1 | #include "grid_remote_connector.hpp" |
---|
2 | #include "client_client_dht_template.hpp" |
---|
3 | #include "leader_process.hpp" |
---|
4 | #include "mpi.hpp" |
---|
5 | #include "element.hpp" |
---|
6 | |
---|
7 | |
---|
8 | |
---|
9 | namespace xios |
---|
10 | { |
---|
11 | /** |
---|
12 | * \brief class constructor. |
---|
13 | * \param srcView List of sources views. |
---|
14 | * \param dstView List of remotes views. |
---|
15 | * \param localComm Local MPI communicator |
---|
16 | * \param remoteSize Size of the remote communicator |
---|
17 | */ |
---|
18 | CGridRemoteConnector::CGridRemoteConnector(vector<shared_ptr<CLocalView>>& srcView, vector<shared_ptr<CDistributedView>>& dstView, MPI_Comm localComm, int remoteSize) |
---|
19 | : srcView_(srcView), dstView_(dstView), localComm_(localComm), remoteSize_(remoteSize) |
---|
20 | {} |
---|
21 | |
---|
22 | /** |
---|
23 | * \brief class constructor. |
---|
24 | * \param srcView List of sources views. |
---|
25 | * \param dstView List of remotes views. |
---|
26 | * \param localComm Local MPI communicator |
---|
27 | * \param remoteSize Size of the remote communicator |
---|
28 | */ |
---|
29 | CGridRemoteConnector::CGridRemoteConnector(vector<shared_ptr<CLocalView>>& srcView, vector< shared_ptr<CLocalView> >& dstView, MPI_Comm localComm, int remoteSize) |
---|
30 | : srcView_(srcView), localComm_(localComm), remoteSize_(remoteSize) |
---|
31 | { |
---|
32 | for(auto& it : dstView) dstView_.push_back((shared_ptr<CDistributedView>) it) ; |
---|
33 | } |
---|
34 | |
---|
35 | |
---|
36 | /** |
---|
37 | * \brief Compute if each view composing the source grid and the remote grid is distributed or not. |
---|
38 | * Result is stored on internal attributes \b isSrcViewDistributed_ and \b isDstViewDistributed_. |
---|
39 | * \detail To compute this, a hash is computed for each array on indices. The hash must permutable, i.e. |
---|
40 | * the order of the list of global indices doesn't influence the value of the hash. So simply a sum of |
---|
41 | * hash of each indices is used for the whole array. After, the computed hash are compared with each other |
---|
42 | * ranks of \b localComm_ MPI communicator using an MPI_ALLReduce. If, for each ranks, the hash is the same |
---|
43 | * then the view is not distributed |
---|
44 | */ |
---|
45 | void CGridRemoteConnector::computeViewDistribution(void) |
---|
46 | { |
---|
47 | HashXIOS<size_t> hashGlobalIndex; // hash function-object |
---|
48 | |
---|
49 | int nDst = dstView_.size() ; |
---|
50 | vector<size_t> hashRank(remoteSize_) ; |
---|
51 | vector<size_t> sizeRank(remoteSize_) ; |
---|
52 | isDstViewDistributed_.resize(nDst) ; |
---|
53 | |
---|
54 | for(int i=0; i<nDst; i++) |
---|
55 | { |
---|
56 | map<int,CArray<size_t,1>> globalIndexView ; |
---|
57 | dstView_[i]->getGlobalIndexView(globalIndexView) ; |
---|
58 | hashRank.assign(remoteSize_,0) ; // everybody ranks to 0 except rank of the remote view I have |
---|
59 | // that would be assign to my local hash |
---|
60 | sizeRank.assign(remoteSize_,0) ; |
---|
61 | |
---|
62 | for(auto& it : globalIndexView) |
---|
63 | { |
---|
64 | int rank=it.first ; |
---|
65 | CArray<size_t,1>& globalIndex = it.second ; |
---|
66 | size_t globalIndexSize = globalIndex.numElements(); |
---|
67 | size_t hashValue=0 ; |
---|
68 | for(size_t ind=0;ind<globalIndexSize;ind++) hashValue += hashGlobalIndex(globalIndex(ind)) ; |
---|
69 | hashRank[rank] += hashValue ; |
---|
70 | sizeRank[rank] += globalIndexSize ; |
---|
71 | } |
---|
72 | // sum all the hash for every process of the local comm. The reduce is on the size of remote view (remoteSize_) |
---|
73 | // after that for each rank of the remote view, we get the hash |
---|
74 | MPI_Allreduce(MPI_IN_PLACE, hashRank.data(), remoteSize_, MPI_SIZE_T, MPI_SUM, localComm_) ; |
---|
75 | MPI_Allreduce(MPI_IN_PLACE, sizeRank.data(), remoteSize_, MPI_SIZE_T, MPI_SUM, localComm_) ; |
---|
76 | size_t value = hashRank[0] ; |
---|
77 | size_t size = sizeRank[0] ; |
---|
78 | isDstViewDistributed_[i]=false ; |
---|
79 | for(int j=0 ; j<remoteSize_ ; j++) |
---|
80 | if (size!=sizeRank[j] || value != hashRank[j]) |
---|
81 | { |
---|
82 | isDstViewDistributed_[i]=true ; |
---|
83 | break ; |
---|
84 | } |
---|
85 | } |
---|
86 | |
---|
87 | int nSrc = srcView_.size() ; |
---|
88 | int commSize,commRank ; |
---|
89 | MPI_Comm_size(localComm_,&commSize) ; |
---|
90 | MPI_Comm_rank(localComm_,&commRank) ; |
---|
91 | hashRank.resize(commSize,0) ; |
---|
92 | isSrcViewDistributed_.resize(nSrc) ; |
---|
93 | |
---|
94 | for(int i=0; i<nSrc; i++) |
---|
95 | { |
---|
96 | CArray<size_t,1> globalIndex ; |
---|
97 | srcView_[i]->getGlobalIndexView(globalIndex) ; |
---|
98 | hashRank.assign(commSize,0) ; // 0 for everybody except my rank |
---|
99 | size_t globalIndexSize = globalIndex.numElements() ; |
---|
100 | |
---|
101 | size_t allEqual ; |
---|
102 | MPI_Allreduce(&globalIndexSize, &allEqual, 1, MPI_SIZE_T, MPI_BXOR, localComm_) ; |
---|
103 | if (allEqual!=0) |
---|
104 | { |
---|
105 | isSrcViewDistributed_[i]=true ; |
---|
106 | break ; |
---|
107 | } |
