1 | /*! |
---|
2 | \file client_client_dht.cpp |
---|
3 | \author Ha NGUYEN |
---|
4 | \since 15 Sep 2015 |
---|
5 | \date 15 Sep 2015 |
---|
6 | |
---|
7 | \brief Distributed hashed table implementation. |
---|
8 | */ |
---|
9 | #include "client_client_dht.hpp" |
---|
10 | #include <limits> |
---|
11 | #include <cmath> |
---|
12 | #include <boost/functional/hash.hpp> |
---|
13 | #include "utils.hpp" |
---|
14 | #include "mpi_tag.hpp" |
---|
15 | |
---|
16 | namespace xios |
---|
17 | { |
---|
18 | |
---|
19 | CClientClientDHT::CClientClientDHT(const boost::unordered_map<size_t,int>& indexInfoMap, |
---|
20 | const MPI_Comm& clientIntraComm, bool isDataDistributed, int hierarLvl) |
---|
21 | : intraCommRoot_(clientIntraComm), commLevel_(), isDataDistributed_(isDataDistributed), |
---|
22 | nbLevel_(hierarLvl), globalIndexToServerMapping_(), globalIndexToInfoMappingLevel_() |
---|
23 | { |
---|
24 | computeMPICommLevel(clientIntraComm); |
---|
25 | int lvl = commLevel_.size() - 1; |
---|
26 | computeDistributedIndex(indexInfoMap, commLevel_[lvl], lvl); |
---|
27 | } |
---|
28 | |
---|
29 | CClientClientDHT::~CClientClientDHT() |
---|
30 | { |
---|
31 | } |
---|
32 | |
---|
33 | /*! |
---|
34 | Calculate MPI communicator for each level of hierarchy. |
---|
35 | \param[in] mpiCommRoot MPI communicator of the level 0 (usually communicator of all clients) |
---|
36 | */ |
---|
37 | void CClientClientDHT::computeMPICommLevel(const MPI_Comm& mpiCommRoot) |
---|
38 | { |
---|
39 | int nbProc; |
---|
40 | MPI_Comm_size(mpiCommRoot,&nbProc); |
---|
41 | if (nbLevel_ > nbProc) nbLevel_ = std::log(nbProc); |
---|
42 | else if (1 > nbLevel_) nbLevel_ = 1; |
---|
43 | |
---|
44 | commLevel_.push_back(mpiCommRoot); |
---|
45 | divideMPICommLevel(mpiCommRoot, nbLevel_); |
---|
46 | } |
---|
47 | |
---|
48 | /*! |
---|
49 | Divide each MPI communicator into sub-communicator. Recursive function |
---|
50 | \param [in] mpiCommLevel MPI communicator of current level |
---|
51 | \param [in] level current level |
---|
52 | */ |
---|
53 | void CClientClientDHT::divideMPICommLevel(const MPI_Comm& mpiCommLevel, int level) |
---|
54 | { |
---|
55 | int clientRank; |
---|
56 | MPI_Comm_rank(mpiCommLevel,&clientRank); |
---|
57 | |
---|
58 | --level; |
---|
59 | if (0 < level) |
---|
60 | { |
---|
61 | int color = clientRank % 2; |
---|
62 | commLevel_.push_back(MPI_Comm()); |
---|
63 | MPI_Comm_split(mpiCommLevel, color, 0, &(commLevel_.back())); |
---|
64 | divideMPICommLevel(commLevel_.back(), level); |
---|
65 | } |
---|
66 | } |
---|
67 | |
---|
68 | /*! |
---|
69 | Compute mapping between indices and information corresponding to these indices |
---|
70 | \param [in] indices indices a proc has |
---|
71 | */ |
---|
72 | void CClientClientDHT::computeServerIndexMapping(const CArray<size_t,1>& indices) |
---|
73 | { |
---|
74 | int lvl = commLevel_.size() - 1; |
---|
75 | computeIndexMapping(indices, commLevel_[lvl], lvl); |
---|
76 | size_t size = indices.numElements(); |
---|
77 | for (size_t idx = 0; idx < size; ++idx) |
---|
78 | { |
---|
79 | int serverIdx = globalIndexToInfoMappingLevel_[indices(idx)]; |
---|
80 | indexGlobalOnServer_[serverIdx].push_back(indices(idx)); |
---|
81 | } |
---|
82 | } |
---|
83 | |
---|
84 | /*! |
---|
85 | Compute mapping between indices and information corresponding to these indices |
---|
86 | for each level of hierarchical DHT. Recursive function |
---|
87 | \param [in] indices indices a proc has |
---|
88 | \param [in] commLevel communicator of current level |
---|
89 | \param [in] level current level |
---|
90 | */ |
---|
91 | void CClientClientDHT::computeIndexMapping(const CArray<size_t,1>& indices, |
---|
92 | const MPI_Comm& commLevel, |
---|
93 | int level) |
---|
94 | { |
---|
95 | int nbClient, clientRank; |
---|
96 | MPI_Comm_size(commLevel,&nbClient); |
---|
97 | MPI_Comm_rank(commLevel,&clientRank); |
