RefineTorusLB.C

Go to the documentation of this file.

/*****************************************************************************
 * $Source: /home/cvs/namd/cvsroot/namd2/src/RefineTorusLB.C,v $
 * $Author: jim $
 * $Date: 2011/04/28 20:47:16 $
 * $Revision: 1.36 $
 *****************************************************************************/

#include "RefineTorusLB.h"
#define EXPAND_INNER_BRICK 2

RefineTorusLB::RefineTorusLB(computeInfo *cs, patchInfo *pas, processorInfo *pes, int ncs, 
int npas, int npes, int flag) : Rebalancer(cs, pas, pes, ncs, npas, npes)
{
  if(flag==1) {
    strategyName = "RefineTorusLB";
    strategy();
    // CREATE THE SPANNING TREE IN THE LOAD BALANCER
#if 0
    if(proxySendSpanning || proxyRecvSpanning) {
    for(int i=0; i<4; i++) {
      decrSTLoad();
      computeAverage();
      createSpanningTree();
      incrSTLoad();
      // for(int i=0; i<P; i++)
      //   delete [] processors[i].proxyUsage;
      InitProxyUsage();
      binaryRefine();
      printLoads();
      createSpanningTree();
    }
    }
#endif
  }
}

RefineTorusLB::~RefineTorusLB() { }

void RefineTorusLB::strategy() {
  int index, realPe;

  const int beginGroup = processors[0].Id;
  const int endGroup = beginGroup + P;
#define INGROUP(PROC) ((PROC) >= beginGroup && (PROC) < endGroup)

  firstAssignInRefine = 0;
  for(int i=0; i<numComputes; i++){
    // HYBRID check if processor is in local group
    realPe = computes[i].oldProcessor;
    if INGROUP(realPe) {
      index = realPe - processors[0].Id;
      assign((computeInfo *) &(computes[i]), (processorInfo *) &(processors[index]));
    }
  }
  firstAssignInRefine = 1;

  printLoads();
  binaryRefine();
  printLoads();
}

void RefineTorusLB::binaryRefine() {
  // compute the max and average load
  computeAverage();
  double max = computeMax();

  double step = 0.01, start = 1.01 + ((double)P)/((double)numComputes);
  double dCurLoad = max/averageLoad;
  int curLoad;
  int minLoad = 0;
  int maxLoad = (int)((dCurLoad - start)/step + 1);
  double dMinLoad = minLoad * step + start;
  double dMaxLoad = maxLoad * step + start;

  // check the two limits of the search: start and dMaxLoad
  int done=0;
  overLoad = dMinLoad;
  if(newRefine())
    done = 1;
  else {
    overLoad = dMaxLoad;
    if(!newRefine()) {
      CkPrintf("Error: Could not refine at max overload\n");
      done = 1;
    }
  } 

  // do a binary search between start and dMaxLoad until we succeed
  while(!done) {
    if(maxLoad - minLoad <= 1)
      done = 1;
    else {
      curLoad = (maxLoad + minLoad)/2;
      overLoad = curLoad * step + start;
      if(newRefine())
        maxLoad = curLoad;
      else
        minLoad = curLoad;
    }
  }
}

int RefineTorusLB::newRefine() {
  int done = 1;
  maxHeap *heavyPes = new maxHeap(P);
  IRSet *lightPes = new IRSet();
  processorInfo *donor, *p, *bestP;
  computeInfo *c;
  Iterator nextC, nextP;
  pcpair good;
  double thresholdLoad = overLoad * averageLoad;
  int index, realPe;

  const int beginGroup = processors[0].Id;
  const int endGroup = beginGroup + P;

  // create a heap and set of heavy and light pes respectively
  for(int i=0; i<P; i++) {
    if (processors[i].load > thresholdLoad)
      heavyPes->insert((InfoRecord *) &(processors[i]));
    else
      lightPes->insert((InfoRecord *) &(processors[i]));
  }