---|
108 | |
---|
109 | // warning : jenkins hash : 0 --> 0 : need to compare number of element for each ranks |
---|
110 | size_t hashValue=0 ; |
---|
111 | for(size_t ind=0;ind<globalIndexSize;ind++) hashValue += hashGlobalIndex(globalIndex(ind)) ; |
---|
112 | MPI_Allreduce(&hashValue, &allEqual, 1, MPI_SIZE_T, MPI_BXOR, localComm_) ; |
---|
113 | if (allEqual!=0) isSrcViewDistributed_[i]=true ; |
---|
114 | else isSrcViewDistributed_[i]=false ; |
---|
115 | } |
---|
116 | |
---|
117 | } |
---|
118 | |
---|
119 | /** |
---|
120 | * \brief Compute the connector, i.e. compute the \b elements_ attribute. |
---|
121 | * \detail Depending of the distributions of the view computed in the computeViewDistribution() call, the connector is computed in computeConnectorMethods(), and to achieve better optimisation |
---|
122 | * some redondant ranks can be removed from the elements_ map. |
---|
123 | */ |
---|
124 | void CGridRemoteConnector::computeConnector(bool eliminateRedundant) |
---|
125 | { |
---|
126 | if (eliminateRedundant) |
---|
127 | { |
---|
128 | computeViewDistribution() ; |
---|
129 | computeConnectorMethods() ; |
---|
130 | computeRedondantRanks() ; |
---|
131 | for(auto& rank : rankToRemove_) |
---|
132 | for(auto& element : elements_) element.erase(rank) ; |
---|
133 | } |
---|
134 | else |
---|
135 | { |
---|
136 | computeViewDistribution() ; |
---|
137 | computeConnectorRedundant() ; |
---|
138 | } |
---|
139 | } |
---|
140 | |
---|
141 | /** |
---|
142 | * \brief Compute the connector, i.e. compute the \b elements_ attribute. |
---|
143 | * \detail In this routine we don't eliminate redundant cells as it it performed in |
---|
144 | * computeConnectorMethods. It can be usefull to perform reduce operation over processes. |
---|
145 | In future, some optimisation could be done considering full redondance of the |
---|
146 | source view or the destination view. |
---|
147 | */ |
---|
148 | void CGridRemoteConnector::computeConnectorRedundant(void) |
---|
149 | { |
---|
150 | vector<shared_ptr<CLocalView>> srcView ; |
---|
151 | vector<shared_ptr<CDistributedView>> dstView ; |
---|
152 | vector<int> indElements ; |
---|
153 | elements_.resize(srcView_.size()) ; |
---|
154 | |
---|
155 | bool srcViewsNonDistributed=true ; // not usefull now but later for optimization |
---|
156 | for(int i=0;i<srcView_.size();i++) srcViewsNonDistributed = srcViewsNonDistributed && !isSrcViewDistributed_[i] ; |
---|
157 | |
---|
158 | bool dstViewsNonDistributed=true ; // not usefull now but later for optimization |
---|
159 | for(int i=0;i<dstView_.size();i++) dstViewsNonDistributed = dstViewsNonDistributed && !isDstViewDistributed_[i] ; |
---|
160 | |
---|
161 | for(int i=0;i<srcView_.size();i++) |
---|
162 | { |
---|
163 | srcView.push_back(srcView_[i]) ; |
---|
164 | dstView.push_back(dstView_[i]) ; |
---|
165 | indElements.push_back(i) ; |
---|
166 | } |
---|
167 | |
---|
168 | computeGenericMethod(srcView, dstView, indElements) ; |
---|
169 | |
---|
170 | map<int,bool> ranks ; |
---|
171 | for(auto& it : elements_[indElements[0]]) |
---|
172 | { |
---|
173 | if (it.second.numElements()==0) ranks[it.first] = false ; |
---|
174 | else ranks[it.first] = true ; |
---|
175 | } |
---|
176 | |
---|
177 | } |
---|
178 | |
---|
179 | |
---|
180 | /** |
---|
181 | * \brief Compute the connector, i.e. compute the \b elements_ attribute. |
---|
182 | * \detail In order to achive better optimisation, |
---|
183 | * we distingute the case when the grid is not distributed on source grid (\bcomputeSrcNonDistributed), |
---|
184 | * or the remote grid (\b computeDstNonDistributed), or the both (\b computeSrcDstNonDistributed). |
---|
185 | * Otherwise the generic method is called computeGenericMethod. Note that in the case, if one element view |
---|
186 | * is not distributed on the source and on the remote grid, then we can used the tensorial product |
---|
187 | * property to computing it independently using \b computeSrcDstNonDistributed method. |
---|
188 | * After that, we call the \b removeRedondantRanks method to supress blocks of data that can be sent |
---|
189 | * redondantly the the remote servers |
---|
190 | */ |
---|
191 | void CGridRemoteConnector::computeConnectorMethods(bool reverse) |
---|
192 | { |
---|
193 | vector<shared_ptr<CLocalView>> srcView ; |
---|
194 | vector<shared_ptr<CDistributedView>> dstView ; |
---|
195 | vector<int> indElements ; |
---|
196 | elements_.resize(srcView_.size()) ; |
---|
197 | |
---|
198 | bool srcViewsNonDistributed=true ; |
---|
199 | for(int i=0;i<srcView_.size();i++) srcViewsNonDistributed = srcViewsNonDistributed && !isSrcViewDistributed_[i] ; |
---|
200 | |
---|
201 | bool dstViewsNonDistributed=true ; |
---|
202 | for(int i=0;i<dstView_.size();i++) dstViewsNonDistributed = dstViewsNonDistributed && !isDstViewDistributed_[i] ; |
---|
203 | |
---|
204 | //***************************************************** |
---|
205 | if (srcViewsNonDistributed && dstViewsNonDistributed) |
---|
206 | { |
---|
207 | int commRank, commSize ; |
---|
208 | MPI_Comm_rank(localComm_,&commRank) ; |
---|
209 | MPI_Comm_size(localComm_,&commSize) ; |