---|
98 | std::vector<size_t> hashedIndex; |
---|
99 | computeHashIndex(hashedIndex, nbClient); |
---|
100 | |
---|
101 | size_t ssize = indices.numElements(), hashedVal; |
---|
102 | |
---|
103 | std::vector<size_t>::const_iterator itbClientHash = hashedIndex.begin(), itClientHash, |
---|
104 | iteClientHash = hashedIndex.end(); |
---|
105 | std::map<int, std::vector<size_t> > client2ClientIndex; |
---|
106 | |
---|
107 | // Number of global index whose mapping server can be found out thanks to index-server mapping |
---|
108 | int nbIndexAlreadyOnClient = 0; |
---|
109 | |
---|
110 | // Number of global index whose mapping server are on other clients |
---|
111 | int nbIndexSendToOthers = 0; |
---|
112 | HashXIOS<size_t> hashGlobalIndex; |
---|
113 | for (int i = 0; i < ssize; ++i) |
---|
114 | { |
---|
115 | size_t index = indices(i); |
---|
116 | hashedVal = hashGlobalIndex(index); |
---|
117 | itClientHash = std::upper_bound(itbClientHash, iteClientHash, hashedVal); |
---|
118 | if (iteClientHash != itClientHash) |
---|
119 | { |
---|
120 | int indexClient = std::distance(itbClientHash, itClientHash)-1; |
---|
121 | { |
---|
122 | client2ClientIndex[indexClient].push_back(index); |
---|
123 | ++nbIndexSendToOthers; |
---|
124 | } |
---|
125 | } |
---|
126 | } |
---|
127 | info << "level " << level << " nbIndexsendtoOther " << nbIndexSendToOthers << std::endl; |
---|
128 | |
---|
129 | int* sendBuff = new int[nbClient]; |
---|
130 | for (int i = 0; i < nbClient; ++i) sendBuff[i] = 0; |
---|
131 | std::map<int, std::vector<size_t> >::iterator itb = client2ClientIndex.begin(), it, |
---|
132 | ite = client2ClientIndex.end(); |
---|
133 | for (it = itb; it != ite; ++it) sendBuff[it->first] = 1; |
---|
134 | int* recvBuff = new int[nbClient]; |
---|
135 | MPI_Allreduce(sendBuff, recvBuff, nbClient, MPI_INT, MPI_SUM, commLevel); |
---|
136 | |
---|
137 | std::list<MPI_Request> sendRequest; |
---|
138 | if (0 != nbIndexSendToOthers) |
---|
139 | for (it = itb; it != ite; ++it) |
---|
140 | sendIndexToClients(it->first, it->second, commLevel, sendRequest); |
---|
141 | |
---|
142 | int nbDemandingClient = recvBuff[clientRank], nbIndexServerReceived = 0; |
---|
143 | |
---|
144 | // Receiving demand as well as the responds from other clients |
---|
145 | // The demand message contains global index; meanwhile the responds have server index information |
---|
146 | // Buffer to receive demand from other clients, it can be allocated or not depending whether it has demand(s) |
---|
147 | // There are some cases we demand duplicate index so need to determine maxium size of demanding buffer |
---|
148 | for (it = itb; it != ite; ++it) sendBuff[it->first] = (it->second).size(); |
---|
149 | MPI_Allreduce(sendBuff, recvBuff, nbClient, MPI_INT, MPI_SUM, commLevel); |
---|
150 | |
---|
151 | unsigned long* recvBuffIndex = 0; |
---|
152 | int maxNbIndexDemandedFromOthers = recvBuff[clientRank]; |
---|
153 | // if (!isDataDistributed_) maxNbIndexDemandedFromOthers = nbDemandingClient * nbIndexSendToOthers; //globalIndexToServerMapping_.size(); // Not very optimal but it's general |
---|
154 | |
---|
155 | if (0 != maxNbIndexDemandedFromOthers) |
---|
156 | recvBuffIndex = new unsigned long[maxNbIndexDemandedFromOthers]; |
---|
157 | |
---|
158 | // Buffer to receive respond from other clients, it can be allocated or not depending whether it demands other clients |
---|
159 | int* recvBuffInfo = 0; |
---|
160 | int nbIndexReceivedFromOthers = nbIndexSendToOthers; |
---|
161 | if (0 != nbIndexReceivedFromOthers) |
---|
162 | recvBuffInfo = new int[nbIndexReceivedFromOthers]; |
---|
163 | |
---|
164 | std::map<int, MPI_Request>::iterator itRequest; |
---|
165 | std::vector<int> demandAlreadyReceived, repondAlreadyReceived; |
---|
166 | |
---|
167 | // Counting of buffer for receiving global index |
---|
