void RecBipartLB::work(LDStats *stats) {
vector<Vertex *> ptrvector;
/** ========================== INITIALIZATION ============================= */
ProcArray *parr = new ProcArray(stats); // Processor Array
ObjGraph *ogr = new ObjGraph(stats); // Object Graph
/** ============================= STRATEGY ================================ */
level=0;
peno=0;
TOTALLOAD=0;
numparts=CkNumPes();
parray=parr;
double avgLoad = parr->getAverageLoad();
int numPes = parr->procs.size();
parr->resetTotalLoad();
for(int i=0;i<ogr->vertices.size();i++)
{
Vertex_helper *helper = new Vertex_helper();
vhelpers.push_back(helper);
ptrvector.push_back((Vertex *)&(ogr->vertices[i]));
}
RecursiveBiPart(ogr,ptrvector,1,numparts);
/** ============================== CLEANUP ================================ */
ogr->convertDecisions(stats); // Send decisions back to LDStats
}
void GraphBFTLB::work(LDStats *stats) {
/** ========================== INITIALIZATION ============================= */
ProcArray *parr = new ProcArray(stats); // Processor Array
ObjGraph *ogr = new ObjGraph(stats); // Object Graph
/** ============================= STRATEGY ================================ */
double avgLoad = parr->getAverageLoad();
int numPes = parr->procs.size();
// CkPrintf("Average Load %g\n\n", avgLoad);
// for(int i=0; i<numPes; i++)
// CkPrintf("PE [%d] %g %g\n", i, parr->procs[i].getTotalLoad(), parr->procs[i].getOverhead());
parr->resetTotalLoad();
int start = 0, nextPe = 0;
std::queue<int> vertexq;
// start at vertex with id 0
vertexq.push(start);
if(parr->procs[nextPe].getTotalLoad() + ogr->vertices[start].getVertexLoad() > avgLoad) {
nextPe++;
avgLoad += (avgLoad - parr->procs[nextPe].getTotalLoad())/(numPes-nextPe);
}
ogr->vertices[start].setNewPe(nextPe);
// CkPrintf("[%d] %d %d %g %g %g\n", start, ogr->vertices[start].getCurrentPe(), ogr->vertices[start].getNewPe(), parr->procs[nextPe].getTotalLoad(), ogr->vertices[start].getVertexLoad(), parr->procs[nextPe].getTotalLoad() + ogr->vertices[start].getVertexLoad());
parr->procs[nextPe].totalLoad() += ogr->vertices[start].getVertexLoad();
int i, nbr;
// breadth first traversal
while(!vertexq.empty()) {
start = vertexq.front();
vertexq.pop();
for(i = 0; i < ogr->vertices[start].sendToList.size(); i++) {
// look at all neighbors of a node in the queue and map them while
// inserting them in the queue (so we can look at their neighbors next)
nbr = ogr->vertices[start].sendToList[i].getNeighborId();
if(ogr->vertices[nbr].getNewPe() == -1) {
vertexq.push(nbr);
if(parr->procs[nextPe].getTotalLoad() + ogr->vertices[nbr].getVertexLoad() > avgLoad) {
nextPe++;
avgLoad += (avgLoad - parr->procs[nextPe].getTotalLoad())/(numPes-nextPe);
}
ogr->vertices[nbr].setNewPe(nextPe);
// CkPrintf("[%d] %d %d %g %g %g\n", nbr, ogr->vertices[nbr].getCurrentPe(), ogr->vertices[nbr].getNewPe(), parr->procs[nextPe].getTotalLoad(), ogr->vertices[start].getVertexLoad(), parr->procs[nextPe].getTotalLoad() + ogr->vertices[start].getVertexLoad());
parr->procs[nextPe].totalLoad() += ogr->vertices[nbr].getVertexLoad();
}
} // end of for loop
} // end of while loop
/** ============================== CLEANUP ================================ */
