RefineTorusLB Class Reference

#include <RefineTorusLB.h>

Inheritance diagram for RefineTorusLB:

List of all members.

Public Member Functions
	RefineTorusLB (computeInfo *cs, patchInfo *pas, processorInfo *pes, int ncs, int npas, int npes, int flag)
	~RefineTorusLB ()
void	binaryRefine ()
int	newRefine ()

---

Constructor & Destructor Documentation

RefineTorusLB::RefineTorusLB (  computeInfo *  cs, patchInfo *  pas, processorInfo *  pes, int  ncs, int  npas, int  npes, int  flag )

Definition at line 18 of file RefineTorusLB.C. References binaryRefine(), Rebalancer::computeAverage(), Rebalancer::createSpanningTree(), Rebalancer::decrSTLoad(), Rebalancer::incrSTLoad(), Rebalancer::InitProxyUsage(), Rebalancer::printLoads(), proxySendSpanning, and Rebalancer::strategy(). : 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 (   )

Definition at line 44 of file RefineTorusLB.C. { }

---

Member Function Documentation

void RefineTorusLB::binaryRefine (   )

Definition at line 56 of file RefineTorusLB.C. References Rebalancer::computeAverage(), Rebalancer::computeMax(), and newRefine(). Referenced by RefineTorusLB(), and TorusLB::TorusLB(). { // compute the max and average load computeAverage(); double max = computeMax(); double step = 0.01, start = 1.05; 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 (   )

Definition at line 97 of file RefineTorusLB.C. References Rebalancer::assign(), Rebalancer::brickDim(), processorInfo::computeSet, Rebalancer::deAssign(), maxHeap::deleteMax(), Iterator::id, Set::insert(), maxHeap::insert(), Set::iterator(), InfoRecord::load, Set::next(), Set::numElements(), computeInfo::patch1, computeInfo::patch2, patchInfo::processor, REASSIGN, and Set::remove(). Referenced by binaryRefine(). { int done = 1; maxHeap *heavyPes = new maxHeap(P); Set *lightPes = new Set(); processorInfo *donor, *p, *bestP; computeInfo *c; Iterator nextC, nextP; pcpair good; double thresholdLoad = overLoad * averageLoad; // 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])); } 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.numElements()) 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 the processors which have the compute's patches first p = &processors[patches[c->patch1].processor]; // patch 1 selectPes(p, c); p = &processors[patches[c->patch2].processor]; // patch 2 selectPes(p, c); // Try the processors which have the patches' proxies p = (processorInfo *)(patches[c->patch1].proxiesOn.iterator((Iterator *)&nextP)); while(p) { // patch 1 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 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 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); p = &processors[pe]; 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); p = &processors[pe]; 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); p = &processors[pe]; 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); p = &processors[pe]; 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 delete heavyPes; delete lightPes; return done; }

---

The documentation for this class was generated from the following files:

• RefineTorusLB.h
• RefineTorusLB.C

---