168 | int countIndex = 0; |
---|
169 | |
---|
170 | // Request returned by MPI_IRecv function about global index |
---|
171 | std::map<int, MPI_Request> requestRecvIndex; |
---|
172 | |
---|
173 | // Mapping client rank and the beginning position of receiving buffer for message of global index from this client |
---|
174 | std::map<int, unsigned long*> indexBuffBegin; |
---|
175 | |
---|
176 | std::map<int,std::vector<size_t> > src2Index; // Temporary mapping contains info of demand (source and associate index) in curren level |
---|
177 | |
---|
178 | CArray<size_t,1> tmpGlobalIndexOnClient(maxNbIndexDemandedFromOthers); |
---|
179 | int k = 0; |
---|
180 | |
---|
181 | while ((0 < nbDemandingClient) || (!sendRequest.empty())) |
---|
182 | { |
---|
183 | // Just check whether a client has any demand from other clients. |
---|
184 | // If it has, then it should send responds to these client(s) |
---|
185 | probeIndexMessageFromClients(recvBuffIndex, maxNbIndexDemandedFromOthers, |
---|
186 | countIndex, indexBuffBegin, |
---|
187 | requestRecvIndex, commLevel); |
---|
188 | if (0 < nbDemandingClient) |
---|
189 | { |
---|
190 | for (itRequest = requestRecvIndex.begin(); |
---|
191 | itRequest != requestRecvIndex.end(); ++itRequest) |
---|
192 | { |
---|
193 | int flagIndexGlobal, count; |
---|
194 | MPI_Status statusIndexGlobal; |
---|
195 | |
---|
196 | MPI_Test(&(itRequest->second), &flagIndexGlobal, &statusIndexGlobal); |
---|
197 | if (true == flagIndexGlobal) |
---|
198 | { |
---|
199 | MPI_Get_count(&statusIndexGlobal, MPI_UNSIGNED_LONG, &count); |
---|
200 | int clientSourceRank = statusIndexGlobal.MPI_SOURCE; |
---|
201 | unsigned long* beginBuff = indexBuffBegin[clientSourceRank]; |
---|
202 | for (int i = 0; i < count; ++i) |
---|
203 | { |
---|
204 | src2Index[clientSourceRank].push_back(*(beginBuff+i)); |
---|
205 | tmpGlobalIndexOnClient(k) = *(beginBuff+i); |
---|
206 | ++k; |
---|
207 | } |
---|
208 | --nbDemandingClient; |
---|
209 | |
---|
210 | demandAlreadyReceived.push_back(clientSourceRank); |
---|
211 | } |
---|
212 | } |
---|
213 | for (int i = 0; i< demandAlreadyReceived.size(); ++i) |
---|
214 | requestRecvIndex.erase(demandAlreadyReceived[i]); |
---|
215 | } |
---|
216 | |
---|
217 | testSendRequest(sendRequest); |
---|
218 | } |
---|
219 | |
---|
220 | if (0 < level) |
---|
221 | { |
---|
222 | --level; |
---|
223 | computeIndexMapping(tmpGlobalIndexOnClient, commLevel_[level], level); |
---|
224 | } |
---|
225 | else |
---|
226 | globalIndexToInfoMappingLevel_ = globalIndexToServerMapping_; |
---|
227 | |
---|
228 | std::map<int, std::vector<int> > client2ClientInfo; |
---|
229 | std::list<MPI_Request> sendInfoRequest; |
---|
230 | std::map<int, std::vector<size_t> >::iterator itbSrc2Idx = src2Index.begin(), itSrc2Idx, |
---|
231 | iteSrc2Idx = src2Index.end(); |
---|
232 | for (itSrc2Idx = itbSrc2Idx; itSrc2Idx != iteSrc2Idx; ++itSrc2Idx) |
---|
233 | { |
---|
234 | int clientSourceRank = itSrc2Idx->first; |
---|
235 | std::vector<size_t>& srcIdx = itSrc2Idx->second; |
---|
236 | for (int idx = 0; idx < srcIdx.size(); ++idx) |
---|
237 | { |
---|
238 | // client2ClientInfo[clientSourceRank].push_back(globalIndexToServerMapping_[srcIdx[idx]]); |
---|
239 | client2ClientInfo[clientSourceRank].push_back(globalIndexToInfoMappingLevel_[srcIdx[idx]]); |
---|
240 | } |
---|
241 | sendInfoToClients(clientSourceRank, client2ClientInfo[clientSourceRank], commLevel, sendInfoRequest); |
---|
242 | } |
---|
243 | |
---|
244 | boost::unordered_map<size_t,int> indexToInfoMapping; |
---|
245 | |
---|
246 | // Counting of buffer for receiving server index |
---|
247 | int countInfo = 0; |
---|
248 | |
---|
249 | std::map<int, MPI_Request> requestRecvInfo; |
---|
250 | |
---|
251 | // Mapping client rank and the begining position of receiving buffer for message of server index from this client |
---|