ogr->convertDecisions(stats); // Send decisions back to LDStats
}
void TempAwareGreedyLB::work(LDStats* stats)
{
CkPrintf("----------------- in TempAwareGreedyLB -----------\n");
/** ========================== INITIALIZATION ============================= */
ProcArray *parr = new ProcArray(stats); // Processor Array
ObjGraph *ogr = new ObjGraph(stats); // Object Graph
/** ============================= STRATEGY ================================ */
parr->resetTotalLoad();
if (_lb_args.debug()>1)
CkPrintf("[%d] In TempAwareGreedyLB strategy\n",CkMyPe());
int vert;
// max heap of objects
std::sort(ogr->vertices.begin(), ogr->vertices.end(), ObjLoadGreater());
// min heap of processors
std::make_heap(parr->procs.begin(), parr->procs.end(), ProcLoadGreater());
for(vert = 0; vert < ogr->vertices.size(); vert++) {
// Pop the least loaded processor
ProcInfo p = parr->procs.front();
std::pop_heap(parr->procs.begin(), parr->procs.end(), ProcLoadGreater());
parr->procs.pop_back();
// Increment the load of the least loaded processor by the load of the
// 'heaviest' unmapped object
p.setTotalLoad(p.getTotalLoad() + ogr->vertices[vert].getVertexLoad());
ogr->vertices[vert].setNewPe(p.getProcId());
// Insert the least loaded processor with load updated back into the heap
parr->procs.push_back(p);
std::push_heap(parr->procs.begin(), parr->procs.end(), ProcLoadGreater());
}
/** ============================== CLEANUP ================================ */
ogr->convertDecisions(stats); // Send decisions back to LDStats
}
void ScotchRefineLB::work(LDStats *stats) {
/** ========================== INITIALIZATION ============================= */
ProcArray *parr = new ProcArray(stats);
ObjGraph *ogr = new ObjGraph(stats);
int cost_array[10] = {64, 256, 512, 1024, 2048, 4096, 8192, 16384, 32768, 65536};
/** ============================= STRATEGY ================================ */
// convert ObjGraph to the Scotch graph
SCOTCH_Num baseval = 0; // starting index of vertices
SCOTCH_Num vertnbr = ogr->vertices.size(); // number of vertices
SCOTCH_Num edgenbr = 0; // number of edges
SCOTCH_Num *oldpemap = (SCOTCH_Num *)malloc(sizeof(SCOTCH_Num) * vertnbr);
double maxLoad = 0.0;
double minLoad = 0.0;
if (vertnbr > 0) {
minLoad = ogr->vertices[baseval].getVertexLoad();
}
long maxBytes = 1;
int i, j, k, vert;
/** remove duplicate edges from recvFrom */
for(i = baseval; i < vertnbr; i++) {
for(j = 0; j < ogr->vertices[i].sendToList.size(); j++) {
vert = ogr->vertices[i].sendToList[j].getNeighborId();
for(k = 0; k < ogr->vertices[i].recvFromList.size(); k++) {
if(ogr->vertices[i].recvFromList[k].getNeighborId() == vert) {
ogr->vertices[i].sendToList[j].setNumBytes(ogr->vertices[i].sendToList[j].getNumBytes() + ogr->vertices[i].recvFromList[k].getNumBytes());
ogr->vertices[i].recvFromList.erase(ogr->vertices[i].recvFromList.begin() + k);
}
}
}
}
/** the object load is normalized to an integer between 0 and 256 */
for(i = baseval; i < vertnbr; i++) {
if(ogr->vertices[i].getVertexLoad() > maxLoad)
maxLoad = ogr->vertices[i].getVertexLoad();
if (ogr->vertices[i].getVertexLoad() < minLoad) {
minLoad = ogr->vertices[i].getVertexLoad();
}
