[1918] | 1 | #include "grid_remote_connector.hpp" |
---|
| 2 | #include "client_client_dht_template.hpp" |
---|
[2179] | 3 | #include "leader_process.hpp" |
---|
[1938] | 4 | #include "mpi.hpp" |
---|
[1918] | 5 | |
---|
| 6 | |
---|
| 7 | |
---|
| 8 | namespace xios |
---|
| 9 | { |
---|
[2179] | 10 | /** |
---|
| 11 | * \brief class constructor. |
---|
| 12 | * \param srcView List of sources views. |
---|
| 13 | * \param dstView List of remotes views. |
---|
| 14 | * \param localComm Local MPI communicator |
---|
| 15 | * \param remoteSize Size of the remote communicator |
---|
| 16 | */ |
---|
[2267] | 17 | CGridRemoteConnector::CGridRemoteConnector(vector<shared_ptr<CLocalView>>& srcView, vector<shared_ptr<CDistributedView>>& dstView, MPI_Comm localComm, int remoteSize) |
---|
[1938] | 18 | : srcView_(srcView), dstView_(dstView), localComm_(localComm), remoteSize_(remoteSize) |
---|
[1918] | 19 | {} |
---|
| 20 | |
---|
[2179] | 21 | /** |
---|
| 22 | * \brief class constructor. |
---|
| 23 | * \param srcView List of sources views. |
---|
| 24 | * \param dstView List of remotes views. |
---|
| 25 | * \param localComm Local MPI communicator |
---|
| 26 | * \param remoteSize Size of the remote communicator |
---|
| 27 | */ |
---|
[2267] | 28 | CGridRemoteConnector::CGridRemoteConnector(vector<shared_ptr<CLocalView>>& srcView, vector< shared_ptr<CLocalView> >& dstView, MPI_Comm localComm, int remoteSize) |
---|
[1999] | 29 | : srcView_(srcView), localComm_(localComm), remoteSize_(remoteSize) |
---|
| 30 | { |
---|
[2267] | 31 | for(auto& it : dstView) dstView_.push_back((shared_ptr<CDistributedView>) it) ; |
---|
[1999] | 32 | } |
---|
| 33 | |
---|
[2179] | 34 | |
---|
| 35 | /** |
---|
| 36 | * \brief Compute if each view composing the source grid and the remote grid is distributed or not. |
---|
| 37 | * Result is stored on internal attributes \b isSrcViewDistributed_ and \b isDstViewDistributed_. |
---|
| 38 | * \detail To compute this, a hash is computed for each array on indices. The hash must permutable, i.e. |
---|
| 39 | * the order of the list of global indices doesn't influence the value of the hash. So simply a sum of |
---|
| 40 | * hash of each indices is used for the whole array. After, the computed hash are compared with each other |
---|
| 41 | * ranks of \b localComm_ MPI communicator using an MPI_ALLReduce. If, for each ranks, the hash is the same |
---|
| 42 | * then the view is not distributed |
---|
| 43 | */ |
---|
| 44 | void CGridRemoteConnector::computeViewDistribution(void) |
---|
| 45 | { |
---|
| 46 | HashXIOS<size_t> hashGlobalIndex; // hash function-object |
---|
| 47 | |
---|
| 48 | int nDst = dstView_.size() ; |
---|
| 49 | vector<size_t> hashRank(remoteSize_) ; |
---|
[2296] | 50 | vector<size_t> sizeRank(remoteSize_) ; |
---|
[2179] | 51 | isDstViewDistributed_.resize(nDst) ; |
---|
| 52 | |
---|
| 53 | for(int i=0; i<nDst; i++) |
---|
| 54 | { |
---|
| 55 | map<int,CArray<size_t,1>> globalIndexView ; |
---|
| 56 | dstView_[i]->getGlobalIndexView(globalIndexView) ; |
---|
| 57 | hashRank.assign(remoteSize_,0) ; // everybody ranks to 0 except rank of the remote view I have |
---|
| 58 | // that would be assign to my local hash |
---|
[2296] | 59 | sizeRank.assign(remoteSize_,0) ; |
---|
| 60 | |
---|
[2179] | 61 | for(auto& it : globalIndexView) |
---|
| 62 | { |
---|
| 63 | int rank=it.first ; |
---|
| 64 | CArray<size_t,1>& globalIndex = it.second ; |
---|
| 65 | size_t globalIndexSize = globalIndex.numElements(); |
---|
| 66 | size_t hashValue=0 ; |
---|
| 67 | for(size_t ind=0;ind<globalIndexSize;ind++) hashValue += hashGlobalIndex(globalIndex(ind)) ; |
---|
| 68 | hashRank[rank] += hashValue ; |
---|
[2296] | 69 | sizeRank[rank] += globalIndexSize ; |
---|
[2179] | 70 | } |
---|
| 71 | // sum all the hash for every process of the local comm. The reduce is on the size of remote view (remoteSize_) |
---|
| 72 | // after that for each rank of the remote view, we get the hash |
---|
| 73 | MPI_Allreduce(MPI_IN_PLACE, hashRank.data(), remoteSize_, MPI_SIZE_T, MPI_SUM, localComm_) ; |
---|
[2296] | 74 | MPI_Allreduce(MPI_IN_PLACE, sizeRank.data(), remoteSize_, MPI_SIZE_T, MPI_SUM, localComm_) ; |
---|
[2179] | 75 | size_t value = hashRank[0] ; |
---|
[2296] | 76 | size_t size = sizeRank[0] ; |
---|
[2179] | 77 | isDstViewDistributed_[i]=false ; |
---|
| 78 | for(int j=0 ; j<remoteSize_ ; j++) |
---|
[2296] | 79 | if (size!=sizeRank[j] || value != hashRank[j]) |
---|
[2179] | 80 | { |
---|