252 | std::map<int, int*> infoBuffBegin; |
---|
253 | |
---|
254 | while ((!sendInfoRequest.empty()) || (nbIndexServerReceived < nbIndexReceivedFromOthers)) |
---|
255 | { |
---|
256 | testSendRequest(sendInfoRequest); |
---|
257 | |
---|
258 | // In some cases, a client need to listen respond from other clients about server information |
---|
259 | // Ok, with the information, a client can fill in its server-global index map. |
---|
260 | probeInfoMessageFromClients(recvBuffInfo, nbIndexReceivedFromOthers, |
---|
261 | countInfo, infoBuffBegin, |
---|
262 | requestRecvInfo, commLevel); |
---|
263 | for (itRequest = requestRecvInfo.begin(); |
---|
264 | itRequest != requestRecvInfo.end(); |
---|
265 | ++itRequest) |
---|
266 | { |
---|
267 | int flagInfo, count; |
---|
268 | MPI_Status statusInfo; |
---|
269 | |
---|
270 | MPI_Test(&(itRequest->second), &flagInfo, &statusInfo); |
---|
271 | if (true == flagInfo) |
---|
272 | { |
---|
273 | MPI_Get_count(&statusInfo, MPI_INT, &count); |
---|
274 | int clientSourceRank = statusInfo.MPI_SOURCE; |
---|
275 | int* beginBuff = infoBuffBegin[clientSourceRank]; |
---|
276 | std::vector<size_t>& globalIndexTmp = client2ClientIndex[clientSourceRank]; |
---|
277 | for (int i = 0; i < count; ++i) |
---|
278 | { |
---|
279 | indexToInfoMapping[globalIndexTmp[i]] = *(beginBuff+i); |
---|
280 | // globalIndexToServerMapping_[globalIndexTmp[i]] = *(beginBuff+i); |
---|
281 | } |
---|
282 | nbIndexServerReceived += count; |
---|
283 | repondAlreadyReceived.push_back(clientSourceRank); |
---|
284 | } |
---|
285 | } |
---|
286 | |
---|
287 | for (int i = 0; i< repondAlreadyReceived.size(); ++i) |
---|
288 | requestRecvInfo.erase(repondAlreadyReceived[i]); |
---|
289 | repondAlreadyReceived.resize(0); |
---|
290 | } |
---|
291 | |
---|
292 | globalIndexToInfoMappingLevel_ = indexToInfoMapping; |
---|
293 | info << "temp " << tmpGlobalIndexOnClient << std::endl; |
---|
294 | if (0 != maxNbIndexDemandedFromOthers) delete [] recvBuffIndex; |
---|
295 | if (0 != nbIndexReceivedFromOthers) delete [] recvBuffInfo; |
---|
296 | delete [] sendBuff; |
---|
297 | delete [] recvBuff; |
---|
298 | } |
---|
299 | |
---|
300 | /*! |
---|
301 | Compute the hash index distribution of whole size_t space then each client will have a range of this distribution |
---|
302 | */ |
---|
303 | void CClientClientDHT::computeHashIndex(std::vector<size_t>& hashedIndex, int nbClient) |
---|
304 | { |
---|
305 | // Compute range of hash index for each client |
---|
306 | hashedIndex.resize(nbClient+1); |
---|
307 | size_t nbHashIndexMax = std::numeric_limits<size_t>::max(); |
---|
308 | size_t nbHashIndex; |
---|
309 | hashedIndex[0] = 0; |
---|
310 | for (int i = 1; i < nbClient; ++i) |
---|
311 | { |
---|
312 | nbHashIndex = nbHashIndexMax / nbClient; |
---|
313 | if (i < (nbHashIndexMax%nbClient)) ++nbHashIndex; |
---|
314 | hashedIndex[i] = hashedIndex[i-1] + nbHashIndex; |
---|
315 | } |
---|
316 | hashedIndex[nbClient] = nbHashIndexMax; |
---|
317 | } |
---|
318 | |
---|
319 | /*! |
---|
320 | Compute distribution of global index for servers |
---|
321 | Each client already holds a piece of information and its attached index. |
---|
322 | This information will be redistributed among processes by projecting indices into size_t space. |
---|
323 | After the redistribution, each client holds rearranged index and its corresponding information. |
---|
324 | \param [in] indexInfoMap index and its corresponding info (usually server index) |
---|
325 | \param [in] commLevel communicator of current level |
---|
326 | \param [in] level current level |
---|
327 | */ |
---|
328 | void CClientClientDHT::computeDistributedIndex(const boost::unordered_map<size_t,int>& indexInfoMap, |
---|
329 | const MPI_Comm& commLevel, |
---|
330 | int level) |
---|
331 | { |
---|
332 | int nbClient, clientRank; |
---|
333 | MPI_Comm_size(commLevel,&nbClient); |