---|
210 | |
---|
211 | map<int,bool> ranks ; |
---|
212 | if (reverse) |
---|
213 | { |
---|
214 | int leaderRank=getLeaderRank(remoteSize_, commSize, commRank) ; |
---|
215 | ranks[leaderRank] = true ; |
---|
216 | } |
---|
217 | else |
---|
218 | { |
---|
219 | list<int> remoteRanks; |
---|
220 | list<int> notUsed ; |
---|
221 | computeLeaderProcess(commRank, commSize, remoteSize_, remoteRanks, notUsed) ; |
---|
222 | for(int rank : remoteRanks) ranks[rank]=true ; |
---|
223 | } |
---|
224 | for(int i=0; i<srcView_.size(); i++) computeSrcDstNonDistributed(i,ranks) ; |
---|
225 | } |
---|
226 | |
---|
227 | //***************************************************** |
---|
228 | else if (srcViewsNonDistributed) |
---|
229 | { |
---|
230 | int commRank, commSize ; |
---|
231 | MPI_Comm_rank(localComm_,&commRank) ; |
---|
232 | MPI_Comm_size(localComm_,&commSize) ; |
---|
233 | list<int> remoteRanks; |
---|
234 | list<int> notUsed ; |
---|
235 | map<int,bool> ranks ; |
---|
236 | |
---|
237 | if (reverse) |
---|
238 | { |
---|
239 | shared_ptr<CLocalElement> voidElement = make_shared<CLocalElement>(commRank, 0, CArray<size_t,1>()) ; |
---|
240 | shared_ptr<CLocalView> voidView = make_shared<CLocalView>(voidElement, CElementView::FULL, CArray<int,1>()) ; |
---|
241 | |
---|
242 | for(int i=0;i<srcView_.size();i++) |
---|
243 | if (isDstViewDistributed_[i]) |
---|
244 | { |
---|
245 | if (commRank==0) srcView.push_back(srcView_[i]) ; |
---|
246 | else srcView.push_back(make_shared<CLocalView>(make_shared<CLocalElement>(commRank, srcView_[i]->getGlobalSize(), CArray<size_t,1>()), |
---|
247 | CElementView::FULL, CArray<int,1>())) ; // void view |
---|
248 | dstView.push_back(dstView_[i]) ; |
---|
249 | indElements.push_back(i) ; |
---|
250 | } |
---|
251 | |
---|
252 | computeGenericMethod(srcView, dstView, indElements) ; |
---|
253 | |
---|
254 | for(int i=0;i<srcView_.size();i++) |
---|
255 | if (isDstViewDistributed_[i]) |
---|
256 | { |
---|
257 | size_t sizeElement ; |
---|
258 | int nRank ; |
---|
259 | if (commRank==0) nRank = elements_[i].size() ; |
---|
260 | MPI_Bcast(&nRank, 1, MPI_INT, 0, localComm_) ; |
---|
261 | |
---|
262 | auto it=elements_[i].begin() ; |
---|
263 | for(int j=0;j<nRank;j++) |
---|
264 | { |
---|
265 | int rank ; |
---|
266 | size_t sizeElement ; |
---|
267 | if (commRank==0) { rank = it->first ; sizeElement=it->second.numElements(); } |
---|
268 | MPI_Bcast(&rank, 1, MPI_INT, 0, localComm_) ; |
---|
269 | MPI_Bcast(&sizeElement, 1, MPI_SIZE_T, 0, localComm_) ; |
---|
270 | if (commRank!=0) elements_[i][rank].resize(sizeElement) ; |
---|
271 | MPI_Bcast(elements_[i][rank].dataFirst(), sizeElement, MPI_SIZE_T, 0, localComm_) ; |
---|
272 | if (commRank==0) ++it ; |
---|
273 | } |
---|
274 | } |
---|
275 | |
---|
276 | for(auto& it : elements_[indElements[0]]) |
---|
277 | { |
---|
278 | if (it.second.numElements()==0) ranks[it.first] = false ; |
---|
279 | else ranks[it.first] = true ; |
---|
280 | } |
---|
281 | |
---|
282 | for(int i=0;i<srcView_.size();i++) |
---|
283 | if (!isDstViewDistributed_[i]) computeSrcDstNonDistributed(i, ranks) ; |
---|
284 | |
---|
285 | } |
---|
286 | else |
---|
287 | { |
---|
288 | computeLeaderProcess(commRank, commSize, remoteSize_, remoteRanks, notUsed) ; |
---|
289 | for(int rank : remoteRanks) ranks[rank]=true ; |
---|
290 | |
---|
291 | for(int i=0; i<srcView_.size(); i++) |
---|
292 | { |
---|
293 | if (isDstViewDistributed_[i]) computeSrcNonDistributed(i) ; |
---|
294 | else computeSrcDstNonDistributed(i, ranks) ; |
---|
295 | } |
---|
296 | } |
---|
297 | |
---|
298 | } |
---|
299 | //***************************************************** |
---|
300 | else if (dstViewsNonDistributed) |
---|
301 | { |
---|
302 | int commRank, commSize ; |
---|
303 | MPI_Comm_rank(localComm_,&commRank) ; |
---|
304 | MPI_Comm_size(localComm_,&commSize) ; |
---|
305 | |
---|
306 | map<int,bool> ranks ; |
---|
307 | if (reverse) |
---|
308 | { |
---|
309 | int leaderRank=getLeaderRank(remoteSize_, commSize, commRank) ; |
---|
310 | ranks[leaderRank] = true ; |
---|
311 | } |
---|
312 | else for(int i=0;i<remoteSize_;i++) ranks[i]=true ; |
---|
313 | |
---|
314 | for(int i=0; i<srcView_.size(); i++) |
---|
315 | { |
---|
316 | if (isSrcViewDistributed_[i]) computeDstNonDistributed(i,ranks) ; |
---|
317 | else computeSrcDstNonDistributed(i,ranks) ; |
---|
318 | } |
---|
319 | } |
---|
320 | //***************************************************** |
---|
321 | else |
---|
322 | { |
---|
323 | for(int i=0;i<srcView_.size();i++) |
---|
324 | if (isSrcViewDistributed_[i] || isDstViewDistributed_[i]) |
---|
325 | { |
---|
326 | srcView.push_back(srcView_[i]) ; |
---|
327 | dstView.push_back(dstView_[i]) ; |
---|
328 | indElements.push_back(i) ; |
---|
329 | } |
---|
330 | |
---|
331 | computeGenericMethod(srcView, dstView, indElements) ; |
---|
332 | |
---|
333 | map<int,bool> ranks ; |
---|
334 | for(auto& it : elements_[indElements[0]]) |
---|
335 | { |
---|
336 | if (it.second.numElements()==0) ranks[it.first] = false ; |
---|
337 | else ranks[it.first] = true ; |
---|
338 | } |
---|
