1 |
|
#ifdef IS_MPI |
2 |
< |
|
2 |
> |
#include <iostream> |
3 |
|
#include <cstdlib> |
4 |
|
#include <cstring> |
5 |
+ |
#include <cmath> |
6 |
|
#include <mpi.h> |
7 |
|
#include <mpi++.h> |
8 |
|
|
9 |
|
#include "mpiSimulation.hpp" |
10 |
|
#include "simError.h" |
11 |
|
#include "fortranWrappers.hpp" |
12 |
+ |
#include "randomSPRNG.hpp" |
13 |
|
|
14 |
+ |
#define BASE_SEED 123456789 |
15 |
|
|
13 |
– |
|
14 |
– |
|
16 |
|
mpiSimulation* mpiSim; |
17 |
|
|
18 |
|
mpiSimulation::mpiSimulation(SimInfo* the_entryPlug) |
22 |
|
|
23 |
|
mpiPlug->numberProcessors = MPI::COMM_WORLD.Get_size(); |
24 |
|
mpiPlug->myNode = worldRank; |
25 |
< |
|
25 |
> |
|
26 |
> |
MolToProcMap = new int[entryPlug->n_mol]; |
27 |
> |
MolComponentType = new int[entryPlug->n_mol]; |
28 |
> |
AtomToProcMap = new int[entryPlug->n_atoms]; |
29 |
> |
|
30 |
|
mpiSim = this; |
31 |
|
wrapMeSimParallel( this ); |
32 |
|
} |
34 |
|
|
35 |
|
mpiSimulation::~mpiSimulation(){ |
36 |
|
|
37 |
+ |
delete[] MolToProcMap; |
38 |
+ |
delete[] MolComponentType; |
39 |
+ |
delete[] AtomToProcMap; |
40 |
+ |
|
41 |
|
delete mpiPlug; |
42 |
|
// perhaps we should let fortran know the party is over. |
43 |
|
|
44 |
|
} |
45 |
|
|
37 |
– |
|
38 |
– |
|
46 |
|
int* mpiSimulation::divideLabor( void ){ |
47 |
|
|
48 |
|
int* globalIndex; |
49 |
|
|
50 |
|
int nComponents; |
51 |
|
MoleculeStamp** compStamps; |
52 |
+ |
randomSPRNG *myRandom; |
53 |
|
int* componentsNmol; |
54 |
+ |
int* AtomsPerProc; |
55 |
|
|
56 |
|
double numerator; |
57 |
|
double denominator; |
58 |
|
double precast; |
59 |
+ |
double x, y, a; |
60 |
+ |
int old_atoms, add_atoms, new_atoms; |
61 |
|
|
62 |
|
int nTarget; |
63 |
|
int molIndex, atomIndex, compIndex, compStart; |
64 |
|
int done; |
65 |
|
int nLocal, molLocal; |
66 |
< |
int i, index; |
66 |
> |
int i, j, loops, which_proc, nmol_local, natoms_local; |
67 |
> |
int nmol_global, natoms_global; |
68 |
> |
int local_index, index; |
69 |
|
int smallDiff, bigDiff; |
70 |
+ |
int baseSeed = BASE_SEED; |
71 |
|
|
72 |
|
int testSum; |
73 |
|
|
74 |
|
nComponents = entryPlug->nComponents; |
75 |
|
compStamps = entryPlug->compStamps; |
76 |
|
componentsNmol = entryPlug->componentsNmol; |
77 |
< |
|
77 |
> |
AtomsPerProc = new int[mpiPlug->numberProcessors]; |
78 |
> |
|
79 |
|
mpiPlug->nAtomsGlobal = entryPlug->n_atoms; |
80 |
|
mpiPlug->nBondsGlobal = entryPlug->n_bonds; |
81 |
|
mpiPlug->nBendsGlobal = entryPlug->n_bends; |
83 |
|
mpiPlug->nSRIGlobal = entryPlug->n_SRI; |
84 |
|
mpiPlug->nMolGlobal = entryPlug->n_mol; |
85 |
|
|
86 |
+ |
myRandom = new randomSPRNG( baseSeed ); |