| 81 | isDstViewDistributed_[i]=true ; |
---|
| 82 | break ; |
---|
| 83 | } |
---|
| 84 | } |
---|
| 85 | |
---|
| 86 | int nSrc = srcView_.size() ; |
---|
| 87 | int commSize,commRank ; |
---|
| 88 | MPI_Comm_size(localComm_,&commSize) ; |
---|
| 89 | MPI_Comm_rank(localComm_,&commRank) ; |
---|
| 90 | hashRank.resize(commSize,0) ; |
---|
| 91 | isSrcViewDistributed_.resize(nSrc) ; |
---|
| 92 | |
---|
| 93 | for(int i=0; i<nSrc; i++) |
---|
| 94 | { |
---|
| 95 | CArray<size_t,1> globalIndex ; |
---|
| 96 | srcView_[i]->getGlobalIndexView(globalIndex) ; |
---|
| 97 | hashRank.assign(commSize,0) ; // 0 for everybody except my rank |
---|
| 98 | size_t globalIndexSize = globalIndex.numElements() ; |
---|
[2296] | 99 | |
---|
| 100 | size_t allEqual ; |
---|
| 101 | MPI_Allreduce(&globalIndexSize, &allEqual, 1, MPI_SIZE_T, MPI_BXOR, localComm_) ; |
---|
| 102 | if (allEqual!=0) |
---|
| 103 | { |
---|
| 104 | isSrcViewDistributed_[i]=true ; |
---|
| 105 | break ; |
---|
| 106 | } |
---|
| 107 | |
---|
| 108 | // warning : jenkins hash : 0 --> 0 : need to compare number of element for each ranks |
---|
[2179] | 109 | size_t hashValue=0 ; |
---|
| 110 | for(size_t ind=0;ind<globalIndexSize;ind++) hashValue += hashGlobalIndex(globalIndex(ind)) ; |
---|
[2296] | 111 | MPI_Allreduce(&hashValue, &allEqual, 1, MPI_SIZE_T, MPI_BXOR, localComm_) ; |
---|
| 112 | if (allEqual!=0) isSrcViewDistributed_[i]=true ; |
---|
| 113 | else isSrcViewDistributed_[i]=false ; |
---|
[2179] | 114 | } |
---|
| 115 | |
---|
| 116 | } |
---|
| 117 | |
---|
| 118 | /** |
---|
| 119 | * \brief Compute the connector, i.e. compute the \b elements_ attribute. |
---|
[2236] | 120 | * \detail Depending of the distributions of the view computed in the computeViewDistribution() call, the connector is computed in computeConnectorMethods(), and to achieve better optimisation |
---|
| 121 | * some redondant ranks can be removed from the elements_ map. |
---|
| 122 | */ |
---|
[2291] | 123 | void CGridRemoteConnector::computeConnector(bool eliminateRedundant) |
---|
[2236] | 124 | { |
---|
[2291] | 125 | if (eliminateRedundant) |
---|
| 126 | { |
---|
| 127 | computeViewDistribution() ; |
---|
| 128 | computeConnectorMethods() ; |
---|
| 129 | computeRedondantRanks() ; |
---|
| 130 | for(auto& rank : rankToRemove_) |
---|
| 131 | for(auto& element : elements_) element.erase(rank) ; |
---|
| 132 | } |
---|
| 133 | else |
---|
| 134 | { |
---|
| 135 | computeViewDistribution() ; |
---|
| 136 | computeConnectorRedundant() ; |
---|
| 137 | } |
---|
[2236] | 138 | } |
---|
[2291] | 139 | |
---|
[2236] | 140 | /** |
---|
| 141 | * \brief Compute the connector, i.e. compute the \b elements_ attribute. |
---|
[2291] | 142 | * \detail In this routine we don't eliminate redundant cells as it it performed in |
---|
| 143 | * computeConnectorMethods. It can be usefull to perform reduce operation over processes. |
---|
| 144 | In future, some optimisation could be done considering full redondance of the |
---|
| 145 | source view or the destination view. |
---|
| 146 | */ |
---|
| 147 | void CGridRemoteConnector::computeConnectorRedundant(void) |
---|
| 148 | { |
---|
| 149 | vector<shared_ptr<CLocalView>> srcView ; |
---|
| 150 | vector<shared_ptr<CDistributedView>> dstView ; |
---|
| 151 | vector<int> indElements ; |
---|
| 152 | elements_.resize(srcView_.size()) ; |
---|
| 153 | |
---|
| 154 | bool srcViewsNonDistributed=true ; // not usefull now but later for optimization |
---|
| 155 | for(int i=0;i<srcView_.size();i++) srcViewsNonDistributed = srcViewsNonDistributed && !isSrcViewDistributed_[i] ; |
---|
| 156 | |
---|
| 157 | bool dstViewsNonDistributed=true ; // not usefull now but later for optimization |
---|
| 158 | for(int i=0;i<dstView_.size();i++) dstViewsNonDistributed = dstViewsNonDistributed && !isDstViewDistributed_[i] ; |
---|
| 159 | |
---|
| 160 | for(int i=0;i<srcView_.size();i++) |
---|
| 161 | { |
---|
| 162 | srcView.push_back(srcView_[i]) ; |
---|
| 163 | dstView.push_back(dstView_[i]) ; |
---|
| 164 | indElements.push_back(i) ; |
---|
| 165 | } |
---|
| 166 | |
---|
| 167 | computeGenericMethod(srcView, dstView, indElements) ; |
---|
| 168 | |
---|
| 169 | map<int,bool> ranks ; |
---|
| 170 | for(auto& it : elements_[indElements[0]]) |
---|
| 171 | { |
---|
| 172 | if (it.second.numElements()==0) ranks[it.first] = false ; |
---|
| 173 | else ranks[it.first] = true ; |
---|
| 174 | } |
---|
| 175 | |
---|
| 176 | } |
---|
| 177 | |
---|
| 178 | |
---|
| 179 | /** |