#if LDB_DEBUG
  iout << "\n Before Refinement Summary\n" << endi;
  printSummary();
#endif

  pcpair pcpairarray[12];
     
  for(int j=0; j<6; j++) {
    bestPe[j] = &pcpairarray[j];    // new pcpair();
    goodPe[j] = &pcpairarray[j+6];  // new pcpair();
  }

  while(1) {
    while(donor = (processorInfo*)heavyPes->deleteMax())
      if(donor->computeSet.hasElements())
        break;
    
    if(!donor) break;

    for(int j=0; j<6; j++) {
      bestPe[j]->reset();
      goodPe[j]->reset();
    }

    nextC.id = 0;
    c = (computeInfo *)donor->computeSet.iterator((Iterator *)&nextC);

    while(c) {
      // Look at pes which have the compute's patches

      // HYBRID check if processor is in local group
#define SELECT_REALPE(X) if INGROUP((X)) { \
        selectPes(&processors[(X) - beginGroup], c); \
      }

      int realPe1 = patches[c->patch1].processor;
      SELECT_REALPE(realPe1)

      int realPe2 = patches[c->patch2].processor;
      if ( realPe2 != realPe1 ) {
        SELECT_REALPE(realPe2)
      }

      // Try the processors which have the patches' proxies
      p = (processorInfo *)(patches[c->patch1].proxiesOn.iterator((Iterator *)&nextP));
      while(p) {                                            // patch 1
        if INGROUP(p->Id) selectPes(p, c);
        p = (processorInfo *)(patches[c->patch1].proxiesOn.next((Iterator *)&nextP));
      }
  
      p = (processorInfo *)(patches[c->patch2].proxiesOn.iterator((Iterator *)&nextP));
      while(p) {                                            //patch 2
        if INGROUP(p->Id) selectPes(p, c);
        p = (processorInfo *)(patches[c->patch2].proxiesOn.next((Iterator *)&nextP));
      }
      
      nextC.id++;
      c = (computeInfo *) donor->computeSet.next((Iterator *)&nextC);
    } // end of compute loop

#define REASSIGN(GRID) if (GRID->c) { deAssign(GRID->c, donor); \
        assign(GRID->c, GRID->p); bestP = GRID->p; }

    bestP = 0;
    // see if we have found a compute processor pair
    REASSIGN(bestPe[5])
#if USE_TOPOMAP
    else REASSIGN(goodPe[5])
#endif
    else REASSIGN(bestPe[4])
#if USE_TOPOMAP
    else REASSIGN(goodPe[4])
#endif
    else REASSIGN(bestPe[3])
#if USE_TOPOMAP
    else REASSIGN(goodPe[3])
#endif
    else REASSIGN(bestPe[1])
#if USE_TOPOMAP
    else REASSIGN(goodPe[1])
#endif
    else REASSIGN(bestPe[2])
#if USE_TOPOMAP
    else REASSIGN(goodPe[2])
#endif
    else REASSIGN(bestPe[0])
#if USE_TOPOMAP
    else REASSIGN(goodPe[0])
#endif

  // Try all pes on the nodes of the home patches
    if ( ! bestP && CmiNumNodes() > 1 ) {  // else not useful
      double minLoad = overLoad * averageLoad;
      good.c = 0; good.p = 0;
      nextC.id = 0;
      c = (computeInfo *)donor->computeSet.iterator((Iterator *)&nextC);
      while(c) {
        int realPe1 = patches[c->patch1].processor;
        int realNode1 = CmiNodeOf(realPe1);
        int nodeSize = CmiNodeSize(realNode1);
        if ( nodeSize > 1 ) {  // else did it already
          int firstpe = CmiNodeFirst(realNode1);
          for ( int rpe = firstpe; rpe < firstpe+nodeSize; ++rpe ) {
            if INGROUP(rpe) {
              p = &processors[rpe - beginGroup];
              if ( p->available && ( p->load + c->load < minLoad ) ) {
                minLoad = p->load + c->load;
                good.c = c;
                good.p = p;
              }
            }
          }
        }
        int realPe2 = patches[c->patch2].processor;
        if ( realPe2 != realPe1 ) {
          int realNode2 = CmiNodeOf(realPe2);
          if ( realNode2 != realNode1 ) {  // else did it already
            nodeSize = CmiNodeSize(realNode2);
            if ( nodeSize > 1 ) {
              int firstpe = CmiNodeFirst(realNode2);
              for ( int rpe = firstpe; rpe < firstpe+nodeSize; ++rpe ) {
                if INGROUP(rpe) {
                  p = &processors[rpe - beginGroup];
                  if ( p->available && ( p->load + c->load < minLoad ) ) {
                    minLoad = p->load + c->load;
                    good.c = c;
                    good.p = p;
                  }
                }
              }
            }
          }
        }
        nextC.id++;
        c = (computeInfo *) donor->computeSet.next((Iterator *)&nextC);
      } // end of compute loop

