tapioca.cpp 23.4 KB
Newer Older
1
#include "tapioca.hpp"
Francois Tessier's avatar
Francois Tessier committed
2

Francois Tessier's avatar
Francois Tessier committed
3
Tapioca::Tapioca ()
Francois Tessier's avatar
Francois Tessier committed
4 5 6 7 8 9
{
  this->SetDefaultValues ();
  this->ParseEnvVariables ();
}


Francois Tessier's avatar
Francois Tessier committed
10
Tapioca::~Tapioca ()
Francois Tessier's avatar
Francois Tessier committed
11 12 13 14
{
}


15 16 17 18 19 20 21
void Tapioca::Finalize ()
{
  this->chunksIndexMatching.clear();
  free (this->chunkCount_);
  free (this->chunkSize_);
  free (this->chunkOffset_);
  this->excludedNode.clear();
Francois Tessier's avatar
Francois Tessier committed
22

23 24
  this->commDataSize_ = 0;

Francois Tessier's avatar
Francois Tessier committed
25 26 27
  MPI_Win_free (&this->RMAWin1);
  MPI_Win_free (&this->RMAWin2);

28 29
  MPI_Comm_free (&this->subComm_);

Francois Tessier's avatar
Francois Tessier committed
30 31 32 33 34
  free (this->buffer1);
  free (this->buffer2);
}


35
void Tapioca::GlobalFence () 
Francois Tessier's avatar
Francois Tessier committed
36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68
{
  int buffer;

  buffer = this->currentRound_ % NBUFFERS;

  switch (buffer)
    {
    case 0:
      MPI_Win_fence (0, this->RMAWin1);
      break;
    case 1:
      MPI_Win_fence (0, this->RMAWin2);
      break;
    }
  
#ifdef TIMING
  this->endAggrTime = MPI_Wtime();
  
  if ( this->startAggrTime != 0 ) {
    this->totAggrTime = this->endAggrTime - this->startAggrTime;
    // fprintf (stdout, "[TIMING][AGG][AGGR] Rank %d, Rnd %d - %.2f ms\n",
    // 	     this->commRank_, this->currentRound_, this->totAggrTime * 1000);
  }
    
  this->startAggrTime = 0;
#endif

}


/***********************/
/*  INITIALIZATION     */
/***********************/
Francois Tessier's avatar
Francois Tessier committed
69
void Tapioca::SetDefaultValues ()
Francois Tessier's avatar
Francois Tessier committed
70 71 72 73
{
  this->rankDataSize_ = 0;
  this->strategy_ = SHORTEST_PATH;
  this->nAggr_ = 8;
74
  this->bufferSize_ = 16777216;
Francois Tessier's avatar
Francois Tessier committed
75
  this->amAnAggr_ = false;
76
  this->commSplit_ = true;
Francois Tessier's avatar
Francois Tessier committed
77
  this->currentRound_ = 0;
78
  this->totalRounds_ = 0;
Francois Tessier's avatar
Francois Tessier committed
79 80 81
  this->currentDataSize_ = 0;
  this->nCommit_ = 0;
  this->writeDevNull_ = false;
82
  this->pipelinedBuffers_ = true;
Francois Tessier's avatar
Francois Tessier committed
83
  /* DEBUG */
Francois Tessier's avatar
Francois Tessier committed
84 85 86
}


Francois Tessier's avatar
Francois Tessier committed
87
void Tapioca::ParseEnvVariables ()
Francois Tessier's avatar
Francois Tessier committed
88
{
89 90 91 92 93
  char *envStrategy   = getenv("TAPIOCA_STRATEGY");
  char *envNAggr      = getenv("TAPIOCA_NBAGGR");
  char *envBufferSize = getenv("TAPIOCA_BUFFERSIZE");
  char *envSplit      = getenv("TAPIOCA_COMMSPLIT");
  char *envDevNull    = getenv("TAPIOCA_DEVNULL");
94
  char *envPipelining = getenv("TAPIOCA_PIPELINING");
Francois Tessier's avatar
Francois Tessier committed
95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120
  
  if (envStrategy != NULL) {
    strcmp(envStrategy, "SHORTEST_PATH")    ? 0 : this->strategy_ = SHORTEST_PATH;
    strcmp(envStrategy, "LONGEST_PATH")     ? 0 : this->strategy_ = LONGEST_PATH;
    strcmp(envStrategy, "TOPOLOGY_AWARE")   ? 0 : this->strategy_ = TOPOLOGY_AWARE;
    strcmp(envStrategy, "CONTENTION_AWARE") ? 0 : this->strategy_ = CONTENTION_AWARE;
    strcmp(envStrategy, "UNIFORM")          ? 0 : this->strategy_ = UNIFORM;
  }