339 | |
---|
340 | for(int i=0;i<srcView_.size();i++) |
---|
341 | if (!isSrcViewDistributed_[i] && !isDstViewDistributed_[i]) computeSrcDstNonDistributed(i, ranks) ; |
---|
342 | } |
---|
343 | |
---|
344 | } |
---|
345 | |
---|
346 | |
---|
347 | /** |
---|
348 | * \brief Compute the connector for the element \b i when the source view is not distributed. |
---|
349 | * After the call element_[i] is defined. |
---|
350 | * \param i Indice of the element composing the source grid. |
---|
351 | */ |
---|
352 | |
---|
353 | void CGridRemoteConnector::computeSrcNonDistributed(int i) |
---|
354 | { |
---|
355 | auto& element = elements_[i] ; |
---|
356 | map<int,CArray<size_t,1>> globalIndexView ; |
---|
357 | dstView_[i]->getGlobalIndexView(globalIndexView) ; |
---|
358 | |
---|
359 | CClientClientDHTTemplate<int>::Index2InfoTypeMap dataInfo; |
---|
360 | |
---|
361 | for(auto& it : globalIndexView) |
---|
362 | { |
---|
363 | auto& globalIndex=it.second ; |
---|
364 | for(size_t ind : globalIndex) dataInfo[ind]=it.first ; |
---|
365 | } |
---|
366 | |
---|
367 | // First we feed the distributed hash map with key (remote global index) |
---|
368 | // associated with the value of the remote rank |
---|
369 | CClientClientDHTTemplate<int> DHT(dataInfo, localComm_) ; |
---|
370 | // after we feed the DHT with the local global indices of the source view |
---|
371 | |
---|
372 | int commRank, commSize ; |
---|
373 | MPI_Comm_rank(localComm_,&commRank) ; |
---|
374 | MPI_Comm_size(localComm_,&commSize) ; |
---|
375 | CArray<size_t,1> srcIndex ; |
---|
376 | // like the source view is not distributed, then only the rank 0 need to feed the DHT |
---|
377 | if (commRank==0) srcView_[i]->getGlobalIndexView(srcIndex) ; |
---|
378 | |
---|
379 | // compute the mapping |
---|
380 | DHT.computeIndexInfoMapping(srcIndex) ; |
---|
381 | auto& returnInfo = DHT.getInfoIndexMap() ; |
---|
382 | |
---|
383 | // returnInfo contains now the map for each global indices to send to a list of remote rank |
---|
384 | // only for the rank=0 because it is the one to feed the DHT |
---|
385 | // so it need to send the list to each server leader i.e. the local process that handle specifically one or more |
---|
386 | // servers |
---|
387 | |
---|
388 | // rankIndGlo : rankIndGlo[rank][indGlo] : list of indice to send the the remote server of rank "rank" |
---|
389 | vector<vector<size_t>> rankIndGlo(remoteSize_) ; |
---|
390 | if (commRank==0) |
---|
391 | for(auto& it1 : returnInfo) |
---|
392 | for(auto& it2 : it1.second) rankIndGlo[it2].push_back(it1.first) ; |
---|
393 | |
---|
394 | |
---|
395 | vector<MPI_Request> requests ; |
---|
396 | |
---|
397 | if (commRank==0) |
---|
398 | { |
---|
399 | requests.resize(remoteSize_) ; |
---|
400 | for(int i=0 ; i<remoteSize_;i++) |
---|
401 | { |
---|
402 | // ok send only the global indices for a server to the "server leader" |
---|
403 | int rank = getLeaderRank(commSize, remoteSize_, i) ; |
---|
404 | MPI_Isend(rankIndGlo[i].data(), rankIndGlo[i].size(), MPI_SIZE_T, rank, i ,localComm_, &requests[i]) ; |
---|
405 | } |
---|
406 | } |
---|
407 | |
---|
408 | list<int> remoteRanks; |
---|
409 | list<int> notUsed ; |
---|
410 | // I am a server leader of which remote ranks ? |
---|
411 | computeLeaderProcess(commRank, commSize, remoteSize_, remoteRanks, notUsed) ; |
---|
412 | |
---|
413 | for(auto remoteRank : remoteRanks) |
---|
414 | { |
---|
415 | MPI_Status status ; |
---|
416 | int size ; |
---|
417 | MPI_Probe(0,remoteRank,localComm_, &status); |
---|
418 | MPI_Get_count(&status, MPI_SIZE_T, &size) ; |
---|
419 | elements_[i][remoteRank].resize(size) ; |
---|
420 | // for each remote ranks receive the global indices from proc 0 |
---|
421 | MPI_Recv(elements_[i][remoteRank].dataFirst(),size, MPI_SIZE_T,0,remoteRank, localComm_,&status) ; |
---|
422 | } |
---|
423 | |
---|
424 | if (commRank==0) |
---|
425 | { |
---|
426 | vector<MPI_Status> status(remoteSize_) ; |
---|
427 | // asynchronous for sender, wait for completion |
---|
428 | MPI_Waitall(remoteSize_, requests.data(), status.data()) ; |
---|
429 | } |
---|
430 | } |
---|
431 | |
---|
432 | /** |
---|
433 | * \brief Compute the connector for the element \b i when the source view is not distributed. |
---|
434 | * After the call element_[i] is defined. |
---|
435 | * \param i Indice of the element composing the source grid. |
---|
436 | */ |
---|
437 | |
---|
438 | void CGridRemoteConnector::computeSrcNonDistributedReverse(int i) |
---|
439 | { |
---|
440 | auto& element = elements_[i] ; |
---|
441 | map<int,CArray<size_t,1>> globalIndexView ; |
---|
442 | dstView_[i]->getGlobalIndexView(globalIndexView) ; |
---|
443 | |
---|
444 | CClientClientDHTTemplate<int>::Index2InfoTypeMap dataInfo; |
---|
445 | |
---|
446 | for(auto& it : globalIndexView) |
---|
447 | { |
---|
448 | auto& globalIndex=it.second ; |
---|
449 | for(size_t ind : globalIndex) dataInfo[ind]=it.first ; |
---|
450 | } |
---|
451 | |
---|
452 | // First we feed the distributed hash map with key (remote global index) |
---|
453 | // associated with the value of the remote rank |
---|