---|
| 180 | * \brief Compute the connector, i.e. compute the \b elements_ attribute. |
---|
[2179] | 181 | * \detail In order to achive better optimisation, |
---|
| 182 | * we distingute the case when the grid is not distributed on source grid (\bcomputeSrcNonDistributed), |
---|
| 183 | * or the remote grid (\b computeDstNonDistributed), or the both (\b computeSrcDstNonDistributed). |
---|
| 184 | * Otherwise the generic method is called computeGenericMethod. Note that in the case, if one element view |
---|
| 185 | * is not distributed on the source and on the remote grid, then we can used the tensorial product |
---|
| 186 | * property to computing it independently using \b computeSrcDstNonDistributed method. |
---|
| 187 | * After that, we call the \b removeRedondantRanks method to supress blocks of data that can be sent |
---|
| 188 | * redondantly the the remote servers |
---|
| 189 | */ |
---|
[2236] | 190 | void CGridRemoteConnector::computeConnectorMethods(void) |
---|
[1918] | 191 | { |
---|
[2267] | 192 | vector<shared_ptr<CLocalView>> srcView ; |
---|
| 193 | vector<shared_ptr<CDistributedView>> dstView ; |
---|
[2179] | 194 | vector<int> indElements ; |
---|
| 195 | elements_.resize(srcView_.size()) ; |
---|
| 196 | |
---|
| 197 | bool srcViewsNonDistributed=true ; |
---|
[2236] | 198 | for(int i=0;i<srcView_.size();i++) srcViewsNonDistributed = srcViewsNonDistributed && !isSrcViewDistributed_[i] ; |
---|
[2179] | 199 | |
---|
| 200 | bool dstViewsNonDistributed=true ; |
---|
[2236] | 201 | for(int i=0;i<dstView_.size();i++) dstViewsNonDistributed = dstViewsNonDistributed && !isDstViewDistributed_[i] ; |
---|
[2179] | 202 | |
---|
| 203 | if (srcViewsNonDistributed) |
---|
| 204 | { |
---|
| 205 | int commRank, commSize ; |
---|
| 206 | MPI_Comm_rank(localComm_,&commRank) ; |
---|
| 207 | MPI_Comm_size(localComm_,&commSize) ; |
---|
| 208 | list<int> remoteRanks; |
---|
| 209 | list<int> notUsed ; |
---|
| 210 | map<int,bool> ranks ; |
---|
| 211 | computeLeaderProcess(commRank, commSize, remoteSize_, remoteRanks, notUsed) ; |
---|
| 212 | for(int rank : remoteRanks) ranks[rank]=true ; |
---|
| 213 | |
---|
| 214 | for(int i=0; i<srcView_.size(); i++) |
---|
| 215 | { |
---|
| 216 | if (isDstViewDistributed_[i]) computeSrcNonDistributed(i) ; |
---|
| 217 | else computeSrcDstNonDistributed(i, ranks) ; |
---|
| 218 | } |
---|
| 219 | } |
---|
| 220 | else if (dstViewsNonDistributed) |
---|
| 221 | { |
---|
| 222 | map<int,bool> ranks ; |
---|
| 223 | for(int i=0;i<remoteSize_;i++) ranks[i]=true ; |
---|
| 224 | for(int i=0; i<srcView_.size(); i++) |
---|
| 225 | { |
---|
| 226 | if (isSrcViewDistributed_[i]) computeDstNonDistributed(i,ranks) ; |
---|
| 227 | else computeSrcDstNonDistributed(i,ranks) ; |
---|
| 228 | } |
---|
| 229 | } |
---|
| 230 | else |
---|
| 231 | { |
---|
| 232 | for(int i=0;i<srcView_.size();i++) |
---|
| 233 | if (isSrcViewDistributed_[i] || isDstViewDistributed_[i]) |
---|
| 234 | { |
---|
| 235 | srcView.push_back(srcView_[i]) ; |
---|
| 236 | dstView.push_back(dstView_[i]) ; |
---|
| 237 | indElements.push_back(i) ; |
---|
| 238 | } |
---|
| 239 | |
---|
| 240 | computeGenericMethod(srcView, dstView, indElements) ; |
---|
| 241 | |
---|
| 242 | map<int,bool> ranks ; |
---|
| 243 | for(auto& it : elements_[indElements[0]]) |
---|
| 244 | { |
---|
| 245 | if (it.second.numElements()==0) ranks[it.first] = false ; |
---|
| 246 | else ranks[it.first] = true ; |
---|
| 247 | } |
---|
| 248 | |
---|
| 249 | for(int i=0;i<srcView_.size();i++) |
---|
| 250 | if (!isSrcViewDistributed_[i] && !isDstViewDistributed_[i]) computeSrcDstNonDistributed(i, ranks) ; |
---|
| 251 | } |
---|
| 252 | |
---|
[1918] | 253 | } |
---|
| 254 | |
---|
[2179] | 255 | |
---|
| 256 | /** |
---|
| 257 | * \brief Compute the connector for the element \b i when the source view is not distributed. |
---|
| 258 | * After the call element_[i] is defined. |
---|
| 259 | * \param i Indice of the element composing the source grid. |
---|
| 260 | */ |
---|
| 261 | |
---|
| 262 | void CGridRemoteConnector::computeSrcNonDistributed(int i) |
---|
[1918] | 263 | { |
---|
[2179] | 264 | auto& element = elements_[i] ; |
---|
| 265 | map<int,CArray<size_t,1>> globalIndexView ; |
---|
| 266 | dstView_[i]->getGlobalIndexView(globalIndexView) ; |
---|
| 267 | |
---|
| 268 | CClientClientDHTTemplate<int>::Index2InfoTypeMap dataInfo; |
---|
| 269 | |
---|
| 270 | for(auto& it : globalIndexView) |
---|
| 271 | { |
---|