  if (envNAggr != NULL) {
    this->nAggr_ = atoi(envNAggr);
  }

  if (envBufferSize != NULL) {
    this->bufferSize_ = atoi(envBufferSize);
  }

  if (envSplit != NULL) {
    strcmp(envSplit, "true")  ? 0 : this->commSplit_ = true;
    strcmp(envSplit, "false") ? 0 : this->commSplit_ = false;
  }

  if (envDevNull != NULL) {
    strcmp(envDevNull, "true")  ? 0 : this->writeDevNull_ = true;
    strcmp(envDevNull, "false") ? 0 : this->writeDevNull_ = false;
  }
121 122 123 124 125

  if (envPipelining != NULL) {
    strcmp(envPipelining, "true")  ? 0 : this->pipelinedBuffers_ = true;
    strcmp(envPipelining, "false") ? 0 : this->pipelinedBuffers_ = false;
  }
Francois Tessier's avatar
Francois Tessier committed
126 127 128
}


Francois Tessier's avatar
Francois Tessier committed
129
void Tapioca::SetCommValues ()
Francois Tessier's avatar
Francois Tessier committed
130 131 132 133 134 135 136 137 138
{
  MPI_Comm_rank (this->subComm_, &this->commRank_);
  MPI_Comm_size (this->subComm_, &this->commSize_);
  MPI_Comm_rank (MPI_COMM_WORLD, &this->worldRank_);

  MPI_Allreduce (&this->rankDataSize_, &this->commDataSize_, 1, MPI_LONG_LONG, MPI_SUM, this->subComm_);
}


Francois Tessier's avatar
Francois Tessier committed
139
void Tapioca::SetOffsets ()
Francois Tessier's avatar
Francois Tessier committed
140 141 142 143 144 145 146 147
{
  MPI_Exscan (&this->rankDataSize_, &this->offsetInAggrData_, 1, 
	      MPI_LONG_LONG, MPI_SUM, this->subComm_);
  if (this->commRank_ == 0)
    this->offsetInAggrData_ = 0;
}


Francois Tessier's avatar
Francois Tessier committed
148
void Tapioca::SetNodesList ()
Francois Tessier's avatar
Francois Tessier committed
149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164
{
  int *coords, *myCoords, i, worldSize, dimensions;
  
  MPI_Comm_size ( MPI_COMM_WORLD, &worldSize );
  coords = (int *)malloc (worldSize * sizeof (int));

  dimensions = topology.NetworkDimensions () + 1;
  myCoords = (int *)malloc (dimensions * sizeof (int));

  topology.RankToCoordinates (this->worldRank_, myCoords);
  this->intCoords_ = this->CoordsToInt (myCoords, dimensions - 1);
  
  MPI_Allgather(&this->intCoords_, 1, MPI_INT, coords, 1, MPI_INT, MPI_COMM_WORLD);

  for ( i = 0; i < worldSize; i++ )
    this->excludedNode[coords[i]] = false;
165 166 167

  free (coords);
  free (myCoords);
Francois Tessier's avatar
Francois Tessier committed
168 169 170 171 172 173
}


/***********************/
/*    AGGREGATION      */
/***********************/
Francois Tessier's avatar
Francois Tessier committed
174
int Tapioca::NumberOfAggregators ()
Francois Tessier's avatar
Francois Tessier committed
175 176 177 178 179
{
  return 0;
}


Francois Tessier's avatar
Francois Tessier committed
180
void Tapioca::IdentifyMyAggregators ()
Francois Tessier's avatar
Francois Tessier committed
181 182 183 184 185
{
  int i, j, c, globalRoundId, upperBound, index = 0;
  int64_t remainingData, offsetInAggrData;
  std::vector<Round_t> rounds;
  
186
  this->totalNeededBuffers_  = ceil ( (double)this->commDataSize_ / (double)this->bufferSize_ );
187
  this->totalRounds_         = ceil ( (double)this->totalNeededBuffers_ / (double)this->nAggr_ );
Francois Tessier's avatar
Francois Tessier committed
188