87 |
|
|
88 |
+ |
a = 3.0 * (double)mpiPlug->nMolGlobal / (double)mpiPlug->nAtomsGlobal; |
89 |
|
|
90 |
+ |
// Initialize things that we'll send out later: |
91 |
+ |
for (i = 0; i < mpiPlug->numberProcessors; i++ ) { |
92 |
+ |
AtomsPerProc[i] = 0; |
93 |
+ |
} |
94 |
+ |
for (i = 0; i < mpiPlug->nMolGlobal; i++ ) { |
95 |
+ |
// default to an error condition: |
96 |
+ |
MolToProcMap[i] = -1; |
97 |
+ |
MolComponentType[i] = -1; |
98 |
+ |
} |
99 |
+ |
for (i = 0; i < mpiPlug->nAtomsGlobal; i++ ) { |
100 |
+ |
// default to an error condition: |
101 |
+ |
AtomToProcMap[i] = -1; |
102 |
+ |
} |
103 |
+ |
|
104 |
+ |
if (mpiPlug->myNode == 0) { |
105 |
+ |
numerator = (double) entryPlug->n_atoms; |
106 |
+ |
denominator = (double) mpiPlug->numberProcessors; |
107 |
+ |
precast = numerator / denominator; |
108 |
+ |
nTarget = (int)( precast + 0.5 ); |
109 |
|
|
110 |
+ |
// Build the array of molecule component types first |
111 |
+ |
molIndex = 0; |
112 |
+ |
for (i=0; i < nComponents; i++) { |
113 |
+ |
for (j=0; j < componentsNmol[i]; j++) { |
114 |
+ |
MolComponentType[molIndex] = i; |
115 |
+ |
molIndex++; |
116 |
+ |
} |
117 |
+ |
} |
118 |
|
|
119 |
+ |
atomIndex = 0; |
120 |
|
|
121 |
+ |
for (i = 0; i < molIndex; i++ ) { |
122 |
|
|
123 |
+ |
done = 0; |
124 |
+ |
loops = 0; |
125 |
|
|
126 |
+ |
while( !done ){ |
127 |
+ |
loops++; |
128 |
+ |
|
129 |
+ |
// Pick a processor at random |
130 |
|
|
131 |
+ |
which_proc = (int) (myRandom->getRandom() * mpiPlug->numberProcessors); |
132 |
|
|
133 |
+ |
// How many atoms does this processor have? |
134 |
+ |
|
135 |
+ |
old_atoms = AtomsPerProc[which_proc]; |
136 |
+ |
add_atoms = compStamps[MolComponentType[i]]->getNAtoms(); |
137 |
+ |
new_atoms = old_atoms + add_atoms; |
138 |
|
|
139 |
< |
|
140 |
< |
numerator = (double) entryPlug->n_atoms; |
141 |
< |
denominator = (double) mpiPlug->numberProcessors; |
142 |
< |
precast = numerator / denominator; |
143 |
< |
nTarget = (int)( precast + 0.5 ); |
144 |
< |
|
145 |
< |
molIndex = 0; |
146 |
< |
atomIndex = 0; |
147 |
< |
compIndex = 0; |
148 |
< |
compStart = 0; |
149 |
< |
for( i=0; i<(mpiPlug->numberProcessors-1); i++){ |
139 |
> |
// If we've been through this loop too many times, we need |
140 |
> |
// to just give up and assign the molecule to this processor |
141 |
> |
// and be done with it. |
142 |
> |
|
143 |
> |
if (loops > 100) { |
144 |
> |
sprintf( painCave.errMsg, |
145 |
> |
"I've tried 100 times to assign molecule %d to a " |
146 |
> |
" processor, but can't find a good spot.\n" |
147 |
> |
"I'm assigning it at random to processor %d.\n", |