454 | CClientClientDHTTemplate<int> DHT(dataInfo, localComm_) ; |
---|
455 | // after we feed the DHT with the local global indices of the source view |
---|
456 | |
---|
457 | int commRank, commSize ; |
---|
458 | MPI_Comm_rank(localComm_,&commRank) ; |
---|
459 | MPI_Comm_size(localComm_,&commSize) ; |
---|
460 | CArray<size_t,1> srcIndex ; |
---|
461 | // like the source view is not distributed, then only the rank 0 need to feed the DHT |
---|
462 | if (commRank==0) srcView_[i]->getGlobalIndexView(srcIndex) ; |
---|
463 | |
---|
464 | // compute the mapping |
---|
465 | DHT.computeIndexInfoMapping(srcIndex) ; |
---|
466 | auto& returnInfo = DHT.getInfoIndexMap() ; |
---|
467 | |
---|
468 | // returnInfo contains now the map for each global indices to send to a list of remote rank |
---|
469 | // only for the rank=0 because it is the one to feed the DHT |
---|
470 | // so it need to send the list to each server leader i.e. the local process that handle specifically one or more |
---|
471 | // servers |
---|
472 | |
---|
473 | // rankIndGlo : rankIndGlo[rank][indGlo] : list of indice to send the the remote server of rank "rank" |
---|
474 | vector<vector<size_t>> rankIndGlo(remoteSize_) ; |
---|
475 | if (commRank==0) |
---|
476 | for(auto& it1 : returnInfo) |
---|
477 | for(auto& it2 : it1.second) rankIndGlo[it2].push_back(it1.first) ; |
---|
478 | |
---|
479 | // bcast the same for each client |
---|
480 | for(int remoteRank=0 ; remoteRank<remoteSize_ ; remoteRank++) |
---|
481 | { |
---|
482 | int remoteDataSize ; |
---|
483 | if (commRank==0) remoteDataSize = rankIndGlo[remoteRank].size() ; |
---|
484 | MPI_Bcast(&remoteDataSize, 1, MPI_INT, 0, localComm_) ; |
---|
485 | |
---|
486 | auto& element = elements_[i][remoteRank] ; |
---|
487 | element.resize(remoteDataSize) ; |
---|
488 | if (commRank==0) for(int j=0 ; j<remoteDataSize; j++) element(j)=rankIndGlo[remoteRank][j] ; |
---|
489 | MPI_Bcast(element.dataFirst(), remoteDataSize, MPI_SIZE_T, 0, localComm_) ; |
---|
490 | } |
---|
491 | } |
---|
492 | |
---|
493 | |
---|
494 | |
---|
495 | /** |
---|
496 | * \brief Compute the remote connector for the element \b i when the remote view is not distributed. |
---|
497 | * After the call, element_[i] is defined. |
---|
498 | * \param i Indice of the element composing the remote grid. |
---|
499 | * \param ranks The list of rank for which the local proc is in charge to compute the connector |
---|
500 | * (if leader server for exemple). if ranks[rank] == false the corresponding elements_ |
---|
501 | * is set to void array (no data to sent) just in order to notify corresponding remote server |
---|
502 | * that the call is collective with each other one |
---|
503 | */ |
---|
504 | void CGridRemoteConnector::computeDstNonDistributed(int i, map<int,bool>& ranks) |
---|
505 | { |
---|
506 | auto& element = elements_[i] ; |
---|
507 | map<int,CArray<size_t,1>> globalIndexView ; |
---|
508 | dstView_[i]->getGlobalIndexView(globalIndexView) ; |
---|
509 | |
---|
510 | |
---|
511 | CClientClientDHTTemplate<int>::Index2InfoTypeMap dataInfo; |
---|
512 | |
---|
513 | // First we feed the distributed hash map with key (remote global index) |
---|
514 | // associated with the value of the remote rank |
---|
515 | for(auto& it : globalIndexView) |
---|
516 | if (it.first==0) // since the remote view is not distributed, insert only the remote rank 0 |
---|
517 | { |
---|
518 | auto& globalIndex=it.second ; |
---|
519 | for(size_t ind : globalIndex) dataInfo[ind]=0 ; // associated the the rank 0 |
---|
520 | } |
---|
521 | |
---|
522 | CClientClientDHTTemplate<int> DHT(dataInfo, localComm_) ; |
---|
523 | // after we feed the DHT with the local global indices of the source view |
---|
524 | |
---|
525 | CArray<size_t,1> srcIndex ; |
---|
526 | srcView_[i]->getGlobalIndexView(srcIndex) ; |
---|
527 | DHT.computeIndexInfoMapping(srcIndex) ; |
---|
528 | auto& returnInfo = DHT.getInfoIndexMap() ; |
---|
529 | |
---|
530 | // returnInfo contains now the map for each global indices to send to a list of remote rank |
---|
531 | // now construct the element_ list of global indices for each rank in my list except if the erray must be empty |
---|
532 | for (auto& rank : ranks) |
---|
533 | { |
---|
534 | if (rank.second) // non empty array => for rank that have not any data to be received |
---|
535 | { |
---|
536 | int size=0 ; |
---|
537 | for(auto& it : returnInfo) if (!it.second.empty()) size++ ; |
---|
538 | auto& array = element[rank.first] ; |
---|
539 | array.resize(size) ; |
---|
540 | size=0 ; |
---|
541 | for(auto& it : returnInfo) |
---|
542 | if (!it.second.empty()) |
---|
543 | { |
---|
544 | array(size)=it.first ; |
---|
545 | size++ ; |
---|
546 | } |
---|
547 | } |
---|
548 | else element[rank.first] = CArray<size_t,1>(0) ; // empty array => for rank that have not any data to be received |
---|
549 | } |
---|
550 | } |
---|
551 | |
---|
552 | /** |
---|
553 | * \brief Compute the remote connector for the element \b i when the source and the remote view are not distributed. |
---|
554 | * After the call, element_[i] is defined. |
---|