189
  for ( i = 0; i < this->totalNeededBuffers_; i++ ) {
Francois Tessier's avatar
Francois Tessier committed
190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222
    Round_t r;
    r.aggr = i % this->nAggr_;
    r.round = i / this->nAggr_;
    rounds.push_back(r);
  }

  for ( c = 0; c < this->nChunks_; c++ ) {
    remainingData    = this->chunkCount_[c] * this->chunkSize_[c];
    offsetInAggrData = this->chunkOffset_[c] - this->offsetInFile_;
    globalRoundId = floor ( offsetInAggrData / this->bufferSize_ );
    upperBound = ( offsetInAggrData % this->bufferSize_ ) + remainingData;
    
    while ( remainingData > 0 ) {
      this->globalAggregatorsRanks.push_back ( rounds[globalRoundId].aggr );
      this->roundsIds.push_back ( rounds[globalRoundId].round );
      this->chunksIndexMatching[c].push_back ( index );
      
      if ( upperBound > this->bufferSize_) {
	remainingData = (upperBound - this->bufferSize_);
	this->dataSize.push_back ( this->chunkCount_[c] * this->chunkSize_[c] - remainingData );
      }
      else {
	this->dataSize.push_back ( remainingData );
	remainingData = 0;
      }
      
      upperBound -= this->bufferSize_;
      globalRoundId++;
      index++;
    }
  }
  
#ifdef DEBUG
Francois Tessier's avatar
Francois Tessier committed
223
  if (this->commRank_ == 4) {
Francois Tessier's avatar
Francois Tessier committed
224
    fprintf (stdout, "[DEBUG] Rounds distrib. on %d aggregators: AGG ", this->nAggr_);
225
    for ( i = 0; i < this->totalNeededBuffers_; i++ )
Francois Tessier's avatar
Francois Tessier committed
226 227
      fprintf (stdout, "%d ", rounds[i].aggr);
    fprintf (stdout, "\n                                          RND ");
228
    for ( i = 0; i < this->totalNeededBuffers_; i++ )
Francois Tessier's avatar
Francois Tessier committed
229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251
      fprintf (stdout, "%d ", rounds[i].round);
    fprintf (stdout, "\n                                          AGG ");
    for ( i = 0; i < this->globalAggregatorsRanks.size(); i++ )
      fprintf (stdout, "%d ", this->globalAggregatorsRanks[i]);
    fprintf (stdout, "\n                                          RID ");
    for ( i = 0; i < this->roundsIds.size(); i++ )
      fprintf (stdout, "%d ", this->roundsIds[i]);
    fprintf (stdout, "\n                                          DAS ");
    for ( i = 0; i < this->dataSize.size(); i++ )
      fprintf (stdout, "%d ", this->dataSize[i]);
    fprintf (stdout, "\n                                          CIM ");
    for ( i = 0; i < this->chunksIndexMatching.size(); i++ ) {
      fprintf (stdout, "{ ");
      for ( j = 0; j < this->chunksIndexMatching[i].size(); j++ )
	fprintf (stdout, "%d ", this->chunksIndexMatching[i][j]);
      fprintf (stdout, "}");
    }
    fprintf (stdout, "\n");
  }
#endif
}


Francois Tessier's avatar
Francois Tessier committed
252
void Tapioca::ElectAggregators ()
Francois Tessier's avatar
Francois Tessier committed
253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371
{
  int aggr, aggrRank, rankAggrComm, sizeAggrComm, aggrRankAggrComm, i, j, aggrCoords, worldSize;
  int64_t color;
  double startTime, endTime, totTime = 0.0;
  MPI_Comm aggrComm;
  MPI_Request request;
  MPI_Status status;
  /* Groups */
  MPI_Group commGroup, aggrGroup;
  int *ranks, *join, *groupRanks, groupSize, groupRank, joinGroup;

  MPI_Comm_size (MPI_COMM_WORLD, &worldSize);

  this->aggregatorsRanks.resize ( this->globalAggregatorsRanks.size() );
  
  for ( aggr = 0; aggr < this->nAggr_; aggr++ ) {
    color = this->DataSizeSentToAggr (aggr);
    
    startTime = MPI_Wtime();

    if ( this->commSplit_ )
      MPI_Comm_split (this->subComm_, color > 0, this->commRank_, &aggrComm);
    else {
      
      MPI_Comm_group (this->subComm_, &commGroup);

      if ( color > 0 ) {
	joinGroup = 1;
	MPI_Allreduce (&joinGroup, &groupSize, 1, MPI_INT, MPI_SUM, this->subComm_ );
      }
      else {
	joinGroup = 0;
	MPI_Allreduce (&joinGroup, &groupSize, 1, MPI_INT, MPI_SUM, this->subComm_ );
	groupSize = this->commSize_ - groupSize;
      }

      join = (int *)malloc (this->commSize_ * sizeof (int));
      MPI_Allgather(&joinGroup, 1, MPI_INT, join, 1, MPI_INT, this->subComm_);
      groupRanks = (int *)malloc (groupSize * sizeof (int));
      
      j = 0;
      for ( i = 0; i < this->commSize_; i++ ) {
	if (join[i] == joinGroup) {
	  groupRanks[j] = i;
	  j++;
	}
      }