148 |
> |
i, which_proc); |
149 |
> |
painCave.isFatal = 0; |
150 |
> |
simError(); |
151 |
> |
|
152 |
> |
MolToProcMap[i] = which_proc; |
153 |
> |
AtomsPerProc[which_proc] += add_atoms; |
154 |
> |
for (j = 0 ; j < add_atoms; j++ ) { |
155 |
> |
AtomToProcMap[atomIndex] = which_proc; |
156 |
> |
atomIndex++; |
157 |
> |
} |
158 |
> |
done = 1; |
159 |
> |
continue; |
160 |
> |
} |
161 |
|
|
162 |
< |
done = 0; |
163 |
< |
nLocal = 0; |
95 |
< |
molLocal = 0; |
96 |
< |
|
97 |
< |
if( i == mpiPlug->myNode ){ |
98 |
< |
mpiPlug->myMolStart = molIndex; |
99 |
< |
mpiPlug->myAtomStart = atomIndex; |
100 |
< |
} |
162 |
> |
// If we can add this molecule to this processor without sending |
163 |
> |
// it above nTarget, then go ahead and do it: |
164 |
|
|
165 |
< |
while( !done ){ |
166 |
< |
|
167 |
< |
if( (molIndex-compStart) >= componentsNmol[compIndex] ){ |
168 |
< |
compStart = molIndex; |
169 |
< |
compIndex++; |
170 |
< |
continue; |
171 |
< |
} |
165 |
> |
if (new_atoms <= nTarget) { |
166 |
> |
MolToProcMap[i] = which_proc; |
167 |
> |
AtomsPerProc[which_proc] += add_atoms; |
168 |
> |
for (j = 0 ; j < add_atoms; j++ ) { |
169 |
> |
AtomToProcMap[atomIndex] = which_proc; |
170 |
> |
atomIndex++; |
171 |
> |
} |
172 |
> |
done = 1; |
173 |
> |
continue; |
174 |
> |
} |
175 |
|
|
176 |
< |
nLocal += compStamps[compIndex]->getNAtoms(); |
177 |
< |
atomIndex += compStamps[compIndex]->getNAtoms(); |
178 |
< |
molIndex++; |
179 |
< |
molLocal++; |
176 |
> |
|
177 |
> |
// The only situation left is when new_atoms > nTarget. We |
178 |
> |
// want to accept this with some probability that dies off the |
179 |
> |
// farther we are from nTarget |
180 |
> |
|
181 |
> |
// roughly: x = new_atoms - nTarget |
182 |
> |
// Pacc(x) = exp(- a * x) |
183 |
> |
// where a = penalty / (average atoms per molecule) |
184 |
> |
|
185 |
> |
x = (double) (new_atoms - nTarget); |
186 |
> |
y = myRandom->getRandom(); |
187 |
|
|
188 |
< |
if ( nLocal == nTarget ) done = 1; |
189 |
< |
|
190 |
< |
else if( nLocal < nTarget ){ |
191 |
< |
smallDiff = nTarget - nLocal; |
192 |
< |
} |
193 |
< |
else if( nLocal > nTarget ){ |
194 |
< |
bigDiff = nLocal - nTarget; |
195 |
< |
|
196 |
< |
if( bigDiff < smallDiff ) done = 1; |
197 |
< |
else{ |
198 |
< |
molIndex--; |
199 |
< |
molLocal--; |
200 |
< |
atomIndex -= compStamps[compIndex]->getNAtoms(); |
128 |
< |
nLocal -= compStamps[compIndex]->getNAtoms(); |
129 |
< |
done = 1; |
130 |
< |
} |
188 |
> |
if (y < exp(- a * x)) { |
189 |
> |
MolToProcMap[i] = which_proc; |
190 |
> |
AtomsPerProc[which_proc] += add_atoms; |
