| 207 |
// Let's all check for sanity: |
| 208 |
|
| 209 |
nmol_local = 0; |
| 210 |
for (i = 0 ; i < parallelData->nMolGlobal; i++ ) { |
| 211 |
if (MolToProcMap[i] == parallelData->myNode) { |
| 212 |
nmol_local++; |
| 213 |
} |
| 214 |
} |
| 215 |
|
| 216 |
|
| 217 |
MPI_Allreduce(&nmol_local,&nmol_global,1,MPI_INT,MPI_SUM, |
| 218 |
MPI_COMM_WORLD); |
| 219 |
|
| 220 |
if( nmol_global != entryPlug->n_mol ){ |
| 221 |
sprintf( painCave.errMsg, |
| 222 |
"The sum of all nmol_local, %d, did not equal the " |
| 223 |
"total number of molecules, %d.\n", |
| 224 |
nmol_global, entryPlug->n_mol ); |
| 225 |
painCave.isFatal = 1; |
| 226 |
simError(); |
| 227 |
} |
| 228 |
|
| 229 |
sprintf( checkPointMsg, |
| 230 |
"Successfully divided the molecules among the processors.\n" ); |
| 231 |
MPIcheckPoint(); |
| 232 |
|
| 233 |
parallelData->nMolLocal = nmol_local; |
| 234 |
parallelData->nAtomsLocal = natoms_local; |
| 235 |
parallelData->nGroupsLocal = ngroups_local; |
| 236 |
|
| 237 |
globalMolIndex.resize(parallelData->nMolLocal); |
| 238 |
local_index = 0; |
| 239 |
for (i = 0; i < parallelData->nMolGlobal; i++) { |
| 240 |
if (MolToProcMap[i] == parallelData->myNode) { |
| 241 |
globalMolIndex[local_index] = i; |
| 242 |
local_index++; |
| 243 |
} |
| 244 |
|
| 245 |
} |
| 246 |
|
| 247 |
} |
| 248 |
|
| 249 |
|
| 250 |
void mpiSimulation::mpiRefresh( void ){ |
| 251 |
|
| 252 |
int isError, i; |
| 253 |
int *localToGlobalAtomIndex = new int[parallelData->nAtomsLocal]; |
| 254 |
int *localToGlobalGroupIndex = new int[parallelData->nGroupsLocal]; |
| 255 |
|
| 256 |
// Fortran indexing needs to be increased by 1 in order to get the 2 |
| 257 |
// languages to not barf |
| 258 |
|
| 259 |
for(i = 0; i < parallelData->nAtomsLocal; i++) |
| 260 |
localToGlobalAtomIndex[i] = globalAtomIndex[i] + 1; |
| 261 |
|
| 262 |
for(i = 0; i < parallelData->nGroupsLocal; i++) |
| 263 |
localToGlobalGroupIndex[i] = globalGroupIndex[i] + 1; |
| 264 |
|
| 265 |
isError = 0; |
| 266 |
|
| 267 |
setFsimParallel( parallelData, |
| 268 |
&(parallelData->nAtomsLocal), localToGlobalAtomIndex, |
| 269 |
&(parallelData->nGroupsLocal), localToGlobalGroupIndex, |
| 270 |
&isError ); |
| 271 |
|
| 272 |
if( isError ){ |
| 273 |
|
| 274 |
sprintf( painCave.errMsg, |
| 275 |
"mpiRefresh errror: fortran didn't like something we gave it.\n" ); |
| 276 |
painCave.isFatal = 1; |
| 277 |
simError(); |
| 278 |
} |
| 279 |
|
| 280 |
delete[] localToGlobalGroupIndex; |
| 281 |
delete[] localToGlobalAtomIndex; |
| 282 |
|
| 283 |
|
| 284 |
sprintf( checkPointMsg, |
| 285 |
" mpiRefresh successful.\n" ); |
| 286 |
MPIcheckPoint(); |
| 287 |
} |
| 288 |
|
| 289 |
|
| 290 |
#endif // is_mpi |