      MPI_Group_incl (commGroup, groupSize, groupRanks, &aggrGroup);
      MPI_Comm_create (this->subComm_, aggrGroup, &aggrComm);
      
      free (join);
      free (groupRanks);
    }
    
    MPI_Comm_rank (aggrComm, &rankAggrComm);
    MPI_Comm_size (aggrComm, &sizeAggrComm);

    endTime = MPI_Wtime();
    totTime += ( endTime - startTime );
    
    if ( color > 0 ) {
      switch ( this->strategy_ )
	{
	case SHORTEST_PATH:
	  aggrRank = this->RankShortestPath (aggrComm, color);
	  break;
	case LONGEST_PATH:
	  aggrRank = this->RankLongestPath (aggrComm, color);
	  break;
	case TOPOLOGY_AWARE:
	  aggrRank = this->RankTopologyAware (aggrComm, color);
	  break;
	case CONTENTION_AWARE:
	  aggrRank = this->RankContentionAware (aggrComm, color);
	  break;
	case UNIFORM:
	  aggrRank = this->RankUniformDistribution (aggrComm, color);
	  break;
	}

      if ( this->commRank_ == aggrRank ) {
	this->totalWrites_     = ceil ( (double)this->aggrDataSize_ / (double)this->bufferSize_);
	this->globalAggrRank_  = aggr;
	aggrCoords             = this->intCoords_;
	MPI_Isend ( &aggrCoords, 1, MPI_INT, 0, 0, MPI_COMM_WORLD, &request );
      }

      for ( i = 0; i < this->aggregatorsRanks.size(); i++ )
	if ( this->globalAggregatorsRanks[i] == aggr )
	  this->aggregatorsRanks[i] = aggrRank;

#ifdef DEBUG
      int coords[topology.NetworkDimensions() + 1];
      if ( this->commRank_ == aggrRank ) {
	topology.RankToCoordinates (this->worldRank_, coords);
	fprintf (stdout, "[DEBUG] AggRank %d, (", this->worldRank_);
	for ( i = 0; i < topology.NetworkDimensions() + 1; i++ )
	  fprintf (stdout, "%u ", coords[i]);
	fprintf (stdout, ") -> %d, part %d, %lld B from %d ranks, %d rounds\n",
		 this->intCoords_, topology.BridgeNodeId(), this->aggrDataSize_, sizeAggrComm, this->totalWrites_);
      }
#endif
    }

    /*
     *  TODO: Compute this... 
     *              - Vesta: 16 nodes/bridge node
     *              - Mira : 64 nodes/bridge node
     */
    for ( i = 0; i < ( (worldSize / topology.ProcessPerNode ()) / (this->commSize_ / topology.ProcessPerNode ()) ); i++ ) {
      if ( this->worldRank_ == 0 )
	MPI_Recv ( &aggrCoords, 1, MPI_INT, MPI_ANY_SOURCE, 0, MPI_COMM_WORLD, &status );
      
      MPI_Bcast ( &aggrCoords, 1, MPI_INT, 0, MPI_COMM_WORLD);
      this->excludedNode[aggrCoords] = true;
    }
  }

Francois Tessier's avatar
Francois Tessier committed
372
#ifdef TIMING
Francois Tessier's avatar
Francois Tessier committed
373
  this->PrintTime( 0, totTime, " |-> Create subcommunicator");
Francois Tessier's avatar
Francois Tessier committed
374
#endif
Francois Tessier's avatar
Francois Tessier committed
375 376 377
}


Francois Tessier's avatar
Francois Tessier committed
378
int64_t Tapioca::DataSizeSentToAggr (int aggrId)
Francois Tessier's avatar
Francois Tessier committed
379 380 381 382 383 384 385 386 387 388 389 390
{
  int i;
  int64_t dataSize = 0;

  for ( i = 0; i < this->aggregatorsRanks.size(); i++ )
    if ( this->globalAggregatorsRanks[i] == aggrId )
      dataSize += this->dataSize[i];

  return dataSize;
}


Francois Tessier's avatar
Francois Tessier committed
391
void Tapioca::InitAggregators () 
Francois Tessier's avatar
Francois Tessier committed
392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427
{
  int aggr, retval;
  
  if ( this->amAnAggr_ ) {
    this->buffer1 = malloc (this->bufferSize_);
    this->buffer2 = malloc (this->bufferSize_);

    retval = MPI_Win_create (this->buffer1, this->bufferSize_, 1, MPI_INFO_NULL, this->subComm_, &this->RMAWin1);
    this->HandleMPIError (retval);
    MPI_Win_create (this->buffer2, this->bufferSize_, 1, MPI_INFO_NULL, this->subComm_, &this->RMAWin2);
    this->HandleMPIError (retval);
  }
  else {
    retval = MPI_Win_create (NULL, 0, 1, MPI_INFO_NULL, this->subComm_, &this->RMAWin1);
    this->HandleMPIError (retval);
    retval = MPI_Win_create (NULL, 0, 1, MPI_INFO_NULL, this->subComm_, &this->RMAWin2);
    this->HandleMPIError (retval);
  }

  retval = MPI_Win_fence (0, this->RMAWin1);
  this->HandleMPIError (retval);
  retval = MPI_Win_fence (0, this->RMAWin2);
  this->HandleMPIError (retval);