555 | * \param i Indice of the element composing the remote grid. |
---|
556 | * \param ranks The list of rank for which the local proc is in charge to compute the connector |
---|
557 | * (if leader server for exemple). if ranks[rank] == false the corresponding elements_ |
---|
558 | * is set to void array (no data to sent) just in order to notify corresponding remote server |
---|
559 | * that the call is collective with each other one |
---|
560 | */ |
---|
561 | |
---|
562 | void CGridRemoteConnector::computeSrcDstNonDistributed(int i, map<int,bool>& ranks) |
---|
563 | { |
---|
564 | auto& element = elements_[i] ; |
---|
565 | map<int,CArray<size_t,1>> globalIndexView ; |
---|
566 | dstView_[i]->getGlobalIndexView(globalIndexView) ; |
---|
567 | |
---|
568 | |
---|
569 | CClientClientDHTTemplate<int>::Index2InfoTypeMap dataInfo; |
---|
570 | // First we feed the distributed hash map with key (remote global index) |
---|
571 | // associated with the value of the remote rank |
---|
572 | |
---|
573 | for(auto& it : globalIndexView) |
---|
574 | if (it.first==0) // insert only the remote rank 0 since the remote view is not distributed |
---|
575 | { |
---|
576 | auto& globalIndex=it.second ; |
---|
577 | for(size_t ind : globalIndex) dataInfo[ind]=0 ; // associated the the rank 0 |
---|
578 | } |
---|
579 | |
---|
580 | CClientClientDHTTemplate<int> DHT(dataInfo, localComm_) ; |
---|
581 | // after we feed the DHT with the local global indices of the source view |
---|
582 | |
---|
583 | int commRank, commSize ; |
---|
584 | MPI_Comm_rank(localComm_,&commRank) ; |
---|
585 | MPI_Comm_size(localComm_,&commSize) ; |
---|
586 | CArray<size_t,1> srcIndex ; |
---|
587 | |
---|
588 | // like the source view is not distributed, then only the rank 0 need to feed the DHT |
---|
589 | if (commRank==0) srcView_[i]->getGlobalIndexView(srcIndex) ; |
---|
590 | DHT.computeIndexInfoMapping(srcIndex) ; |
---|
591 | auto& returnInfo = DHT.getInfoIndexMap() ; |
---|
592 | |
---|
593 | vector<size_t> indGlo ; |
---|
594 | if (commRank==0) |
---|
595 | for(auto& it1 : returnInfo) |
---|
596 | for(auto& it2 : it1.second) indGlo.push_back(it1.first) ; |
---|
597 | |
---|
598 | // now local rank 0 know which indices to send to remote rank 0, but all the server |
---|
599 | // must receive the same information. So only the leader rank will sent this. |
---|
600 | // So local rank 0 must broadcast the information to all leader. |
---|
601 | // for this we create a new communicator composed of local process that must send data |
---|
602 | // to a remote rank, data are broadcasted, and element_[i] is construction for each remote |
---|
603 | // rank in charge |
---|
604 | int color=0 ; |
---|
605 | if (ranks.empty()) color=0 ; |
---|
606 | else color=1 ; |
---|
607 | if (commRank==0) color=1 ; |
---|
608 | MPI_Comm newComm ; |
---|
609 | MPI_Comm_split(localComm_, color, commRank, &newComm) ; |
---|
610 | if (color==1) |
---|
611 | { |
---|
612 | // ok, I am part of the process that must send something to one or more remote server |
---|
613 | // so I get the list of global indices from rank 0 |
---|
614 | int dataSize ; |
---|
615 | if (commRank==0) dataSize=indGlo.size() ; |
---|
616 | MPI_Bcast(&dataSize,1,MPI_INT, 0, newComm) ; |
---|
617 | indGlo.resize(dataSize) ; |
---|
618 | MPI_Bcast(indGlo.data(),dataSize,MPI_SIZE_T,0,newComm) ; |
---|
619 | } |
---|
620 | MPI_Comm_free(&newComm) ; |
---|
621 | |
---|
622 | // construct element_[i] from indGlo |
---|
623 | for(auto& rank : ranks) |
---|
624 | { |
---|
625 | if (rank.second) |
---|
626 | { |
---|
627 | int dataSize=indGlo.size(); |
---|
628 | auto& element = elements_[i][rank.first] ; |
---|
629 | element.resize(dataSize) ; |
---|
630 | for(int i=0;i<dataSize; i++) element(i)=indGlo[i] ; |
---|
631 | } |
---|
632 | else element[rank.first] = CArray<size_t,1>(0) ; |
---|
633 | } |
---|
634 | |
---|
635 | } |
---|
636 | |
---|
637 | |
---|
638 | /** |
---|
639 | * \brief Generic method the compute the grid remote connector. Only distributed elements are specifed in the source view and remote view. |
---|
640 | * Connector for non distributed elements are computed separatly to improve performance and memory consumption. After the call, |
---|
641 | * \b elements_ is defined. |
---|
642 | * \param srcView List of the source views composing the grid, without non distributed views |
---|
643 | * \param dstView List of the remote views composing the grid, without non distributed views |
---|
644 | * \param indElements Index of the view making the correspondance between all views and views distributed (that are in input) |
---|
645 | */ |
---|
646 | void CGridRemoteConnector::computeGenericMethod(vector<shared_ptr<CLocalView>>& srcView, vector<shared_ptr<CDistributedView>>& dstView, vector<int>& indElements) |
---|
647 | { |
---|
648 | // generic method, every element can be distributed |
---|
649 | int nDst = dstView.size() ; |
---|
650 | vector<size_t> dstSliceSize(nDst) ; |
---|
651 | dstSliceSize[0] = 1 ; |
---|
652 | for(int i=1; i<nDst; i++) dstSliceSize[i] = dstView[i-1]->getGlobalSize()*dstSliceSize[i-1] ; |
---|
653 | |
---|