---|
334 | MPI_Comm_rank(commLevel,&clientRank); |
---|
335 | std::vector<size_t> hashedIndex; |
---|
336 | computeHashIndex(hashedIndex, nbClient); |
---|
337 | |
---|
338 | int* sendBuff = new int[nbClient]; |
---|
339 | int* sendNbIndexBuff = new int[nbClient]; |
---|
340 | for (int i = 0; i < nbClient; ++i) |
---|
341 | { |
---|
342 | sendBuff[i] = 0; sendNbIndexBuff[i] = 0; |
---|
343 | } |
---|
344 | |
---|
345 | // Compute size of sending and receving buffer |
---|
346 | std::map<int, std::vector<size_t> > client2ClientIndex; |
---|
347 | std::map<int, std::vector<int> > client2ClientInfo; |
---|
348 | |
---|
349 | std::vector<size_t>::const_iterator itbClientHash = hashedIndex.begin(), itClientHash, |
---|
350 | iteClientHash = hashedIndex.end(); |
---|
351 | boost::unordered_map<size_t,int>::const_iterator it = indexInfoMap.begin(), |
---|
352 | ite = indexInfoMap.end(); |
---|
353 | HashXIOS<size_t> hashGlobalIndex; |
---|
354 | for (; it != ite; ++it) |
---|
355 | { |
---|
356 | size_t hashIndex = hashGlobalIndex(it->first); |
---|
357 | itClientHash = std::upper_bound(itbClientHash, iteClientHash, hashIndex); |
---|
358 | if (itClientHash != iteClientHash) |
---|
359 | { |
---|
360 | int indexClient = std::distance(itbClientHash, itClientHash)-1; |
---|
361 | { |
---|
362 | sendBuff[indexClient] = 1; |
---|
363 | ++sendNbIndexBuff[indexClient]; |
---|
364 | client2ClientIndex[indexClient].push_back(it->first); |
---|
365 | client2ClientInfo[indexClient].push_back(it->second); |
---|
366 | } |
---|
367 | } |
---|
368 | } |
---|
369 | |
---|
370 | // Calculate from how many clients each client receive message. |
---|
371 | int* recvBuff = new int[nbClient]; |
---|
372 | MPI_Allreduce(sendBuff, recvBuff, nbClient, MPI_INT, MPI_SUM, commLevel); |
---|
373 | int recvNbClient = recvBuff[clientRank]; |
---|
374 | |
---|
375 | // Calculate size of buffer for receiving message |
---|
376 | int* recvNbIndexBuff = new int[nbClient]; |
---|
377 | MPI_Allreduce(sendNbIndexBuff, recvNbIndexBuff, nbClient, MPI_INT, MPI_SUM, commLevel); |
---|
378 | int recvNbIndexCount = recvNbIndexBuff[clientRank]; |
---|
379 | unsigned long* recvIndexBuff = new unsigned long[recvNbIndexCount]; |
---|
380 | int* recvInfoBuff = new int[recvNbIndexCount]; |
---|
381 | |
---|
382 | // If a client holds information about index and the corresponding which don't belong to it, |
---|
383 | // it will send a message to the correct clients. |
---|
384 | // Contents of the message are index and its corresponding informatioin |
---|
385 | std::list<MPI_Request> sendRequest; |
---|
386 | std::map<int, std::vector<size_t> >::iterator itIndex = client2ClientIndex.begin(), |
---|
387 | iteIndex = client2ClientIndex.end(); |
---|
388 | for (; itIndex != iteIndex; ++itIndex) |
---|
389 | sendIndexToClients(itIndex->first, itIndex->second, commLevel, sendRequest); |
---|
390 | std::map<int, std::vector<int> >::iterator itInfo = client2ClientInfo.begin(), |
---|
391 | iteInfo = client2ClientInfo.end(); |
---|
392 | for (; itInfo != iteInfo; ++itInfo) |
---|
393 | sendInfoToClients(itInfo->first, itInfo->second, commLevel, sendRequest); |
---|
394 | |
---|
395 | std::map<int, MPI_Request>::iterator itRequestIndex, itRequestInfo; |
---|
396 | std::map<int, int> countBuffInfo, countBuffIndex; |
---|
397 | std::vector<int> processedList; |
---|
398 | |
---|
399 | bool isFinished = (0 == recvNbClient) ? true : false; |
---|
400 | |
---|
401 | // Counting of buffer for receiving global index |
---|
402 | int countIndex = 0; |
---|
403 | |
---|
404 | // Counting of buffer for receiving server index |
---|
405 | int countInfo = 0; |
---|
406 | |
---|
407 | // Request returned by MPI_IRecv function about global index |
---|
408 | std::map<int, MPI_Request> requestRecvIndex, requestRecvInfo; |
---|
409 | |
---|