#ifdef DEBUG
  if (this->commRank_ == MASTER) {
    fprintf (stdout, "[DEBUG] %d RMA windows created (%d aggr., %d buffers)\n", 
	     NBUFFERS, this->nAggr_, NBUFFERS);
  }
#endif
}


/***********************/
/*     PLACEMENT       */
/***********************/
Francois Tessier's avatar
Francois Tessier committed
428
int Tapioca::RankShortestPath (MPI_Comm aggrComm, int64_t dataSize)
Francois Tessier's avatar
Francois Tessier committed
429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460
{
  int commRank, aggrRank, aggrPrank, ppn, nodeId;
  struct { int hops; int rank; } hopsToIONnode, shortestPath;

  MPI_Comm_rank (aggrComm, &commRank);
  
  hopsToIONnode.hops = topology.DistanceToIONode ( this->worldRank_ );
  hopsToIONnode.rank = commRank;

  if ( this->excludedNode[this->intCoords_] )
    hopsToIONnode.hops = INT_MAX;

  MPI_Allreduce ( &hopsToIONnode, &shortestPath, 1, MPI_2INTEGER, MPI_MINLOC, aggrComm );
  MPI_Reduce ( &dataSize, &this->aggrDataSize_, 1, MPI_LONG_LONG, MPI_SUM, shortestPath.rank, aggrComm );

  if ( shortestPath.rank == commRank ) {
    aggrRank = this->commRank_;
    this->amAnAggr_ = true;
  }

#ifdef DEBUG
  if ( shortestPath.rank == commRank )
    fprintf (stdout, "[DEBUG] Aggr. rank %d in aggrComm, distance to I/O node %d hops\n",
	     shortestPath.rank, shortestPath.hops);
#endif
  
  MPI_Bcast ( &aggrRank, 1, MPI_INT, shortestPath.rank, aggrComm);

  return aggrRank;
}


Francois Tessier's avatar
Francois Tessier committed
461
int Tapioca::RankLongestPath (MPI_Comm aggrComm, int64_t dataSize)
Francois Tessier's avatar
Francois Tessier committed
462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493
{
  int commRank, aggrRank;
  struct { int hops; int rank; } hopsToIONnode, longestPath;

  MPI_Comm_rank (aggrComm, &commRank);
  
  hopsToIONnode.hops = topology.DistanceToIONode ( this->worldRank_ );
  hopsToIONnode.rank = commRank;

  if ( this->excludedNode[this->intCoords_] )
    hopsToIONnode.hops = INT_MIN;
  
  MPI_Allreduce ( &hopsToIONnode, &longestPath, 1, MPI_2INTEGER, MPI_MAXLOC, aggrComm );
  MPI_Reduce ( &dataSize, &this->aggrDataSize_, 1, MPI_LONG_LONG, MPI_SUM, longestPath.rank, aggrComm );

  if ( longestPath.rank == commRank ) {
    aggrRank  = this->commRank_;
    this->amAnAggr_ = true;
  }

#ifdef DEBUG
  if ( longestPath.rank == commRank )
    fprintf (stdout, "[DEBUG] Aggr. rank %d in aggrComm, distance to I/O node %d hops\n",
	     longestPath.rank, longestPath.hops);
#endif
  
  MPI_Bcast ( &aggrRank, 1, MPI_INT, longestPath.rank, aggrComm);

  return aggrRank;
}


Francois Tessier's avatar
Francois Tessier committed
494
int Tapioca::RankTopologyAware (MPI_Comm aggrComm, int64_t dataSize)
Francois Tessier's avatar
Francois Tessier committed
495 496
{
  struct { double cost; int rank; } aggrCost, minCost;
497
  int aggrCommRank, aggrCommSize, worldRank, rank, distance, dim, hops, aggrRank, nIOnodes;
Francois Tessier's avatar
Francois Tessier committed
498
  int64_t *dataDistrib, aggregatedData = 0;
499
  int *srcCoords, *destCoords, *globalRanks, *IOnodesList;
Francois Tessier's avatar
Francois Tessier committed
500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516
  MPI_Comm_rank (aggrComm, &aggrCommRank);
  MPI_Comm_size (aggrComm, &aggrCommSize);
  MPI_Comm_rank (MPI_COMM_WORLD, &worldRank);