      REASSIGN((&good))
    }

  // Try all pes on the physical nodes of the home patches
    if ( ! bestP && ( CmiNumPhysicalNodes() > 1 ) &&
         ( CmiNumPhysicalNodes() < CmiNumNodes() ) ) {  // else not useful
      double minLoad = overLoad * averageLoad;
      good.c = 0; good.p = 0;
      nextC.id = 0;
      c = (computeInfo *)donor->computeSet.iterator((Iterator *)&nextC);
      while(c) {
        int realPe1 = patches[c->patch1].processor;
        int realNode1 = CmiPhysicalNodeID(realPe1);
        int *rpelist;
        int nodeSize;
        CmiGetPesOnPhysicalNode(realNode1, &rpelist, &nodeSize);
        if ( nodeSize > 1 ) {  // else did it already
          for ( int ipe = 0; ipe < nodeSize; ++ipe ) {
            int rpe = rpelist[ipe];
            if INGROUP(rpe) {
              p = &processors[rpe - beginGroup];
              if ( p->available && ( p->load + c->load < minLoad ) ) {
                minLoad = p->load + c->load;
                good.c = c;
                good.p = p;
              }
            }
          }
        }
        int realPe2 = patches[c->patch2].processor;
        if ( realPe2 != realPe1 ) {
          int realNode2 = CmiPhysicalNodeID(realPe2);
          if ( realNode2 != realNode1 ) {  // else did it already
            CmiGetPesOnPhysicalNode(realNode2, &rpelist, &nodeSize);
            if ( nodeSize > 1 ) {  // else did it already
              for ( int ipe = 0; ipe < nodeSize; ++ipe ) {
                int rpe = rpelist[ipe];
                if INGROUP(rpe) {
                  p = &processors[rpe - beginGroup];
                  if ( p->available && ( p->load + c->load < minLoad ) ) {
                    minLoad = p->load + c->load;
                    good.c = c;
                    good.p = p;
                  }
                }
              }
            }
          }
        }
        nextC.id++;
        c = (computeInfo *) donor->computeSet.next((Iterator *)&nextC);
      } // end of compute loop

      REASSIGN((&good))
    }

    if(bestP) {
      if(bestP->load > averageLoad) {
        // CkPrintf("Acceptor %d became heavy%f %f\n", bestP->Id, bestP->load, overLoad*averageLoad);
        lightPes->remove(bestP);
      } else {
        // CkPrintf("Acceptor %d still light %f %f\n", bestP->Id, bestP->load, overLoad*averageLoad);
      }
      if(donor->load > overLoad*averageLoad) {
        // CkPrintf("Donor %d still heavy %f %f\n", donor->Id, donor->load, overLoad*averageLoad);
        heavyPes->insert((InfoRecord *) donor);
      }
      else {
        // CkPrintf("Donor %d became light %f %f\n", donor->Id, donor->load, overLoad*averageLoad);
        lightPes->insert((InfoRecord *) donor);
      }
      
      continue;
    }
    //else
      //CkPrintf("1st try failed\n");

    int found = 0;
#if USE_TOPOMAP
    // if this fails, look at the inner brick
    int p1, p2, pe, x1, x2, xm, xM, y1, y2, ym, yM, z1, z2, zm, zM, t1, t2;
    int dimNX, dimNY, dimNZ, dimNT;
    double minLoad;

    good.c = 0; good.p = 0;
    minLoad = overLoad*averageLoad;
    nextC.id = 0;
    c = (computeInfo *)donor->computeSet.iterator((Iterator *)&nextC);
 
    while(c) {    
      p1 = patches[c->patch1].processor;
      p2 = patches[c->patch2].processor;

      tmgr.rankToCoordinates(p1, x1, y1, z1, t1);
      tmgr.rankToCoordinates(p2, x2, y2, z2, t2);
      dimNX = tmgr.getDimNX();
      dimNY = tmgr.getDimNY();
      dimNZ = tmgr.getDimNZ();
      dimNT = tmgr.getDimNT();

      brickDim(x1, x2, dimNX, xm, xM);
      brickDim(y1, y2, dimNY, ym, yM);
      brickDim(z1, z2, dimNZ, zm, zM);