edgenbr += ogr->vertices[i].sendToList.size() + ogr->vertices[i].recvFromList.size();
oldpemap[i] = ogr->vertices[i].getCurrentPe();
}
for(i = baseval; i < vertnbr; i++) {
for(j = 0; j < ogr->vertices[i].sendToList.size(); j++) {
if (ogr->vertices[i].sendToList[j].getNumBytes() > maxBytes) {
maxBytes = ogr->vertices[i].sendToList[j].getNumBytes();
}
}
for(j = 0; j < ogr->vertices[i].recvFromList.size(); j++) {
if (ogr->vertices[i].recvFromList[j].getNumBytes() > maxBytes) {
maxBytes = ogr->vertices[i].recvFromList[j].getNumBytes();
}
}
}
/* adjacency list */
SCOTCH_Num *verttab = (SCOTCH_Num *)malloc(sizeof(SCOTCH_Num) * (vertnbr+1));
/* loads of vertices */
SCOTCH_Num *velotab = (SCOTCH_Num *)malloc(sizeof(SCOTCH_Num) * vertnbr);
/* id of the neighbors */
SCOTCH_Num *edgetab = (SCOTCH_Num *)malloc(sizeof(SCOTCH_Num) * edgenbr);
/* number of bytes exchanged */
SCOTCH_Num *edlotab = (SCOTCH_Num *)malloc(sizeof(SCOTCH_Num) * edgenbr);
int edgeNum = 0;
double ratio = 256.0/maxLoad;
double byteRatio = 1024.0/maxBytes;
for(i = baseval; i < vertnbr; i++) {
verttab[i] = edgeNum;
velotab[i] = (int)ceil(ogr->vertices[i].getVertexLoad() * ratio);
for(j = 0; j < ogr->vertices[i].sendToList.size(); j++) {
edgetab[edgeNum] = ogr->vertices[i].sendToList[j].getNeighborId();
edlotab[edgeNum] = (int) ceil(ogr->vertices[i].sendToList[j].getNumBytes()
* byteRatio);
edgeNum++;
}
for(j = 0; j < ogr->vertices[i].recvFromList.size(); j++) {
edgetab[edgeNum] = ogr->vertices[i].recvFromList[j].getNeighborId();
edlotab[edgeNum] = (int)
ceil(ogr->vertices[i].recvFromList[j].getNumBytes() * byteRatio);
edgeNum++;
}
}
verttab[i] = edgeNum;
CkAssert(edgeNum == edgenbr);
SCOTCH_Graph graph; // Graph to partition
SCOTCH_Strat strat; // Strategy to achieve partitioning
/* Initialize data structures */
SCOTCH_graphInit (&graph);
SCOTCH_stratInit (&strat);
SCOTCH_graphBuild (&graph, baseval, vertnbr, verttab, NULL, velotab, NULL, edgenbr, edgetab, edlotab);
SCOTCH_graphCheck (&graph);
double migration_cost = 1024.0;
if (step() == 0) {
SCOTCH_stratGraphMapBuild (&strat, SCOTCH_STRATBALANCE, parr->procs.size (), 0.01);
} else {
SCOTCH_stratGraphMapBuild (&strat, SCOTCH_STRATBALANCE | SCOTCH_STRATREMAP, parr->procs.size (), 0.01);
}
SCOTCH_Num *pemap = (SCOTCH_Num *)malloc(sizeof(SCOTCH_Num) * vertnbr);
// Takes as input the graph, arch graph, strategy, migration cost in
// double, old mapping and new mapping
if (step() == 0) {
SCOTCH_graphPart(&graph, parr->procs.size(), &strat, pemap);
} else {
SCOTCH_graphRepart(&graph, parr->procs.size(), oldpemap, migration_cost, NULL, &strat, pemap);
}
SCOTCH_graphExit (&graph);
SCOTCH_stratExit (&strat);
free(verttab);
free(velotab);
free(edgetab);
free(edlotab);
for(i = baseval; i < vertnbr; i++) {
if(pemap[i] != ogr->vertices[i].getCurrentPe())
ogr->vertices[i].setNewPe(pemap[i]);
}
free(pemap);
free(oldpemap);
/** ============================== CLEANUP ================================ */
ogr->convertDecisions(stats);
delete parr;
delete ogr;
}
void RefineSwapLB::work(LDStats* stats)
{
/** ========================== INITIALIZATION ============================= */
ProcArray *parr = new ProcArray(stats); // Processor Array