410 | // Mapping client rank and the beginning position of receiving buffer for message of global index from this client |
---|
411 | std::map<int, unsigned long*> indexBuffBegin; |
---|
412 | |
---|
413 | // Mapping client rank and the begining position of receiving buffer for message of server index from this client |
---|
414 | std::map<int, int*> infoBuffBegin; |
---|
415 | |
---|
416 | boost::unordered_map<size_t,int> indexToInfoMapping; |
---|
417 | |
---|
418 | // Now each client trys to listen to demand from others. |
---|
419 | // If they have message, it processes: pushing global index and corresponding server to its map |
---|
420 | while (!isFinished || (!sendRequest.empty())) |
---|
421 | { |
---|
422 | testSendRequest(sendRequest); |
---|
423 | probeIndexMessageFromClients(recvIndexBuff, recvNbIndexCount, |
---|
424 | countIndex, indexBuffBegin, |
---|
425 | requestRecvIndex, commLevel); |
---|
426 | // Processing complete request |
---|
427 | for (itRequestIndex = requestRecvIndex.begin(); |
---|
428 | itRequestIndex != requestRecvIndex.end(); |
---|
429 | ++itRequestIndex) |
---|
430 | { |
---|
431 | int rank = itRequestIndex->first; |
---|
432 | int count = computeBuffCountIndexGlobal(itRequestIndex->second); |
---|
433 | if (0 != count) |
---|
434 | countBuffIndex[rank] = count; |
---|
435 | } |
---|
436 | |
---|
437 | probeInfoMessageFromClients(recvInfoBuff, recvNbIndexCount, |
---|
438 | countInfo, infoBuffBegin, |
---|
439 | requestRecvInfo, commLevel); |
---|
440 | for (itRequestInfo = requestRecvInfo.begin(); |
---|
441 | itRequestInfo != requestRecvInfo.end(); |
---|
442 | ++itRequestInfo) |
---|
443 | { |
---|
444 | int rank = itRequestInfo->first; |
---|
445 | int count = computeBuffCountIndexServer(itRequestInfo->second); |
---|
446 | if (0 != count) |
---|
447 | countBuffInfo[rank] = count; |
---|
448 | } |
---|
449 | |
---|
450 | for (std::map<int, int>::iterator it = countBuffIndex.begin(); |
---|
451 | it != countBuffIndex.end(); ++it) |
---|
452 | { |
---|
453 | int rank = it->first; |
---|
454 | if ((countBuffInfo.end() != countBuffInfo.find(rank)) && |
---|
455 | (countBuffIndex.end() != countBuffIndex.find(rank))) |
---|
456 | { |
---|
457 | int count = it->second; |
---|
458 | for (int i = 0; i < count; ++i) |
---|
459 | indexToInfoMapping.insert(std::make_pair<size_t,int>(*(indexBuffBegin[rank]+i),*(infoBuffBegin[rank]+i))); |
---|
460 | processedList.push_back(rank); |
---|
461 | --recvNbClient; |
---|
462 | } |
---|
463 | } |
---|
464 | |
---|
465 | for (int i = 0; i < processedList.size(); ++i) |
---|
466 | { |
---|
467 | requestRecvInfo.erase(processedList[i]); |
---|
468 | requestRecvIndex.erase(processedList[i]); |
---|
469 | countBuffIndex.erase(processedList[i]); |
---|
470 | countBuffInfo.erase(processedList[i]); |
---|
471 | } |
---|
472 | |
---|
473 | if (0 == recvNbClient) isFinished = true; |
---|
474 | } |
---|
475 | |
---|
476 | delete [] sendBuff; |
---|
477 | delete [] sendNbIndexBuff; |
---|
478 | delete [] recvBuff; |
---|
479 | delete [] recvNbIndexBuff; |
---|
480 | delete [] recvIndexBuff; |
---|
481 | delete [] recvInfoBuff; |
---|
482 | |
---|
483 | // Ok, now do something recursive |
---|
484 | if (0 < level) |
---|
485 | { |
---|
486 | --level; |
---|
487 | computeDistributedIndex(indexToInfoMapping, commLevel_[level], level); |
---|
488 | } |
---|
489 | else |
---|
490 | globalIndexToServerMapping_ = indexToInfoMapping; |
---|
491 | } |
---|
492 | |
---|
493 | /*! |
---|
494 | Probe and receive message containg global index from other clients. |
---|
495 | Each client can send a message of global index to other clients to fulfill their maps. |
---|
496 | Each client probes message from its queue then if the message is ready, it will be put into the receiving buffer |
---|
497 | \param [in] recvIndexBuff buffer dedicated for receiving global index |
---|