| 272 | auto& globalIndex=it.second ; |
---|
| 273 | for(size_t ind : globalIndex) dataInfo[ind]=it.first ; |
---|
| 274 | } |
---|
| 275 | |
---|
| 276 | // First we feed the distributed hash map with key (remote global index) |
---|
| 277 | // associated with the value of the remote rank |
---|
| 278 | CClientClientDHTTemplate<int> DHT(dataInfo, localComm_) ; |
---|
| 279 | // after we feed the DHT with the local global indices of the source view |
---|
| 280 | |
---|
| 281 | int commRank, commSize ; |
---|
| 282 | MPI_Comm_rank(localComm_,&commRank) ; |
---|
| 283 | MPI_Comm_size(localComm_,&commSize) ; |
---|
| 284 | CArray<size_t,1> srcIndex ; |
---|
| 285 | // like the source view is not distributed, then only the rank 0 need to feed the DHT |
---|
| 286 | if (commRank==0) srcView_[i]->getGlobalIndexView(srcIndex) ; |
---|
| 287 | |
---|
| 288 | // compute the mapping |
---|
| 289 | DHT.computeIndexInfoMapping(srcIndex) ; |
---|
| 290 | auto& returnInfo = DHT.getInfoIndexMap() ; |
---|
| 291 | |
---|
| 292 | // returnInfo contains now the map for each global indices to send to a list of remote rank |
---|
| 293 | // only for the rank=0 because it is the one to feed the DHT |
---|
| 294 | // so it need to send the list to each server leader i.e. the local process that handle specifically one or more |
---|
| 295 | // servers |
---|
| 296 | |
---|
| 297 | // rankIndGlo : rankIndGlo[rank][indGlo] : list of indice to send the the remote server of rank "rank" |
---|
| 298 | vector<vector<size_t>> rankIndGlo(remoteSize_) ; |
---|
| 299 | if (commRank==0) |
---|
| 300 | for(auto& it1 : returnInfo) |
---|
| 301 | for(auto& it2 : it1.second) rankIndGlo[it2].push_back(it1.first) ; |
---|
| 302 | |
---|
| 303 | |
---|
| 304 | vector<MPI_Request> requests ; |
---|
| 305 | |
---|
| 306 | if (commRank==0) |
---|
| 307 | { |
---|
| 308 | requests.resize(remoteSize_) ; |
---|
| 309 | for(int i=0 ; i<remoteSize_;i++) |
---|
| 310 | { |
---|
| 311 | // ok send only the global indices for a server to the "server leader" |
---|
| 312 | int rank = getLeaderRank(commSize, remoteSize_, i) ; |
---|
| 313 | MPI_Isend(rankIndGlo[i].data(), rankIndGlo[i].size(), MPI_SIZE_T, rank, i ,localComm_, &requests[i]) ; |
---|
| 314 | } |
---|
| 315 | } |
---|
| 316 | |
---|
| 317 | list<int> remoteRanks; |
---|
| 318 | list<int> notUsed ; |
---|
| 319 | // I am a server leader of which remote ranks ? |
---|
| 320 | computeLeaderProcess(commRank, commSize, remoteSize_, remoteRanks, notUsed) ; |
---|
| 321 | |
---|
| 322 | for(auto remoteRank : remoteRanks) |
---|
| 323 | { |
---|
| 324 | MPI_Status status ; |
---|
| 325 | int size ; |
---|
| 326 | MPI_Probe(0,remoteRank,localComm_, &status); |
---|
| 327 | MPI_Get_count(&status, MPI_SIZE_T, &size) ; |
---|
| 328 | elements_[i][remoteRank].resize(size) ; |
---|
| 329 | // for each remote ranks receive the global indices from proc 0 |
---|
| 330 | MPI_Recv(elements_[i][remoteRank].dataFirst(),size, MPI_SIZE_T,0,remoteRank, localComm_,&status) ; |
---|
| 331 | } |
---|
| 332 | |
---|
| 333 | if (commRank==0) |
---|
| 334 | { |
---|
| 335 | vector<MPI_Status> status(remoteSize_) ; |
---|
| 336 | // asynchronous for sender, wait for completion |
---|
| 337 | MPI_Waitall(remoteSize_, requests.data(), status.data()) ; |
---|
| 338 | } |
---|
| 339 | } |
---|
| 340 | |
---|
| 341 | /** |
---|
| 342 | * \brief Compute the remote connector for the element \b i when the remote view is not distributed. |
---|
| 343 | * After the call, element_[i] is defined. |
---|
| 344 | * \param i Indice of the element composing the remote grid. |
---|
| 345 | * \param ranks The list of rank for which the local proc is in charge to compute the connector |
---|
| 346 | * (if leader server for exemple). if ranks[rank] == false the corresponding elements_ |
---|
| 347 | * is set to void array (no data to sent) just in order to notify corresponding remote server |
---|
| 348 | * that the call is collective with each other one |
---|
| 349 | */ |
---|
| 350 | void CGridRemoteConnector::computeDstNonDistributed(int i, map<int,bool>& ranks) |
---|
| 351 | { |
---|
| 352 | auto& element = elements_[i] ; |
---|
| 353 | map<int,CArray<size_t,1>> globalIndexView ; |
---|
| 354 | dstView_[i]->getGlobalIndexView(globalIndexView) ; |
---|
| 355 | |
---|
| 356 | |
---|
| 357 | CClientClientDHTTemplate<int>::Index2InfoTypeMap dataInfo; |
---|
| 358 | |
---|
| 359 | // First we feed the distributed hash map with key (remote global index) |
---|