ObjGraph *ogr = new ObjGraph(stats); // Object Graph
/** ============================= STRATEGY ================================ */
if (_lb_args.debug()>1)
CkPrintf("[%d] In RefineSwapLB strategy\n",CkMyPe());
int vert;
double avg_load = parr->getAverageLoad();
double threshold = avg_load * 0.01;
double lower_bound_load = avg_load - threshold;
double upper_bound_load = avg_load + threshold;
cout <<"Average load " << avg_load << endl;
std::vector<int> min_pe_heap;
std::vector<int> max_pe_heap;
std::vector<int>* pe_obj = new std::vector<int>[parr->procs.size()];
// Create a datastructure to store the objects in a processor
for (int i = 0; i < ogr->vertices.size(); i++) {
pe_obj[ogr->vertices[i].getCurrentPe()].push_back(i);
// CkPrintf("%d pe %d: %lf\n", i, ogr->vertices[i].getCurrentPe(), ogr->vertices[i].getVertexLoad());
}
// Construct max heap of overloaded processors and min heap of underloaded
// processors.
for (int i = 0; i < parr->procs.size(); i++) {
//CkPrintf("%d : %lf\n", i, parr->procs[i].getTotalLoad());
if (parr->procs[i].getTotalLoad() > upper_bound_load) {
max_pe_heap.push_back(i);
} else if (parr->procs[i].getTotalLoad() < lower_bound_load) {
min_pe_heap.push_back(i);
}
}
std::make_heap(max_pe_heap.begin(), max_pe_heap.end(), ProcLoadGreaterIndex(parr));
while (max_pe_heap.size() != 0 && min_pe_heap.size() != 0) {
int p_index = getMax(parr, max_pe_heap);
ProcInfo &pinfo = parr->procs[p_index];
bool success = refine(parr, ogr, max_pe_heap, min_pe_heap, pe_obj, p_index, avg_load, threshold);
if (!success) {
// Swap with something.
if (!refineSwap(parr, ogr, max_pe_heap, min_pe_heap, pe_obj, p_index, avg_load,
threshold)) {
max_pe_heap.push_back(p_index);
std::push_heap(max_pe_heap.begin(), max_pe_heap.end(),
ProcLoadGreaterIndex(parr));
break;
}
}
}
/** ============================== CLEANUP ================================ */
ogr->convertDecisions(stats); // Send decisions back to LDStats
delete[] pe_obj;
delete parr;
delete ogr;
}
void MetisLB::work(LDStats* stats)
{
/** ========================== INITIALIZATION ============================= */
ProcArray *parr = new ProcArray(stats);
ObjGraph *ogr = new ObjGraph(stats);
/** ============================= STRATEGY ================================ */
if (_lb_args.debug() >= 2) {
CkPrintf("[%d] In MetisLB Strategy...\n", CkMyPe());
}
// convert ObjGraph to the adjacency structure
int numVertices = ogr->vertices.size(); // number of vertices
int numEdges = 0; // number of edges
double maxLoad = 0.0;
int i, j, k, vert;
/** remove duplicate edges from recvFrom */
for(i = 0; i < numVertices; i++) {
for(j = 0; j < ogr->vertices[i].sendToList.size(); j++) {
vert = ogr->vertices[i].sendToList[j].getNeighborId();
for(k = 0; k < ogr->vertices[i].recvFromList.size(); k++) {
if(ogr->vertices[i].recvFromList[k].getNeighborId() == vert) {
ogr->vertices[i].sendToList[j].setNumBytes(ogr->vertices[i].sendToList[j].getNumBytes() + ogr->vertices[i].recvFromList[k].getNumBytes());
ogr->vertices[i].recvFromList.erase(ogr->vertices[i].recvFromList.begin() + k);
}
}
}
}
/** the object load is normalized to an integer between 0 and 256 */
for(i = 0; i < numVertices; i++) {
