// - retrive the stack base address
static VOID OnThreadStart(THREADID, CONTEXT *ctxt, INT32, VOID *){
ADDRINT stackBase = PIN_GetContextReg(ctxt, REG_STACK_PTR);
ProcInfo *pInfo = ProcInfo::getInstance();
pInfo->addThreadStackAddress(stackBase);
pInfo->addThreadTebAddress();
//MYINFO("-----------------a NEW Thread started!--------------------\n");
}
inline void handleTransfer(int randomly_obj_id, ProcInfo& p, int possible_pe, std::vector<int>* parr_objs, ObjGraph *ogr, ProcArray* parr) {
ogr->vertices[randomly_obj_id].setNewPe(possible_pe);
parr_objs[possible_pe].push_back(randomly_obj_id);
ProcInfo &possible_pe_procinfo = parr->procs[possible_pe];
p.totalLoad() -= ogr->vertices[randomly_obj_id].getVertexLoad();
possible_pe_procinfo.totalLoad() += ogr->vertices[randomly_obj_id].getVertexLoad();
eraseObjFromParrObjs(parr_objs[p.getProcId()], randomly_obj_id);
//CkPrintf("After transfered %d from %d : Load %E to %d : Load %E\n", randomly_obj_id, p.getProcId(), p.getTotalLoad(),
// possible_pe, possible_pe_procinfo.getTotalLoad());
}
VOID VirtualFreeHook(UINT32 address_to_free){
MYINFO("Have to free the address %08x\n" , address_to_free);
ProcInfo *pInfo = ProcInfo::getInstance();
std::vector<HeapZone> HeapMap = pInfo->getHeapMap();
int index_to_remove = -1;
MYINFO("HeapZone before free");
for(unsigned index=0; index < HeapMap.size(); index++) {
if(address_to_free == pInfo->getHeapZoneByIndex(index)->begin){
index_to_remove = index;
}
}
if(index_to_remove != -1){
pInfo->deleteHeapZone(index_to_remove);
}
MYINFO("HeapZone after free");
}
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
}
bool operator()(ProcInfo p1, ProcInfo p2) {
return (p1.getTotalLoad() > p2.getTotalLoad());
}
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);
}