| 360 | // associated with the value of the remote rank |
---|
| 361 | for(auto& it : globalIndexView) |
---|
| 362 | if (it.first==0) // since the remote view is not distributed, insert only the remote rank 0 |
---|
| 363 | { |
---|
| 364 | auto& globalIndex=it.second ; |
---|
| 365 | for(size_t ind : globalIndex) dataInfo[ind]=0 ; // associated the the rank 0 |
---|
| 366 | } |
---|
| 367 | |
---|
| 368 | CClientClientDHTTemplate<int> DHT(dataInfo, localComm_) ; |
---|
| 369 | // after we feed the DHT with the local global indices of the source view |
---|
| 370 | |
---|
| 371 | CArray<size_t,1> srcIndex ; |
---|
| 372 | srcView_[i]->getGlobalIndexView(srcIndex) ; |
---|
| 373 | DHT.computeIndexInfoMapping(srcIndex) ; |
---|
| 374 | auto& returnInfo = DHT.getInfoIndexMap() ; |
---|
| 375 | |
---|
| 376 | // returnInfo contains now the map for each global indices to send to a list of remote rank |
---|
| 377 | // now construct the element_ list of global indices for each rank in my list except if the erray must be empty |
---|
| 378 | for (auto& rank : ranks) |
---|
| 379 | { |
---|
| 380 | if (rank.second) // non empty array => for rank that have not any data to be received |
---|
| 381 | { |
---|
| 382 | int size=0 ; |
---|
| 383 | for(auto& it : returnInfo) if (!it.second.empty()) size++ ; |
---|
| 384 | auto& array = element[rank.first] ; |
---|
| 385 | array.resize(size) ; |
---|
| 386 | size=0 ; |
---|
| 387 | for(auto& it : returnInfo) |
---|
| 388 | if (!it.second.empty()) |
---|
| 389 | { |
---|
| 390 | array(size)=it.first ; |
---|
| 391 | size++ ; |
---|
| 392 | } |
---|
| 393 | } |
---|
| 394 | else element[rank.first] = CArray<size_t,1>(0) ; // empty array => for rank that have not any data to be received |
---|
| 395 | } |
---|
| 396 | } |
---|
| 397 | |
---|
| 398 | /** |
---|
| 399 | * \brief Compute the remote connector for the element \b i when the source and the remote view are not distributed. |
---|
| 400 | * After the call, element_[i] is defined. |
---|
| 401 | * \param i Indice of the element composing the remote grid. |
---|
| 402 | * \param ranks The list of rank for which the local proc is in charge to compute the connector |
---|
| 403 | * (if leader server for exemple). if ranks[rank] == false the corresponding elements_ |
---|
| 404 | * is set to void array (no data to sent) just in order to notify corresponding remote server |
---|
| 405 | * that the call is collective with each other one |
---|
| 406 | */ |
---|
| 407 | |
---|
| 408 | void CGridRemoteConnector::computeSrcDstNonDistributed(int i, map<int,bool>& ranks) |
---|
| 409 | { |
---|
| 410 | auto& element = elements_[i] ; |
---|
| 411 | map<int,CArray<size_t,1>> globalIndexView ; |
---|
| 412 | dstView_[i]->getGlobalIndexView(globalIndexView) ; |
---|
| 413 | |
---|
| 414 | |
---|
| 415 | CClientClientDHTTemplate<int>::Index2InfoTypeMap dataInfo; |
---|
| 416 | // First we feed the distributed hash map with key (remote global index) |
---|
| 417 | // associated with the value of the remote rank |
---|
| 418 | |
---|
| 419 | for(auto& it : globalIndexView) |
---|
| 420 | if (it.first==0) // insert only the remote rank 0 since the remote view is not distributed |
---|
| 421 | { |
---|
| 422 | auto& globalIndex=it.second ; |
---|
| 423 | for(size_t ind : globalIndex) dataInfo[ind]=0 ; // associated the the rank 0 |
---|
| 424 | } |
---|
| 425 | |
---|
| 426 | CClientClientDHTTemplate<int> DHT(dataInfo, localComm_) ; |
---|
| 427 | // after we feed the DHT with the local global indices of the source view |
---|
| 428 | |
---|
| 429 | int commRank, commSize ; |
---|
| 430 | MPI_Comm_rank(localComm_,&commRank) ; |
---|
| 431 | MPI_Comm_size(localComm_,&commSize) ; |
---|
| 432 | CArray<size_t,1> srcIndex ; |
---|
| 433 | |
---|
| 434 | // like the source view is not distributed, then only the rank 0 need to feed the DHT |
---|
| 435 | if (commRank==0) srcView_[i]->getGlobalIndexView(srcIndex) ; |
---|
| 436 | DHT.computeIndexInfoMapping(srcIndex) ; |
---|
| 437 | auto& returnInfo = DHT.getInfoIndexMap() ; |
---|
| 438 | |
---|
| 439 | vector<size_t> indGlo ; |
---|
| 440 | if (commRank==0) |
---|
| 441 | for(auto& it1 : returnInfo) |
---|
| 442 | for(auto& it2 : it1.second) indGlo.push_back(it1.first) ; |
---|
| 443 | |
---|
| 444 | // now local rank 0 know which indices to seed to remote rank 0, but all the server |
---|
| 445 | // must receive the same information. So only the leader rank will sent this. |
---|
| 446 | // So local rank 0 must broadcast the information to all leader. |