if(ogr->vertices[i].getVertexLoad() > maxLoad)
maxLoad = ogr->vertices[i].getVertexLoad();
numEdges = numEdges + ogr->vertices[i].sendToList.size() + ogr->vertices[i].recvFromList.size();
}
/* adjacency list */
idx_t *xadj = new idx_t[numVertices + 1];
/* id of the neighbors */
idx_t *adjncy = new idx_t[numEdges];
/* weights of the vertices */
idx_t *vwgt = new idx_t[numVertices];
/* weights of the edges */
idx_t *adjwgt = new idx_t[numEdges];
int edgeNum = 0;
double ratio = 256.0/maxLoad;
for(i = 0; i < numVertices; i++) {
xadj[i] = edgeNum;
vwgt[i] = (int)ceil(ogr->vertices[i].getVertexLoad() * ratio);
for(j = 0; j < ogr->vertices[i].sendToList.size(); j++) {
adjncy[edgeNum] = ogr->vertices[i].sendToList[j].getNeighborId();
adjwgt[edgeNum] = ogr->vertices[i].sendToList[j].getNumBytes();
edgeNum++;
}
for(j = 0; j < ogr->vertices[i].recvFromList.size(); j++) {
adjncy[edgeNum] = ogr->vertices[i].recvFromList[j].getNeighborId();
adjwgt[edgeNum] = ogr->vertices[i].recvFromList[j].getNumBytes();
edgeNum++;
}
}
xadj[i] = edgeNum;
CkAssert(edgeNum == numEdges);
idx_t edgecut; // number of edges cut by the partitioning
idx_t *pemap;
idx_t options[METIS_NOPTIONS];
METIS_SetDefaultOptions(options);
//options[METIS_OPTION_PTYPE] = METIS_PTYPE_RB;
// C style numbering
options[METIS_OPTION_NUMBERING] = 0;
// number of constrains
idx_t ncon = 1;
// number of partitions
idx_t numPes = parr->procs.size();
real_t ubvec[ncon];
// allow 10% imbalance
ubvec[0] = 1.1;
// mapping of objs to partitions
pemap = new idx_t[numVertices];
// Specifies size of vertices for computing the total communication volume
idx_t *vsize = NULL;
// This array of size nparts specifies the desired weight for each partition
// and setting it to NULL indicates graph should be equally divided among
// partitions
real_t *tpwgts = NULL;
int option = 0;
if (WEIGHTED == option) {
// set up the different weights between 0 and 1
tpwgts = new real_t[numPes];
for (i = 0; i < numPes; i++) {
tpwgts[i] = 1.0/(real_t)numPes;
}
} else if (MULTI_CONSTRAINT == option) {
CkAbort("Multiple constraints not implemented.\n");
}
// numVertices: num vertices in the graph; ncon: num balancing constrains
// xadj, adjncy: of size n+1 and adjncy of 2m, adjncy[xadj[i]] through and
// including adjncy[xadj[i+1]-1];
// vwgt: weight of the vertices; vsize: amt of data that needs to be sent
// for ith vertex is vsize[i]
// adjwght: the weight of edges; numPes: total parts
// tpwghts: target partition weight, can pass NULL to equally divide
// ubvec: of size ncon to indicate allowed load imbalance tolerance (> 1.0)
// options: array of options; edgecut: stores the edgecut; pemap: mapping
METIS_PartGraphRecursive(&numVertices, &ncon, xadj, adjncy, vwgt, vsize, adjwgt,
&numPes, tpwgts, ubvec, options, &edgecut, pemap);
delete[] xadj;
delete[] adjncy;
delete[] vwgt;
delete[] adjwgt;
delete[] vsize;
delete[] tpwgts;
if (_lb_args.debug() >= 1) {
CkPrintf("[%d] MetisLB done! \n", CkMyPe());
}
for(i = 0; i < numVertices; i++) {
if(pemap[i] != ogr->vertices[i].getCurrentPe())
ogr->vertices[i].setNewPe(pemap[i]);
}
delete[] pemap;
/** ============================== CLEANUP ================================ */