  aggrCost.rank = aggrCommRank;
  aggrCost.cost = 0;

  dataDistrib = (int64_t*) malloc (aggrCommSize * sizeof(int64_t));
  globalRanks = (int *) malloc (aggrCommSize * sizeof(int));
  MPI_Allgather(&worldRank, 1, MPI_INT, globalRanks, 1, MPI_INT, aggrComm);
  MPI_Allgather(&dataSize, 1, MPI_LONG_LONG, dataDistrib, 1, MPI_LONG_LONG, aggrComm);

  for ( rank = 0; rank < aggrCommSize; rank++ ) {
    aggregatedData += dataDistrib[rank];
    if ( rank != aggrCommRank ) {
      distance = topology.DistanceBetweenRanks ( globalRanks[rank], worldRank );

517 518 519
      // aggrCost.cost = std::max ( distance * LATENCY + (double)dataDistrib[rank] / BANDWIDTH,
      // 			 aggrCost.cost );
      aggrCost.cost += (distance * LATENCY + (double)dataDistrib[rank] / BANDWIDTH);
Francois Tessier's avatar
Francois Tessier committed
520 521
    }
  }
522

523 524
  // IOnodesList = (int *) malloc ( MAX_IONODES * sizeof ( int ) );
  // nIOnodes = topology.IONodesPerFile (this->filename_, IOnodesList);
525

526 527 528 529 530 531 532
  // if ( this->commRank_ == 0 ) {
  //   fprintf (stdout, "[LUSTRE] nLnet = %d\n", nIOnodes);
  //   fprintf (stdout, "[LUSTRE] list = ");
  //   for ( int i = 0; i < nIOnodes; i++ )
  //     fprintf (stdout, "%d ", IOnodesList[i]);
  //   fprintf (stdout, "\n");
  // }
533 534
    
  //aggrCost.cost += topology.DistanceToIONode ( worldRank ) * LATENCY + (double)aggregatedData / BANDWIDTH;
Francois Tessier's avatar
Francois Tessier committed
535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559

  if ( this->excludedNode[this->intCoords_] )
    aggrCost.cost = DBL_MAX;

  MPI_Allreduce ( &aggrCost, &minCost, 1, MPI_DOUBLE_INT, MPI_MINLOC, aggrComm );

  MPI_Reduce ( &dataSize, &this->aggrDataSize_, 1, MPI_LONG_LONG, MPI_SUM, minCost.rank, aggrComm );

  if ( minCost.rank == aggrCommRank ) {
    aggrRank = this->commRank_;
    this->amAnAggr_ = true;
  }

  MPI_Bcast ( &aggrRank, 1, MPI_INT, minCost.rank, aggrComm );

#ifdef DEBUG
  if ( minCost.rank == aggrCommRank )
    fprintf (stdout, "[DEBUG] Aggr. rank %d in aggrComm, distance to I/O node %d hops, cost: %.4f\n",
	     minCost.rank, topology.DistanceToIONode ( worldRank ), minCost.cost );
#endif

  return aggrRank;
}


Francois Tessier's avatar
Francois Tessier committed
560
int Tapioca::RankUniformDistribution (MPI_Comm aggrComm, int64_t dataSize)
Francois Tessier's avatar
Francois Tessier committed
561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591
{
  int aggrRank, aggrCommRank, rootRank = 0, rootCoords;
  
  MPI_Comm_rank (aggrComm, &aggrCommRank);

  if ( aggrCommRank == rootRank )
    rootCoords = this->intCoords_;
  
  MPI_Bcast ( &rootCoords, 1, MPI_INT, rootRank, aggrComm);
  
  while ( this->excludedNode[rootCoords] ) {
    rootRank += topology.ProcessPerNode ();
    if ( aggrCommRank == rootRank )
      rootCoords = this->intCoords_;

    MPI_Bcast ( &rootCoords, 1, MPI_INT, rootRank, aggrComm);
  }

  MPI_Reduce ( &dataSize, &this->aggrDataSize_, 1, MPI_LONG_LONG, MPI_SUM, rootRank, aggrComm );

  if ( aggrCommRank == rootRank ) {
    aggrRank = this->commRank_;
    this->amAnAggr_ = true;
  }    
  