---|
| 447 | // for this we create a new communicator composed of local process that must send data |
---|
| 448 | // to a remote rank, data are broadcasted, and element_[i] is construction for each remote |
---|
| 449 | // rank in charge |
---|
| 450 | int color=0 ; |
---|
| 451 | if (ranks.empty()) color=0 ; |
---|
| 452 | else color=1 ; |
---|
| 453 | if (commRank==0) color=1 ; |
---|
| 454 | MPI_Comm newComm ; |
---|
| 455 | MPI_Comm_split(localComm_, color, commRank, &newComm) ; |
---|
| 456 | if (color==1) |
---|
| 457 | { |
---|
| 458 | // ok, I am part of the process that must send something to one or more remote server |
---|
| 459 | // so I get the list of global indices from rank 0 |
---|
| 460 | int dataSize ; |
---|
| 461 | if (commRank==0) dataSize=indGlo.size() ; |
---|
| 462 | MPI_Bcast(&dataSize,1,MPI_INT, 0, newComm) ; |
---|
| 463 | indGlo.resize(dataSize) ; |
---|
| 464 | MPI_Bcast(indGlo.data(),dataSize,MPI_SIZE_T,0,newComm) ; |
---|
| 465 | } |
---|
| 466 | MPI_Comm_free(&newComm) ; |
---|
| 467 | |
---|
| 468 | // construct element_[i] from indGlo |
---|
| 469 | for(auto& rank : ranks) |
---|
| 470 | { |
---|
| 471 | if (rank.second) |
---|
| 472 | { |
---|
| 473 | int dataSize=indGlo.size(); |
---|
| 474 | auto& element = elements_[i][rank.first] ; |
---|
| 475 | element.resize(dataSize) ; |
---|
| 476 | for(int i=0;i<dataSize; i++) element(i)=indGlo[i] ; |
---|
| 477 | } |
---|
| 478 | else element[rank.first] = CArray<size_t,1>(0) ; |
---|
| 479 | } |
---|
| 480 | |
---|
| 481 | } |
---|
| 482 | |
---|
[2291] | 483 | |
---|
[2179] | 484 | /** |
---|
| 485 | * \brief Generic method the compute the grid remote connector. Only distributed elements are specifed in the source view and remote view. |
---|
| 486 | * Connector for non distributed elements are computed separatly to improve performance and memory consumption. After the call, |
---|
| 487 | * \b elements_ is defined. |
---|
| 488 | * \param srcView List of the source views composing the grid, without non distributed views |
---|
| 489 | * \param dstView List of the remote views composing the grid, without non distributed views |
---|
| 490 | * \param indElements Index of the view making the correspondance between all views and views distributed (that are in input) |
---|
| 491 | */ |
---|
[2267] | 492 | void CGridRemoteConnector::computeGenericMethod(vector<shared_ptr<CLocalView>>& srcView, vector<shared_ptr<CDistributedView>>& dstView, vector<int>& indElements) |
---|
[2179] | 493 | { |
---|
[1918] | 494 | // generic method, every element can be distributed |
---|
[2179] | 495 | int nDst = dstView.size() ; |
---|
[1930] | 496 | vector<size_t> dstSliceSize(nDst) ; |
---|
| 497 | dstSliceSize[0] = 1 ; |
---|
[2179] | 498 | for(int i=1; i<nDst; i++) dstSliceSize[i] = dstView[i-1]->getGlobalSize()*dstSliceSize[i-1] ; |
---|
[1930] | 499 | |
---|
| 500 | CClientClientDHTTemplate<int>::Index2VectorInfoTypeMap dataInfo ; |
---|
[2179] | 501 | CClientClientDHTTemplate<size_t>::Index2VectorInfoTypeMap info ; // info map |
---|
[1930] | 502 | |
---|
[2179] | 503 | // first, we need to feed the DHT with the global index of the remote server |
---|
| 504 | // for that : |
---|
| 505 | // First the first element insert the in a DHT with key as the rank and value the list of global index associated |
---|
| 506 | // Then get the previously stored index associate with the remote rank I am in charge and reinsert the global index |
---|
| 507 | // corresponding to the position of the element in the remote view suing tensorial product |
---|
| 508 | // finaly we get only the list of remote global index I am in charge for the whole remote grid |
---|
| 509 | |
---|
[1930] | 510 | for(int pos=0; pos<nDst; pos++) |
---|
| 511 | { |
---|
| 512 | size_t sliceSize=dstSliceSize[pos] ; |
---|
| 513 | map<int,CArray<size_t,1>> globalIndexView ; |
---|
[2179] | 514 | dstView[pos]->getGlobalIndexView(globalIndexView) ; |
---|
[1930] | 515 | |
---|
| 516 | CClientClientDHTTemplate<size_t>::Index2VectorInfoTypeMap lastInfo(info) ; |
---|
| 517 | |
---|
| 518 | if (pos>0) |
---|
| 519 | { |
---|
| 520 | CArray<size_t,1> ranks(globalIndexView.size()) ; |
---|
| 521 | auto it=globalIndexView.begin() ; |
---|
| 522 | for(int i=0 ; i<ranks.numElements();i++,it++) ranks(i)=it->first ; |
---|
| 523 | CClientClientDHTTemplate<size_t> dataRanks(info, localComm_) ; |
---|
| 524 | dataRanks.computeIndexInfoMapping(ranks) ; |
---|
| 525 | lastInfo = dataRanks.getInfoIndexMap() ; |