ogr->convertDecisions(stats);
delete parr;
delete ogr;
}
void TreeMatchLB::work(BaseLB::LDStats* stats)
{
/** ========================= 1st Do Load Balancing =======================*/
/** ========================== INITIALIZATION ============================= */
ProcArray *parr = new ProcArray(stats); // Processor Array
ObjGraph *ogr = new ObjGraph(stats); // Object Graph
/** ============================= STRATEGY ================================ */
parr->resetTotalLoad();
if (_lb_args.debug()>1)
CkPrintf("[%d] In GreedyLB strategy\n",CkMyPe());
int vert;
// max heap of objects
std::sort(ogr->vertices.begin(), ogr->vertices.end(), ObjLoadGreater());
// min heap of processors
std::make_heap(parr->procs.begin(), parr->procs.end(), ProcLoadGreater());
for(vert = 0; vert < ogr->vertices.size(); vert++) {
// Pop the least loaded processor
ProcInfo p = parr->procs.front();
std::pop_heap(parr->procs.begin(), parr->procs.end(), ProcLoadGreater());
parr->procs.pop_back();
// Increment the load of the least loaded processor by the load of the
// 'heaviest' unmapped object
p.totalLoad() += ogr->vertices[vert].getVertexLoad();
ogr->vertices[vert].setNewPe(p.getProcId());
// Insert the least loaded processor with load updated back into the heap
parr->procs.push_back(p);
std::push_heap(parr->procs.begin(), parr->procs.end(), ProcLoadGreater());
}
/** ============================== CLEANUP ================================ */
ogr->convertDecisions(stats); // Send decisions back to LDStats
/** ====================== 2nd do Topology aware mapping ====================*/
int nb_procs;
double **comm_mat;
int i;
int *object_mapping, *permutation;
/* get number of processors and teh greedy load balancing*/
nb_procs = stats->nprocs();
object_mapping=stats->to_proc.getVec();
stats->makeCommHash();
// allocate communication matrix
comm_mat=(double**)malloc(sizeof(double*)*nb_procs);
for(i=0;i<nb_procs;i++){
comm_mat[i]=(double*)calloc(nb_procs,sizeof(double));
}
/* Build the communicartion matrix*/
for(i=0;i<stats->n_comm;i++){
LDCommData &commData = stats->commData[i];
if((!commData.from_proc())&&(commData.recv_type()==LD_OBJ_MSG)){
/* object_mapping[i] is the processors of object i*/
int from = object_mapping[stats->getHash(commData.sender)];
int to = object_mapping[stats->getHash(commData.receiver.get_destObj())];
if(from!=to){
comm_mat[from][to]+=commData.bytes;
comm_mat[to][from]+=commData.bytes;
}
}
}
/* build the topology of the hardware (abe machine here)*/
tm_topology_t *topology=build_abe_topology(nb_procs);
display_topology(topology);
/* compute the affinity tree */
tree_t *comm_tree=build_tree_from_topology(topology,comm_mat,nb_procs,NULL,NULL);
/* Compute the processor permutation*/
permutation=(int*)malloc(sizeof(int)*nb_procs);
map_topology_simple(topology,comm_tree,permutation,NULL);
/*
Apply this perutation to all objects
Side effect: object_mapping points to the stats->to_proc.getVec()
So, these lines change also stats->to_proc.getVec()
*/
for(i=0;i<nb_procs;i++)
object_mapping[i]=permutation[object_mapping[i]];
// free communication matrix;
for(i=0;i<nb_procs;i++){
free(comm_mat[i]);
}
free(comm_mat);
free_topology(topology);
}