498 | \param [in] recvNbIndexCount size of the buffer |
---|
499 | \param [in] countIndex number of received index |
---|
500 | \param [in] indexBuffBegin beginning of index buffer for each source rank |
---|
501 | \param [in] requestRecvIndex request of receving index |
---|
502 | \param [in] intraComm communicator |
---|
503 | */ |
---|
504 | void CClientClientDHT::probeIndexMessageFromClients(unsigned long* recvIndexBuff, |
---|
505 | const int recvNbIndexCount, |
---|
506 | int& countIndex, |
---|
507 | std::map<int, unsigned long*>& indexBuffBegin, |
---|
508 | std::map<int, MPI_Request>& requestRecvIndex, |
---|
509 | const MPI_Comm& intraComm) |
---|
510 | { |
---|
511 | MPI_Status statusIndexGlobal; |
---|
512 | int flagIndexGlobal, count; |
---|
513 | |
---|
514 | // Probing for global index |
---|
515 | MPI_Iprobe(MPI_ANY_SOURCE, MPI_DHT_INDEX, intraComm, &flagIndexGlobal, &statusIndexGlobal); |
---|
516 | if ((true == flagIndexGlobal) && (countIndex < recvNbIndexCount)) |
---|
517 | { |
---|
518 | MPI_Get_count(&statusIndexGlobal, MPI_UNSIGNED_LONG, &count); |
---|
519 | indexBuffBegin.insert(std::make_pair<int, unsigned long*>(statusIndexGlobal.MPI_SOURCE, recvIndexBuff+countIndex)); |
---|
520 | MPI_Irecv(recvIndexBuff+countIndex, count, MPI_UNSIGNED_LONG, |
---|
521 | statusIndexGlobal.MPI_SOURCE, MPI_DHT_INDEX, intraComm, |
---|
522 | &requestRecvIndex[statusIndexGlobal.MPI_SOURCE]); |
---|
523 | countIndex += count; |
---|
524 | } |
---|
525 | } |
---|
526 | |
---|
527 | /*! |
---|
528 | Probe and receive message containg server index from other clients. |
---|
529 | Each client can send a message of server index to other clients to fulfill their maps. |
---|
530 | Each client probes message from its queue then if the message is ready, it will be put into the receiving buffer |
---|
531 | \param [in] recvInfoBuff buffer dedicated for receiving server index |
---|
532 | \param [in] recvNbIndexCount size of the buffer |
---|
533 | \param [in] countInfo number of received info |
---|
534 | \param [in] infoBuffBegin beginning of index buffer for each source rank |
---|
535 | \param [in] requestRecvInfo request of receving index |
---|
536 | \param [in] intraComm communicator |
---|
537 | */ |
---|
538 | void CClientClientDHT::probeInfoMessageFromClients(int* recvInfoBuff, |
---|
539 | const int recvNbIndexCount, |
---|
540 | int& countInfo, |
---|
541 | std::map<int, int*>& infoBuffBegin, |
---|
542 | std::map<int, MPI_Request>& requestRecvInfo, |
---|
543 | const MPI_Comm& intraComm) |
---|
544 | { |
---|
545 | MPI_Status statusInfo; |
---|
546 | int flagInfo, count; |
---|
547 | |
---|
548 | // Probing for server index |
---|
549 | MPI_Iprobe(MPI_ANY_SOURCE, MPI_DHT_INFO, intraComm, &flagInfo, &statusInfo); |
---|
550 | if ((true == flagInfo) && (countInfo < recvNbIndexCount)) |
---|
551 | { |
---|
552 | MPI_Get_count(&statusInfo, MPI_INT, &count); |
---|
553 | infoBuffBegin.insert(std::make_pair<int, int*>(statusInfo.MPI_SOURCE, recvInfoBuff+countInfo)); |
---|
554 | MPI_Irecv(recvInfoBuff+countInfo, count, MPI_INT, |
---|
555 | statusInfo.MPI_SOURCE, MPI_DHT_INFO, intraComm, |
---|
556 | &requestRecvInfo[statusInfo.MPI_SOURCE]); |
---|
557 | |
---|
558 | countInfo += count; |
---|
559 | } |
---|
560 | } |
---|
561 | |
---|
562 | /*! |
---|
563 | Send message containing index to clients |
---|
564 | \param [in] clientDestRank rank of destination client |
---|
565 | \param [in] indices index to send |
---|
566 | \param [in] clientIntraComm communication group of client |
---|
567 | \param [in] requestSendIndex list of sending request |
---|
568 | */ |
---|
569 | void CClientClientDHT::sendIndexToClients(int clientDestRank, std::vector<size_t>& indices, |
---|
570 | const MPI_Comm& clientIntraComm, |
---|
571 | std::list<MPI_Request>& requestSendIndex) |
---|
572 | { |
---|
573 | MPI_Request request; |
---|