  MPI_Bcast ( &aggrRank, 1, MPI_INT, rootRank, aggrComm);
  
  return aggrRank;
}


Francois Tessier's avatar
Francois Tessier committed
592
int Tapioca::RankContentionAware (MPI_Comm aggrComm, int64_t dataSize)
Francois Tessier's avatar
Francois Tessier committed
593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717
{
  struct { double cost; int rank; } aggrCost, minCost;
  int aggrCommRank, aggrCommSize, worldRank, rank, distance, interRanks;
  int aggrRank, dim, hops, ppn, intaggrRank, node, i;
  int64_t *dataDistrib, aggregatedData = 0;
  int *srcCoords, *destCoords, *globalRanks;
  
  char *split;
  int srcNode, destNode, interNode, srcLink, destLink;
  std::map< int, std::map<int, int> > links;
  std::map< int, std::vector<int> > routes;
  std::map< int, int > routeCost;
  int *path;

  MPI_Comm_rank (aggrComm, &aggrCommRank);
  MPI_Comm_size (aggrComm, &aggrCommSize);
  MPI_Comm_rank (MPI_COMM_WORLD, &worldRank);

  aggrCost.rank = aggrCommRank;
  aggrCost.cost = 0;

  ppn = topology.ProcessPerNode ();

  dataDistrib = (int64_t*) malloc (aggrCommSize * sizeof(int64_t));
  globalRanks = (int *) malloc (aggrCommSize * sizeof(int));
  MPI_Allgather(&worldRank, 1, MPI_INT, globalRanks, 1, MPI_INT, aggrComm);
  MPI_Allgather(&dataSize, 1, MPI_LONG_LONG, dataDistrib, 1, MPI_LONG_LONG, aggrComm);

  destNode = globalRanks[aggrCommRank]/ppn;
  
  /*
   *  Contention:
   *     - routes: map < srcNode, {<intermediate nodes>}
   *     - links: map < node, < node, weight > >
   */
  for ( rank = 0; rank < aggrCommSize; rank++ ) {
    srcNode = globalRanks[rank]/ppn;
    srcLink = srcNode;
    path = (int *)calloc (50, sizeof(int));
    
    if ( srcNode != destNode && routes.find(srcNode) == routes.end() ) {
      interRanks = topology.RouteBetweenRanks (globalRanks[rank], globalRanks[aggrCommRank], path);
      routeCost[srcNode] = 0;
      
      for ( i = 1; i < interRanks; i++ ) {
	interNode = path[i] / ppn;
	routes[srcNode].push_back (interNode);
	links[srcLink][interNode] = 0;
	srcLink = interNode;
      }
    }
    free (path);
  }

  /* I/O Node */
  path = (int *)calloc (50, sizeof(int));
  interRanks = topology.RouteToIONode (this->worldRank_, path );
  srcNode = this->worldRank_ / ppn;
  srcLink = srcNode;

  for ( i = 1; i < interRanks; i++ ) {
    interNode = path[i] / ppn;
    routes[srcNode].push_back (interNode);
    links[srcLink][interNode] = 0;
    srcLink = interNode;
  }
  free (path);

 
  for ( rank = 0; rank < aggrCommSize; rank++ ) {
    srcNode = globalRanks[rank]/ppn;
    srcLink = srcNode;
    
    if ( srcNode != destNode ) {
      for ( node = 0; node < routes[srcNode].size (); node++ ) {
	links[srcLink][routes[srcNode][node]]++;
	routeCost[srcNode] = std::max(routeCost[srcNode], links[srcLink][routes[srcNode][node]]);
	srcLink = routes[srcNode][node];
      }
    }
    else
      routeCost[srcNode] = 1;
  }

  
  for ( rank = 0; rank < aggrCommSize; rank++ ) {
    aggregatedData += dataDistrib[rank];
    srcNode = globalRanks[rank]/ppn;
    
    if ( rank != aggrCommRank ) {
      
      aggrCost.cost = std::max ( (double)dataDistrib[rank] / ( BANDWIDTH / routeCost[srcNode] ),
				 aggrCost.cost );
    }
  }

  /* I/O Node */
  srcNode = this->worldRank_ / ppn;
  aggrCost.cost += aggregatedData / ( BANDWIDTH / routeCost[srcNode] );
 
  
  if ( this->excludedNode[this->intCoords_] )
    aggrCost.cost = DBL_MAX;
  
  MPI_Allreduce ( &aggrCost, &minCost, 1, MPI_DOUBLE_INT, MPI_MINLOC, aggrComm );

  MPI_Reduce ( &dataSize, &this->aggrDataSize_, 1, MPI_LONG_LONG, MPI_SUM, minCost.rank, aggrComm );

  if ( minCost.rank == aggrCommRank ) {
    aggrRank = this->commRank_;
    this->amAnAggr_ = true;
  }