654 | CClientClientDHTTemplate<int>::Index2VectorInfoTypeMap dataInfo ; |
---|
655 | CClientClientDHTTemplate<size_t>::Index2VectorInfoTypeMap info ; // info map |
---|
656 | |
---|
657 | // first, we need to feed the DHT with the global index of the remote server |
---|
658 | // for that : |
---|
659 | // First the first element insert the in a DHT with key as the rank and value the list of global index associated |
---|
660 | // Then get the previously stored index associate with the remote rank I am in charge and reinsert the global index |
---|
661 | // corresponding to the position of the element in the remote view suing tensorial product |
---|
662 | // finaly we get only the list of remote global index I am in charge for the whole remote grid |
---|
663 | |
---|
664 | for(int pos=0; pos<nDst; pos++) |
---|
665 | { |
---|
666 | size_t sliceSize=dstSliceSize[pos] ; |
---|
667 | map<int,CArray<size_t,1>> globalIndexView ; |
---|
668 | dstView[pos]->getGlobalIndexView(globalIndexView) ; |
---|
669 | |
---|
670 | CClientClientDHTTemplate<size_t>::Index2VectorInfoTypeMap lastInfo(info) ; |
---|
671 | |
---|
672 | if (pos>0) |
---|
673 | { |
---|
674 | CArray<size_t,1> ranks(globalIndexView.size()) ; |
---|
675 | auto it=globalIndexView.begin() ; |
---|
676 | for(int i=0 ; i<ranks.numElements();i++,it++) ranks(i)=it->first ; |
---|
677 | CClientClientDHTTemplate<size_t> dataRanks(info, localComm_) ; |
---|
678 | dataRanks.computeIndexInfoMapping(ranks) ; |
---|
679 | lastInfo = dataRanks.getInfoIndexMap() ; |
---|
680 | } |
---|
681 | |
---|
682 | info.clear() ; |
---|
683 | for(auto& it : globalIndexView) |
---|
684 | { |
---|
685 | int rank = it.first ; |
---|
686 | auto& globalIndex = it.second ; |
---|
687 | auto& inf = info[rank] ; |
---|
688 | if (pos==0) for(int i=0;i<globalIndex.numElements();i++) inf.push_back(globalIndex(i)) ; |
---|
689 | else |
---|
690 | { |
---|
691 | auto& lastGlobalIndex = lastInfo[rank] ; |
---|
692 | for(size_t lastGlobalInd : lastGlobalIndex) |
---|
693 | { |
---|
694 | for(int i=0;i<globalIndex.numElements();i++) inf.push_back(globalIndex(i)*sliceSize+lastGlobalInd) ; |
---|
695 | } |
---|
696 | } |
---|
697 | } |
---|
698 | |
---|
699 | if (pos==nDst-1) |
---|
700 | { |
---|
701 | for(auto& it : info) |
---|
702 | { |
---|
703 | int rank=it.first ; |
---|
704 | auto& globalIndex = it.second ; |
---|
705 | for(auto globalInd : globalIndex) dataInfo[globalInd].push_back(rank) ; |
---|
706 | } |
---|
707 | } |
---|
708 | } |
---|
709 | |
---|
710 | // we feed the DHT with the remote global index |
---|
711 | CClientClientDHTTemplate<int> dataRanks(dataInfo, localComm_) ; |
---|
712 | |
---|
713 | // generate list of global index for src view |
---|
714 | int nSrc = srcView.size() ; |
---|
715 | vector<size_t> srcSliceSize(nSrc) ; |
---|
716 | |
---|
717 | srcSliceSize[0] = 1 ; |
---|
718 | for(int i=1; i<nSrc; i++) srcSliceSize[i] = srcView[i-1]->getGlobalSize()*srcSliceSize[i-1] ; |
---|
719 | |
---|
720 | vector<size_t> srcGlobalIndex ; |
---|
721 | size_t sliceIndex=0 ; |
---|
722 | srcView[nSrc-1]->getGlobalIndex(srcGlobalIndex, sliceIndex, srcSliceSize.data(), srcView.data(), nSrc-1) ; |
---|
723 | // now we have the global index of the source grid in srcGlobalIndex |
---|
724 | // we feed the DHT with the src global index (if we have) |
---|
725 | if (srcGlobalIndex.size()>0) |
---|
726 | { |
---|
727 | CArray<size_t,1> srcGlobalIndexArray(srcGlobalIndex.data(), shape(srcGlobalIndex.size()),neverDeleteData) ; |
---|
728 | dataRanks.computeIndexInfoMapping(srcGlobalIndexArray) ; |
---|
729 | } |
---|
730 | else |
---|
731 | { |
---|
732 | CArray<size_t,1> srcGlobalIndexArray ; |
---|
733 | dataRanks.computeIndexInfoMapping(srcGlobalIndexArray) ; |
---|
734 | } |
---|
735 | const auto& returnInfo = dataRanks.getInfoIndexMap() ; |
---|
736 | // returnInfo contains now the map for each global indices to send to a list of remote rank |
---|
737 | // but we want to use the tensorial product property to get the same information using only global |
---|
738 | // index of element view. So the idea is to reverse the information : for a global index of the grid |
---|
739 | // to send to the remote server, what is the global index of each element composing the grid ? |
---|
740 | |
---|
741 | vector<map<int, set<size_t>>> elements(nSrc) ; // internal representation of elements composing the grid |
---|
742 | |
---|
743 | for(auto& indRanks : returnInfo) |
---|
744 | { |
---|
745 | size_t gridIndexGlo=indRanks.first ; |
---|
746 | auto& ranks = indRanks.second ; |
---|
747 | for(int i=nSrc-1; i>=0; i--) |
---|
748 | { |
---|
749 | auto& element = elements[i] ; |
---|
750 | size_t localIndGlo = gridIndexGlo / srcSliceSize[i] ; |
---|
751 | gridIndexGlo = gridIndexGlo % srcSliceSize[i] ; |
---|
752 | for(int rank : ranks) element[rank].insert(localIndGlo) ; |
---|
753 | } |
---|
754 | } |
---|
755 | |
---|
756 | // elements_.resize(nSrc) ; |
---|
757 | for(int i=0 ; i<nSrc; i++) |
---|
758 | { |
---|
759 | auto& element=elements[i] ; |
---|
760 | for(auto& rankInd : element) |
---|
761 | { |
---|
762 | int rank=rankInd.first ; |