---|
| 526 | } |
---|
| 527 | |
---|
| 528 | info.clear() ; |
---|
| 529 | for(auto& it : globalIndexView) |
---|
| 530 | { |
---|
| 531 | int rank = it.first ; |
---|
| 532 | auto& globalIndex = it.second ; |
---|
| 533 | auto& inf = info[rank] ; |
---|
| 534 | if (pos==0) for(int i=0;i<globalIndex.numElements();i++) inf.push_back(globalIndex(i)) ; |
---|
| 535 | else |
---|
| 536 | { |
---|
| 537 | auto& lastGlobalIndex = lastInfo[rank] ; |
---|
| 538 | for(size_t lastGlobalInd : lastGlobalIndex) |
---|
| 539 | { |
---|
| 540 | for(int i=0;i<globalIndex.numElements();i++) inf.push_back(globalIndex(i)*sliceSize+lastGlobalInd) ; |
---|
| 541 | } |
---|
| 542 | } |
---|
| 543 | } |
---|
| 544 | |
---|
| 545 | if (pos==nDst-1) |
---|
| 546 | { |
---|
| 547 | for(auto& it : info) |
---|
| 548 | { |
---|
| 549 | int rank=it.first ; |
---|
| 550 | auto& globalIndex = it.second ; |
---|
| 551 | for(auto globalInd : globalIndex) dataInfo[globalInd].push_back(rank) ; |
---|
| 552 | } |
---|
| 553 | } |
---|
| 554 | } |
---|
| 555 | |
---|
[2179] | 556 | // we feed the DHT with the remote global index |
---|
[1930] | 557 | CClientClientDHTTemplate<int> dataRanks(dataInfo, localComm_) ; |
---|
[1938] | 558 | |
---|
[1918] | 559 | // generate list of global index for src view |
---|
[2179] | 560 | int nSrc = srcView.size() ; |
---|
[1918] | 561 | vector<size_t> srcSliceSize(nSrc) ; |
---|
[1930] | 562 | |
---|
| 563 | srcSliceSize[0] = 1 ; |
---|
[2179] | 564 | for(int i=1; i<nSrc; i++) srcSliceSize[i] = srcView[i-1]->getGlobalSize()*srcSliceSize[i-1] ; |
---|
[1930] | 565 | |
---|
[1918] | 566 | vector<size_t> srcGlobalIndex ; |
---|
| 567 | size_t sliceIndex=0 ; |
---|
[2179] | 568 | srcView[nSrc-1]->getGlobalIndex(srcGlobalIndex, sliceIndex, srcSliceSize.data(), srcView.data(), nSrc-1) ; |
---|
| 569 | // now we have the global index of the source grid in srcGlobalIndex |
---|
| 570 | // we feed the DHT with the src global index (if we have) |
---|
[1984] | 571 | if (srcGlobalIndex.size()>0) |
---|
| 572 | { |
---|
| 573 | CArray<size_t,1> srcGlobalIndexArray(srcGlobalIndex.data(), shape(srcGlobalIndex.size()),neverDeleteData) ; |
---|
| 574 | dataRanks.computeIndexInfoMapping(srcGlobalIndexArray) ; |
---|
| 575 | } |
---|
| 576 | else |
---|
| 577 | { |
---|
| 578 | CArray<size_t,1> srcGlobalIndexArray ; |
---|
| 579 | dataRanks.computeIndexInfoMapping(srcGlobalIndexArray) ; |
---|
| 580 | } |
---|
[1918] | 581 | const auto& returnInfo = dataRanks.getInfoIndexMap() ; |
---|
[2179] | 582 | // returnInfo contains now the map for each global indices to send to a list of remote rank |
---|
| 583 | // but we want to use the tensorial product property to get the same information using only global |
---|
| 584 | // index of element view. So the idea is to reverse the information : for a global index of the grid |
---|
| 585 | // to send to the remote server, what is the global index of each element composing the grid ? |
---|
[1918] | 586 | |
---|
| 587 | vector<map<int, set<size_t>>> elements(nSrc) ; // internal representation of elements composing the grid |
---|
| 588 | |
---|
| 589 | for(auto& indRanks : returnInfo) |
---|
| 590 | { |
---|
| 591 | size_t gridIndexGlo=indRanks.first ; |
---|
| 592 | auto& ranks = indRanks.second ; |
---|
[1930] | 593 | for(int i=nSrc-1; i>=0; i--) |
---|
[1918] | 594 | { |
---|
| 595 | auto& element = elements[i] ; |
---|
[1930] | 596 | size_t localIndGlo = gridIndexGlo / srcSliceSize[i] ; |
---|
| 597 | gridIndexGlo = gridIndexGlo % srcSliceSize[i] ; |
---|
[1918] | 598 | for(int rank : ranks) element[rank].insert(localIndGlo) ; |
---|
| 599 | } |
---|
| 600 | } |
---|
| 601 | |
---|
[2179] | 602 | // elements_.resize(nSrc) ; |
---|
[1918] | 603 | for(int i=0 ; i<nSrc; i++) |
---|
| 604 | { |
---|
| 605 | auto& element=elements[i] ; |
---|
| 606 | for(auto& rankInd : element) |
---|
| 607 | { |
---|
| 608 | int rank=rankInd.first ; |
---|
| 609 | set<size_t>& indGlo = rankInd.second ; |
---|
[2179] | 610 | CArray<size_t,1>& indGloArray = elements_[indElements[i]][rank] ; |
---|
[1918] | 611 | indGloArray.resize(indGlo.size()) ; |
---|
| 612 | int j=0 ; |
---|
| 613 | for (auto index : indGlo) { indGloArray(j) = index ; j++; } |
---|
| 614 | } |
---|
| 615 | } |
---|
[1938] | 616 | |
---|
| 617 | // So what about when there is some server that have no data to receive |
---|
| 618 | // they must be inform they receive an event with no data. |
---|
| 619 | // So find remote servers with no data, and one client will take in charge |