  MPI_Bcast ( &aggrRank, 1, MPI_INT, minCost.rank, aggrComm );

#ifdef DEBUG
  if ( minCost.rank == aggrCommRank )
    fprintf (stdout, "[DEBUG] Aggr. rank %d in aggrComm, distance to I/O node %d hops, cost: %.4f\n",
	     minCost.rank, topology.DistanceToIONode ( this->worldRank_ ), minCost.cost );
#endif

  return aggrRank;
}


Francois Tessier's avatar
Francois Tessier committed
718
int Tapioca::CoordsToInt (int *coords, int dim)
Francois Tessier's avatar
Francois Tessier committed
719 720 721 722 723 724 725 726 727 728 729 730
{
  int i, res = 0;

  for ( i = 0; i < dim; i++ )
    res += coords[i] * pow (10.0, (double)i);

  return res;
}

/***********************/
/*        MISC.        */
/***********************/
Francois Tessier's avatar
Francois Tessier committed
731
const char* Tapioca::getStrategyName ()
Francois Tessier's avatar
Francois Tessier committed
732 733 734 735 736 737 738 739 740 741 742 743 744
{
  switch (this->strategy_)
    {
    case SHORTEST_PATH:    return "Shortest path";
    case LONGEST_PATH:     return "Longest path";
    case TOPOLOGY_AWARE:   return "Topology-aware placement";
    case CONTENTION_AWARE: return "Contention-aware placement";
    case UNIFORM:          return "Uniform placement";
    default:               return "No placement strategy defined!";
    }
}


Francois Tessier's avatar
Francois Tessier committed
745
void Tapioca::HandleMPIError (int retval) 
Francois Tessier's avatar
Francois Tessier committed
746 747 748 749 750 751
{
#ifdef DEBUG
  char msg[MPI_MAX_ERROR_STRING];
  int resultlen;
  if (retval != MPI_SUCCESS) {
    MPI_Error_string(retval, msg, &resultlen);
752
    fprintf(stderr, "[ERR] %s\n", msg);
Francois Tessier's avatar
Francois Tessier committed
753 754 755 756 757 758 759
  }
#endif

  return;
}


Francois Tessier's avatar
Francois Tessier committed
760
void Tapioca::PrintTime ( double startTime, double endTime, char* func )
Francois Tessier's avatar
Francois Tessier committed
761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780
{
  double totTime, avgTime, minTime, maxTime;
  int commSize, commRank;
  
  MPI_Comm_rank(MPI_COMM_WORLD, &commRank);
  MPI_Comm_size(MPI_COMM_WORLD, &commSize);
  
  totTime = endTime - startTime;
  totTime = totTime * 1000;
  MPI_Reduce(&totTime, &avgTime, 1, MPI_DOUBLE, MPI_SUM, 0, MPI_COMM_WORLD);
  MPI_Reduce(&totTime, &maxTime, 1, MPI_DOUBLE, MPI_MAX, 0, MPI_COMM_WORLD);
  MPI_Reduce(&totTime, &minTime, 1, MPI_DOUBLE, MPI_MIN, 0, MPI_COMM_WORLD);
  if (commRank == 0) {
    avgTime = avgTime / commSize;
    fprintf (stdout, "[TIMING][AGG] %s: %.2f ms [%.2f ; %.2f]\n",
             func, avgTime, minTime, maxTime);
  }
}


Francois Tessier's avatar
Francois Tessier committed
781
void Tapioca::MPIIOInfo ( MPI_File fileHandle )
Francois Tessier's avatar
Francois Tessier committed
782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807
{
  MPI_Info info;
  int flag;
  char value[1024];

  if ( this->worldRank_ == 0 ) {
    MPI_File_get_info ( fileHandle, &info );
    
    fprintf ( stdout, "[INFO] MPI Two-phases I/O\n");
    
    MPI_Info_get ( info, "cb_buffer_size", 1024, value, &flag );
    fprintf ( stdout, "[INFO] cb_buffer_size = %s\n", value );
    
    MPI_Info_get ( info, "cb_nodes", 1024, value, &flag );
    fprintf ( stdout, "[INFO] cb_nodes = %s\n", value );

    MPI_Info_get ( info, "romio_cb_read", 1024, value, &flag );
    fprintf ( stdout, "[INFO] romio_cb_read = %s\n", value );
    
    MPI_Info_get ( info, "romio_cb_write", 1024, value, &flag );
    fprintf ( stdout, "[INFO] romio_cb_write = %s\n", value );
    
    MPI_Info_get ( info, "romio_no_indep_rw", 1024, value, &flag );
    fprintf ( stdout, "[INFO] romio_no_indep_rw = %s\n", value );
  }
}