      // to expand the inner brick by some hops
#if 0
      if(xm>=EXPAND_INNER_BRICK) xm=xm-EXPAND_INNER_BRICK; else xm=0;
      if(ym>=EXPAND_INNER_BRICK) ym=ym-EXPAND_INNER_BRICK; else ym=0;
      if(zm>=EXPAND_INNER_BRICK) zm=zm-EXPAND_INNER_BRICK; else zm=0;

      xM=xM+EXPAND_INNER_BRICK;
      yM=yM+EXPAND_INNER_BRICK;
      zM=zM+EXPAND_INNER_BRICK;
#endif

      // first go over the processors inside the brick and choose the least 
      for(int i=xm; i<=xM; i++)
        for(int j=ym; j<=yM; j++)
          for(int k=zm; k<=zM; k++)
            for(int l=0; l<dimNT; l++)
            {
              pe = tmgr.coordinatesToRank(i%dimNX, j%dimNY, k%dimNZ, l);
              if ( ! INGROUP(pe) ) continue;
              p = &processors[pe - beginGroup];
              if(c->load + p->load < minLoad) {
                minLoad = c->load + p->load;
                good.c = c;
                good.p = p;
              }
            }
      nextC.id++;
      c = (computeInfo *) donor->computeSet.next((Iterator *)&nextC);
    }

    if(good.c) {
      found = 1;
      //CkPrintf("2nd try succeeded\n");
    }
    else {
      found = 0;
      //CkPrintf("2nd try failed\n");
    }

    // if that also fails, look at the outer brick
    minLoad = overLoad * averageLoad;
    if(found==0) {
      good.c = 0; good.p = 0;
      p = 0;

      nextC.id = 0;
      c = (computeInfo *)donor->computeSet.iterator((Iterator *)&nextC);
      while(c) {
        p1 = patches[c->patch1].processor;
        p2 = patches[c->patch2].processor;

        tmgr.rankToCoordinates(p1, x1, y1, z1, t1);
        tmgr.rankToCoordinates(p2, x2, y2, z2, t2);
        dimNX = tmgr.getDimNX();
        dimNY = tmgr.getDimNY();
        dimNZ = tmgr.getDimNZ();
        dimNT = tmgr.getDimNT();

        brickDim(x1, x2, dimNX, xm, xM);
        brickDim(y1, y2, dimNY, ym, yM);
        brickDim(z1, z2, dimNZ, zm, zM);

        for(int i=xM+1; i<xm+dimNX; i++)
          for(int j=0; j<dimNY; j++)
            for(int k=0; k<dimNZ; k++)
              for(int l=0; l<dimNT; l++)
              {
                pe = tmgr.coordinatesToRank(i%dimNX, j%dimNY, k%dimNZ, l);
                if ( ! INGROUP(pe) ) continue;
                p = &processors[pe - beginGroup];
                if(c->load + p->load < minLoad) {
                  good.c = c;
                  good.p = p;
                  found = 1; break;
                }
              }

        if(found==1)
          break;
        else {
          for(int j=yM+1; j<ym+dimNY; j++)
            for(int i=xm; i<=xM; i++)
              for(int k=0; k<dimNZ; k++)
                for(int l=0; l<dimNT; l++)
                {
                  pe = tmgr.coordinatesToRank(i%dimNX, j%dimNY, k%dimNZ, l);
                  if ( ! INGROUP(pe) ) continue;
                  p = &processors[pe - beginGroup];
                  if(c->load + p->load < minLoad) {
                    good.c = c;
                    good.p = p;
                    found = 1; break;
                  }
                }
        }
 
        if(found==1)
          break;
        else {
          for(int k=zM+1; k<zm+dimNZ; k++)
            for(int i=xm; i<=xM; i++)
              for(int j=ym; j<=yM; j++)
                for(int l=0; l<dimNT; l++)
                {
                  pe = tmgr.coordinatesToRank(i%dimNX, j%dimNY, k%dimNZ, l);
                  if ( ! INGROUP(pe) ) continue;
                  p = &processors[pe - beginGroup];
                  if(c->load + p->load < minLoad) {
                    good.c = c;
                    good.p = p;
                    found = 1; break;
                  }
                }
        }

        if(found==1) break;

        nextC.id++;
        c = (computeInfo *) donor->computeSet.next((Iterator *)&nextC);
      } 
    }

    if(found == 1) {
      deAssign(good.c, donor);
      assign(good.c, good.p);
      if (good.p->load > averageLoad) lightPes->remove(good.p);
      if (donor->load > overLoad*averageLoad)
        heavyPes->insert((InfoRecord *) donor);
      else
        lightPes->insert((InfoRecord *) donor);
      continue;
    }