191 |
> |
for (j = 0 ; j < add_atoms; j++ ) { |
192 |
> |
AtomToProcMap[atomIndex] = which_proc; |
193 |
> |
atomIndex++; |
194 |
> |
} |
195 |
> |
done = 1; |
196 |
> |
continue; |
197 |
> |
} else { |
198 |
> |
continue; |
199 |
> |
} |
200 |
> |
|
201 |
|
} |
202 |
|
} |
203 |
+ |
|
204 |
+ |
// Spray out this nonsense to all other processors: |
205 |
+ |
|
206 |
+ |
MPI::COMM_WORLD.Bcast(MolToProcMap, mpiPlug->nMolGlobal, |
207 |
+ |
MPI_INT, 0); |
208 |
+ |
|
209 |
+ |
MPI::COMM_WORLD.Bcast(AtomToProcMap, mpiPlug->nAtomsGlobal, |
210 |
+ |
MPI_INT, 0); |
211 |
+ |
|
212 |
+ |
MPI::COMM_WORLD.Bcast(MolComponentType, mpiPlug->nMolGlobal, |
213 |
+ |
MPI_INT, 0); |
214 |
+ |
|
215 |
+ |
MPI::COMM_WORLD.Bcast(AtomsPerProc, mpiPlug->numberProcessors, |
216 |
+ |
MPI_INT, 0); |
217 |
+ |
} else { |
218 |
+ |
|
219 |
+ |
// Listen to your marching orders from processor 0: |
220 |
|
|
221 |
< |
if( i == mpiPlug->myNode ){ |
222 |
< |
mpiPlug->myMolEnd = (molIndex - 1); |
136 |
< |
mpiPlug->myAtomEnd = (atomIndex - 1); |
137 |
< |
mpiPlug->myNlocal = nLocal; |
138 |
< |
mpiPlug->myMol = molLocal; |
139 |
< |
} |
221 |
> |
MPI::COMM_WORLD.Bcast(MolToProcMap, mpiPlug->nMolGlobal, |
222 |
> |
MPI_INT, 0); |
223 |
|
|
224 |
< |
numerator = (double)( entryPlug->n_atoms - atomIndex ); |
225 |
< |
denominator = (double)( mpiPlug->numberProcessors - (i+1) ); |
226 |
< |
precast = numerator / denominator; |
227 |
< |
nTarget = (int)( precast + 0.5 ); |
224 |
> |
MPI::COMM_WORLD.Bcast(AtomToProcMap, mpiPlug->nAtomsGlobal, |
225 |
> |
MPI_INT, 0); |
226 |
> |
|
227 |
> |
MPI::COMM_WORLD.Bcast(MolComponentType, mpiPlug->nMolGlobal, |
228 |
> |
MPI_INT, 0); |
229 |
> |
|
230 |
> |
MPI::COMM_WORLD.Bcast(AtomsPerProc, mpiPlug->numberProcessors, |
231 |
> |
MPI_INT, 0); |
232 |
> |
|
233 |
> |
|
234 |
|
} |
146 |
– |
|
147 |
– |
if( mpiPlug->myNode == mpiPlug->numberProcessors-1 ){ |
148 |
– |
mpiPlug->myMolStart = molIndex; |
149 |
– |
mpiPlug->myAtomStart = atomIndex; |
235 |
|
|
151 |
– |
nLocal = 0; |
152 |
– |
molLocal = 0; |
153 |
– |
while( compIndex < nComponents ){ |
154 |
– |
|
155 |
– |
if( (molIndex-compStart) >= componentsNmol[compIndex] ){ |
156 |
– |
compStart = molIndex; |
157 |
– |
compIndex++; |
158 |
– |
continue; |
159 |
– |
} |
236 |
|
|
237 |
< |
nLocal += compStamps[compIndex]->getNAtoms(); |
238 |
< |
atomIndex += compStamps[compIndex]->getNAtoms(); |
239 |
< |
molIndex++; |
240 |
< |
molLocal++; |
241 |
< |
} |
242 |
< |
|
243 |
< |
mpiPlug->myMolEnd = (molIndex - 1); |
168 |
< |
mpiPlug->myAtomEnd = (atomIndex - 1); |
169 |
< |