---|
| 620 | // that it receive global index with no data (0-size) |
---|
| 621 | vector<int> ranks(remoteSize_,0) ; |
---|
[2179] | 622 | for(auto& it : elements_[indElements[0]]) ranks[it.first] = 1 ; |
---|
[1938] | 623 | MPI_Allreduce(MPI_IN_PLACE, ranks.data(), remoteSize_, MPI_INT, MPI_SUM, localComm_) ; |
---|
| 624 | int commRank, commSize ; |
---|
| 625 | MPI_Comm_rank(localComm_,&commRank) ; |
---|
| 626 | MPI_Comm_size(localComm_,&commSize) ; |
---|
| 627 | int pos=0 ; |
---|
| 628 | for(int i=0; i<remoteSize_ ; i++) |
---|
| 629 | if (ranks[i]==0) |
---|
| 630 | { |
---|
[2179] | 631 | if (pos%commSize==commRank) |
---|
| 632 | for(int j=0 ; j<nSrc; j++) elements_[indElements[j]][i] = CArray<size_t,1>(0) ; |
---|
[1938] | 633 | pos++ ; |
---|
| 634 | } |
---|
[1918] | 635 | } |
---|
| 636 | |
---|
[2179] | 637 | /** |
---|
[2236] | 638 | * \brief Once the connector is computed (compute \b elements_), redondant data can be avoid to be sent to the server. |
---|
| 639 | * This call compute the redondant rank and store them in \b rankToRemove_ attribute. |
---|
[2179] | 640 | * The goal of this method is to make a hash of each block of indice that determine wich data to send to a |
---|
| 641 | * of a specific server rank using a hash method. So data to send to a rank is associated to a hash. |
---|
| 642 | * After we compare hash between local rank and remove redondant data corresponding to the same hash. |
---|
| 643 | */ |
---|
[2236] | 644 | void CGridRemoteConnector::computeRedondantRanks(void) |
---|
[2179] | 645 | { |
---|
| 646 | int commRank ; |
---|
| 647 | MPI_Comm_rank(localComm_,&commRank) ; |
---|
[1938] | 648 | |
---|
[2179] | 649 | set<int> ranks; |
---|
| 650 | for(auto& element : elements_) |
---|
| 651 | for(auto& it : element) ranks.insert(it.first) ; |
---|
| 652 | |
---|
| 653 | for(auto& element : elements_) |
---|
| 654 | for(auto& it : element) |
---|
| 655 | if (ranks.count(it.first)==0) ERROR("void CGridRemoteConnector::removeRedondantRanks(void)",<<"number of ranks in elements is not coherent between each element") ; |
---|
| 656 | |
---|
| 657 | HashXIOS<size_t> hashGlobalIndex; |
---|
| 658 | |
---|
| 659 | map<int,size_t> hashRanks ; |
---|
| 660 | for(auto& element : elements_) |
---|
| 661 | for(auto& it : element) |
---|
| 662 | { |
---|
| 663 | auto& globalIndex=it.second ; |
---|
| 664 | int rank=it.first ; |
---|
| 665 | size_t hash ; |
---|
| 666 | hash=0 ; |
---|
| 667 | for(int i=0; i<globalIndex.numElements(); i++) hash+=hashGlobalIndex(globalIndex(i)) ; |
---|
[2296] | 668 | if (globalIndex.numElements()>0) |
---|
| 669 | { |
---|
| 670 | if (hashRanks.count(rank)==0) hashRanks[rank]=hash ; |
---|
| 671 | else hashRanks[rank]=hashGlobalIndex.hashCombine(hashRanks[rank],hash) ; |
---|
| 672 | } |
---|
[2179] | 673 | } |
---|
| 674 | // a hash is now computed for data block I will sent to the server. |
---|
| 675 | |
---|
| 676 | CClientClientDHTTemplate<int>::Index2InfoTypeMap info ; |
---|
| 677 | |
---|
| 678 | map<size_t,int> hashRank ; |
---|
| 679 | HashXIOS<int> hashGlobalIndexRank; |
---|
| 680 | for(auto& it : hashRanks) |
---|
| 681 | { |
---|
| 682 | it.second = hashGlobalIndexRank.hashCombine(it.first,it.second) ; |
---|
| 683 | info[it.second]=commRank ; |
---|
| 684 | hashRank[it.second]=it.first ; |
---|
| 685 | } |
---|
| 686 | |
---|
| 687 | // we feed a DHT map with key : hash, value : myrank |
---|
| 688 | CClientClientDHTTemplate<int> dataHash(info, localComm_) ; |
---|
| 689 | CArray<size_t,1> hashList(hashRank.size()) ; |
---|
| 690 | |
---|
| 691 | int i=0 ; |
---|
| 692 | for(auto& it : hashRank) { hashList(i)=it.first ; i++; } |
---|
| 693 | |
---|
| 694 | // now who are the ranks that have the same hash : feed the DHT with my list of hash |
---|
| 695 | dataHash.computeIndexInfoMapping(hashList) ; |
---|
| 696 | auto& hashRankList = dataHash.getInfoIndexMap() ; |
---|
| 697 | |
---|
| 698 | |
---|
| 699 | for(auto& it : hashRankList) |
---|
| 700 | { |
---|
| 701 | size_t hash = it.first ; |
---|
| 702 | auto& ranks = it.second ; |
---|
| 703 | |
---|
| 704 | bool first=true ; |
---|
| 705 | // only the process with the lowest rank get in charge of sendinf data to remote server |
---|
| 706 | for(int rank : ranks) if (commRank>rank) first=false ; |
---|
[2236] | 707 | if (!first) rankToRemove_.insert(hashRank[hash]) ; |
---|
[2179] | 708 | } |
---|
[2236] | 709 | } |
---|
[2291] | 710 | |
---|
[1918] | 711 | } |
---|