---|
763 | set<size_t>& indGlo = rankInd.second ; |
---|
764 | CArray<size_t,1>& indGloArray = elements_[indElements[i]][rank] ; |
---|
765 | indGloArray.resize(indGlo.size()) ; |
---|
766 | int j=0 ; |
---|
767 | for (auto index : indGlo) { indGloArray(j) = index ; j++; } |
---|
768 | } |
---|
769 | } |
---|
770 | |
---|
771 | // So what about when there is some server that have no data to receive |
---|
772 | // they must be inform they receive an event with no data. |
---|
773 | // So find remote servers with no data, and one client will take in charge |
---|
774 | // that it receive global index with no data (0-size) |
---|
775 | vector<int> ranks(remoteSize_,0) ; |
---|
776 | for(auto& it : elements_[indElements[0]]) ranks[it.first] = 1 ; |
---|
777 | MPI_Allreduce(MPI_IN_PLACE, ranks.data(), remoteSize_, MPI_INT, MPI_SUM, localComm_) ; |
---|
778 | int commRank, commSize ; |
---|
779 | MPI_Comm_rank(localComm_,&commRank) ; |
---|
780 | MPI_Comm_size(localComm_,&commSize) ; |
---|
781 | int pos=0 ; |
---|
782 | for(int i=0; i<remoteSize_ ; i++) |
---|
783 | if (ranks[i]==0) |
---|
784 | { |
---|
785 | if (pos%commSize==commRank) |
---|
786 | for(int j=0 ; j<nSrc; j++) elements_[indElements[j]][i] = CArray<size_t,1>(0) ; |
---|
787 | pos++ ; |
---|
788 | } |
---|
789 | } |
---|
790 | |
---|
791 | /** |
---|
792 | * \brief Once the connector is computed (compute \b elements_), redondant data can be avoid to be sent to the server. |
---|
793 | * This call compute the redondant rank and store them in \b rankToRemove_ attribute. |
---|
794 | * The goal of this method is to make a hash of each block of indice that determine wich data to send to a |
---|
795 | * of a specific server rank using a hash method. So data to send to a rank is associated to a hash. |
---|
796 | * After we compare hash between local rank and remove redondant data corresponding to the same hash. |
---|
797 | */ |
---|
798 | void CGridRemoteConnector::computeRedondantRanks(bool reverse) |
---|
799 | { |
---|
800 | int commRank ; |
---|
801 | MPI_Comm_rank(localComm_,&commRank) ; |
---|
802 | |
---|
803 | set<int> ranks; |
---|
804 | for(auto& element : elements_) |
---|
805 | for(auto& it : element) ranks.insert(it.first) ; |
---|
806 | |
---|
807 | for(auto& element : elements_) |
---|
808 | for(auto& it : element) |
---|
809 | if (ranks.count(it.first)==0) ERROR("void CGridRemoteConnector::removeRedondantRanks(void)",<<"number of ranks in elements is not coherent between each element") ; |
---|
810 | |
---|
811 | HashXIOS<size_t> hashGlobalIndex; |
---|
812 | |
---|
813 | map<int,size_t> hashRanks ; |
---|
814 | for(auto& element : elements_) |
---|
815 | for(auto& it : element) |
---|
816 | { |
---|
817 | auto& globalIndex=it.second ; |
---|
818 | int rank=it.first ; |
---|
819 | size_t hash ; |
---|
820 | hash=0 ; |
---|
821 | for(int i=0; i<globalIndex.numElements(); i++) hash+=hashGlobalIndex(globalIndex(i)) ; |
---|
822 | if (globalIndex.numElements()>0) |
---|
823 | { |
---|
824 | if (hashRanks.count(rank)==0) hashRanks[rank]=hash ; |
---|
825 | else hashRanks[rank]=hashGlobalIndex.hashCombine(hashRanks[rank],hash) ; |
---|
826 | } |
---|
827 | } |
---|
828 | |
---|
829 | if (reverse) |
---|
830 | { |
---|
831 | set<size_t> hashs ; |
---|
832 | //easy because local |
---|
833 | for(auto& hashRank : hashRanks) |
---|
834 | { |
---|
835 | if (hashs.count(hashRank.second)==0) hashs.insert(hashRank.second) ; |
---|
836 | else rankToRemove_.insert(hashRank.first) ; |
---|
837 | } |
---|
838 | |
---|
839 | } |
---|
840 | else |
---|
841 | { |
---|
842 | // a hash is now computed for data block I will sent to the server. |
---|
843 | |
---|
844 | CClientClientDHTTemplate<int>::Index2InfoTypeMap info ; |
---|
845 | |
---|
846 | map<size_t,int> hashRank ; |
---|
847 | HashXIOS<int> hashGlobalIndexRank; |
---|
848 | for(auto& it : hashRanks) |
---|
849 | { |
---|
850 | it.second = hashGlobalIndexRank.hashCombine(it.first,it.second) ; |
---|
851 | info[it.second]=commRank ; |
---|
852 | hashRank[it.second]=it.first ; |
---|
853 | } |
---|
854 | |
---|
855 | // we feed a DHT map with key : hash, value : myrank |
---|
856 | CClientClientDHTTemplate<int> dataHash(info, localComm_) ; |
---|
857 | CArray<size_t,1> hashList(hashRank.size()) ; |
---|
858 | |
---|
859 | int i=0 ; |
---|
860 | for(auto& it : hashRank) { hashList(i)=it.first ; i++; } |
---|
861 | |
---|
862 | // now who are the ranks that have the same hash : feed the DHT with my list of hash |
---|
863 | dataHash.computeIndexInfoMapping(hashList) ; |
---|
864 | auto& hashRankList = dataHash.getInfoIndexMap() ; |
---|
865 | |
---|
866 | |
---|
867 | for(auto& it : hashRankList) |
---|
868 | { |
---|
869 | size_t hash = it.first ; |
---|
870 | auto& ranks = it.second ; |
---|
871 | |
---|
872 | bool first=true ; |
---|
873 | // only the process with the lowest rank get in charge of sendinf data to remote server |
---|
874 | for(int rank : ranks) if (commRank>rank) first=false ; |
---|
875 | if (!first) rankToRemove_.insert(hashRank[hash]) ; |
---|
876 | } |
---|
877 | } |
---|
878 | } |
---|
879 | |
---|
880 | } |
---|