#endif /* USE_TOPOMAP */

    // find the first processor to place the compute on
    p = (processorInfo *)lightPes->iterator((Iterator *) &nextP);
    if(found == 0) {
     while (p)
      {
        nextC.id = 0;
        c = (computeInfo *)donor->computeSet.iterator((Iterator *)&nextC);
        while (c)
        {
          selectPes(p, c);
          nextC.id++;
          c = (computeInfo *) donor->computeSet.next((Iterator *)&nextC);
        }
        p = (processorInfo *)lightPes->next((Iterator *) &nextP);
      }

      bestP = 0;
      REASSIGN(bestPe[5])
#if USE_TOPOMAP
      else REASSIGN(goodPe[5])
#endif
      else REASSIGN(bestPe[4])
#if USE_TOPOMAP
      else REASSIGN(goodPe[4])
#endif
      else REASSIGN(bestPe[3])
#if USE_TOPOMAP
      else REASSIGN(goodPe[3])
#endif
      else REASSIGN(bestPe[1])
#if USE_TOPOMAP
      else REASSIGN(goodPe[1])
#endif
      else REASSIGN(bestPe[2])
#if USE_TOPOMAP
      else REASSIGN(goodPe[2])
#endif
      else REASSIGN(bestPe[0])
#if USE_TOPOMAP
      else REASSIGN(goodPe[0])
#endif
    }

    if(bestP) {
      if(bestP->load > averageLoad) lightPes->remove(bestP);
      if(donor->load > overLoad*averageLoad)
        heavyPes->insert((InfoRecord *) donor);
      else
        lightPes->insert((InfoRecord *) donor);
      continue;
    }
    else {
      done = 0;
      break;
    }
 
  } // end of while loop

#if LDB_DEBUG
  iout << "After Refinement Summary\n" << endi;
  printSummary();
#endif

  delete heavyPes;
  delete lightPes;

  return done;
}

void RefineTorusLB::selectPes(processorInfo *p, computeInfo *c) {
  if (p->available == CmiFalse)
    return;

  // find the position in bestPe/goodPe to place this pair
  // HP HP HP HP HP HP
  // 02 11 20 01 10 00
  //  5  4  3  2  1  0
  int numPatches, numProxies, /* badForComm, */ index;
  numAvailable(c, p, &numPatches, &numProxies, 0 /* &badForComm */); 
  int numEither = numPatches + numProxies;
  index = (numEither*(numEither+1))/2 + numProxies;

#if USE_TOPOMAP
  int x, y, z, t;
  int p1, p2, pe, x1, x2, xm, xM, y1, y2, ym, yM, z1, z2, zm, zM, t1, t2;
  int dimNX, dimNY, dimNZ, dimNT;
  double minLoad;
  p1 = patches[c->patch1].processor;
  p2 = patches[c->patch2].processor;

  tmgr.rankToCoordinates(p1, x1, y1, z1, t1);
  tmgr.rankToCoordinates(p2, x2, y2, z2, t2);
  dimNX = tmgr.getDimNX();
  dimNY = tmgr.getDimNY();
  dimNZ = tmgr.getDimNZ();
  dimNT = tmgr.getDimNT();

  brickDim(x1, x2, dimNX, xm, xM);
  brickDim(y1, y2, dimNY, ym, yM);
  brickDim(z1, z2, dimNZ, zm, zM);
#endif

  if (p->load + c->load < overLoad * averageLoad) {
#if USE_TOPOMAP
    tmgr.rankToCoordinates(p->Id, x, y, z, t);
    int wB = withinBrick(x, y, z, xm, xM, dimNX, ym, yM, dimNY, zm, zM, dimNZ);
    if (wB) {
#endif
      pcpair* &oldp = bestPe[index];

      if (!(oldp->p) || ((p->load + c->load) < (oldp->p->load + oldp->c->load))) {
        oldp->p = p;
        oldp->c = c;
      } 
#if USE_TOPOMAP
    } else {
      pcpair* &oldp = goodPe[index];
      double loadDiff = 0.0;

      if (!(oldp->p)) {
        oldp->p = p;
        oldp->c = c;
      } else {
        loadDiff = oldp->p->load + oldp->c->load - p->load - c->load;
        if ( (loadDiff > 0.4) || (loadDiff > 0.0 && (tmgr.getHopsBetweenRanks(p->Id, p1) + tmgr.getHopsBetweenRanks(p->Id, p2) < tmgr.getHopsBetweenRanks((oldp->p)->Id, p1) + tmgr.getHopsBetweenRanks((oldp->p)->Id, p2))) ) {
          oldp->p = p;
          oldp->c = c;
        }
      }
    }
#endif
  }
}

---