mpiPlug->myNlocal = nLocal; |
170 |
< |
mpiPlug->myMol = molLocal; |
237 |
> |
// Let's all check for sanity: |
238 |
> |
|
239 |
> |
nmol_local = 0; |
240 |
> |
for (i = 0 ; i < mpiPlug->nMolGlobal; i++ ) { |
241 |
> |
if (MolToProcMap[i] == mpiPlug->myNode) { |
242 |
> |
nmol_local++; |
243 |
> |
} |
244 |
|
} |
245 |
|
|
246 |
+ |
natoms_local = 0; |
247 |
+ |
for (i = 0; i < mpiPlug->nAtomsGlobal; i++) { |
248 |
+ |
if (AtomToProcMap[i] == mpiPlug->myNode) { |
249 |
+ |
natoms_local++; |
250 |
+ |
} |
251 |
+ |
} |
252 |
|
|
253 |
< |
MPI_Allreduce( &nLocal, &testSum, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD ); |
253 |
> |
MPI::COMM_WORLD.Allreduce(&nmol_local,&nmol_global,1,MPI_INT,MPI_SUM); |
254 |
> |
MPI::COMM_WORLD.Allreduce(&natoms_local,&natoms_global,1,MPI_INT,MPI_SUM); |
255 |
|
|
256 |
< |
if( mpiPlug->myNode == 0 ){ |
257 |
< |
if( testSum != entryPlug->n_atoms ){ |
258 |
< |
sprintf( painCave.errMsg, |
259 |
< |
"The summ of all nLocals, %d, did not equal the total number of atoms, %d.\n", |
260 |
< |
testSum, entryPlug->n_atoms ); |
261 |
< |
painCave.isFatal = 1; |
262 |
< |
simError(); |
183 |
< |
} |
256 |
> |
if( nmol_global != entryPlug->n_mol ){ |
257 |
> |
sprintf( painCave.errMsg, |
258 |
> |
"The sum of all nmol_local, %d, did not equal the " |
259 |
> |
"total number of molecules, %d.\n", |
260 |
> |
nmol_global, entryPlug->n_mol ); |
261 |
> |
painCave.isFatal = 1; |
262 |
> |
simError(); |
263 |
|
} |
264 |
+ |
|
265 |
+ |
if( natoms_global != entryPlug->n_atoms ){ |
266 |
+ |
sprintf( painCave.errMsg, |
267 |
+ |
"The sum of all natoms_local, %d, did not equal the " |
268 |
+ |
"total number of atoms, %d.\n", |
269 |
+ |
natoms_global, entryPlug->n_atoms ); |
270 |
+ |
painCave.isFatal = 1; |
271 |
+ |
simError(); |
272 |
+ |
} |
273 |
|
|
274 |
|
sprintf( checkPointMsg, |
275 |
|
"Successfully divided the molecules among the processors.\n" ); |
276 |
|
MPIcheckPoint(); |
277 |
|
|
278 |
< |
// lets create the identity array |
278 |
> |
mpiPlug->myNMol = nmol_local; |
279 |
> |
mpiPlug->myNlocal = natoms_local; |
280 |
|
|
281 |
|
globalIndex = new int[mpiPlug->myNlocal]; |
282 |
< |
index = mpiPlug->myAtomStart; |
283 |
< |
for( i=0; i<mpiPlug->myNlocal; i++){ |
284 |
< |
globalIndex[i] = index; |
285 |
< |
index++; |
282 |
> |
local_index = 0; |
283 |
> |
for (i = 0; i < mpiPlug->nAtomsGlobal; i++) { |
284 |
> |
if (AtomToProcMap[i] == mpiPlug->myNode) { |
285 |
> |
globalIndex[local_index] = i; |
286 |
> |
local_index++; |
287 |
> |
} |
288 |
|
} |
289 |
< |
|
289 |
> |
|
290 |
|
return globalIndex; |
291 |
|
} |
292 |
|
|