574 | requestSendIndex.push_back(request); |
---|
575 | MPI_Isend(&(indices)[0], (indices).size(), MPI_UNSIGNED_LONG, |
---|
576 | clientDestRank, MPI_DHT_INDEX, clientIntraComm, &(requestSendIndex.back())); |
---|
577 | } |
---|
578 | |
---|
579 | /*! |
---|
580 | Send message containing information to clients |
---|
581 | \param [in] clientDestRank rank of destination client |
---|
582 | \param [in] info server index to send |
---|
583 | \param [in] clientIntraComm communication group of client |
---|
584 | \param [in] requestSendInfo list of sending request |
---|
585 | */ |
---|
586 | void CClientClientDHT::sendInfoToClients(int clientDestRank, std::vector<int>& info, |
---|
587 | const MPI_Comm& clientIntraComm, |
---|
588 | std::list<MPI_Request>& requestSendInfo) |
---|
589 | { |
---|
590 | MPI_Request request; |
---|
591 | requestSendInfo.push_back(request); |
---|
592 | MPI_Isend(&(info)[0], (info).size(), MPI_INT, |
---|
593 | clientDestRank, MPI_DHT_INFO, clientIntraComm, &(requestSendInfo.back())); |
---|
594 | } |
---|
595 | |
---|
596 | /*! |
---|
597 | Verify status of sending request |
---|
598 | \param [in] sendRequest sending request to verify |
---|
599 | */ |
---|
600 | void CClientClientDHT::testSendRequest(std::list<MPI_Request>& sendRequest) |
---|
601 | { |
---|
602 | int flag = 0; |
---|
603 | MPI_Status status; |
---|
604 | std::list<MPI_Request>::iterator itRequest; |
---|
605 | int sizeListRequest = sendRequest.size(); |
---|
606 | int idx = 0; |
---|
607 | while (idx < sizeListRequest) |
---|
608 | { |
---|
609 | bool isErased = false; |
---|
610 | for (itRequest = sendRequest.begin(); itRequest != sendRequest.end(); ++itRequest) |
---|
611 | { |
---|
612 | MPI_Test(&(*itRequest), &flag, &status); |
---|
613 | if (true == flag) |
---|
614 | { |
---|
615 | isErased = true; |
---|
616 | break; |
---|
617 | } |
---|
618 | } |
---|
619 | if (true == isErased) sendRequest.erase(itRequest); |
---|
620 | ++idx; |
---|
621 | } |
---|
622 | } |
---|
623 | |
---|
624 | /*! |
---|
625 | Process the received request. Pushing global index and server index into map |
---|
626 | \param[in] buffIndexGlobal pointer to the begining of buffer containing global index |
---|
627 | \param[in] buffIndexServer pointer to the begining of buffer containing server index |
---|
628 | \param[in] count size of received message |
---|
629 | */ |
---|
630 | //void CClientClientDHT::processReceivedRequest(unsigned long* buffIndexGlobal, int* buffIndexServer, int count) |
---|
631 | //{ |
---|
632 | // for (int i = 0; i < count; ++i) |
---|
633 | // globalIndexToServerMapping_.insert(std::make_pair<size_t,int>(*(buffIndexGlobal+i),*(buffIndexServer+i))); |
---|
634 | //} |
---|
635 | |
---|
636 | /*! |
---|
637 | Compute size of message containing global index |
---|
638 | \param[in] requestRecv request of message |
---|
639 | */ |
---|
640 | int CClientClientDHT::computeBuffCountIndexGlobal(MPI_Request& requestRecv) |
---|
641 | { |
---|
642 | int flag, count = 0; |
---|
643 | MPI_Status status; |
---|
644 | |
---|
645 | MPI_Test(&requestRecv, &flag, &status); |
---|
646 | if (true == flag) |
---|
647 | { |
---|
648 | MPI_Get_count(&status, MPI_UNSIGNED_LONG, &count); |
---|
649 | } |
---|
650 | |
---|
651 | return count; |
---|
652 | } |
---|
653 | |
---|
654 | /*! |
---|
655 | Compute size of message containing server index |
---|
656 | \param[in] requestRecv request of message |
---|
657 | */ |
---|
658 | int CClientClientDHT::computeBuffCountIndexServer(MPI_Request& requestRecv) |
---|
659 | { |
---|
660 | int flag, count = 0; |
---|
661 | MPI_Status status; |
---|
662 | |
---|
663 | MPI_Test(&requestRecv, &flag, &status); |
---|
664 | if (true == flag) |
---|
665 | { |
---|
666 | MPI_Get_count(&status, MPI_INT, &count); |
---|
667 | } |
---|
668 | |
---|
669 | return count; |
---|
670 | } |
---|
671 | |
---|
672 | } |
---|