/** swap Walkers with Recv/Send
*
* The algorithm ensures that the load per node can differ only by one walker.
* The communication is one-dimensional.
*/
void GlobalWalkerControl::swapWalkersBlocked(MCWalkerConfiguration& W) {
NumSwaps++;
OffSet[0]=0;
for(int i=0; i<NumContexts; i++) OffSet[i+1]=OffSet[i]+NumPerNode[i];
FairPartition(Cur_pop,NumContexts,FairOffSet);
int toLeft=FairOffSet[MyContext]-OffSet[MyContext];
int toRight=FairOffSet[MyContext+1]-OffSet[MyContext+1];
Walker_t& wRef(*W[0]);
vector<Walker_t*> newW;
int num_deleted=0;
int last=NumPerNode[MyContext]-1;
//scehdule irecv
if(toLeft<0) { // recv from node-1
int dn=-toLeft;
OOMPI_Packed recvBuffer(dn*wRef.byteSize(),OOMPI_COMM_WORLD);
OOMPI_COMM_WORLD[MyContext-1].Recv(recvBuffer);
while(dn) {
Walker_t *awalker= new Walker_t(wRef);
awalker->getMessage(recvBuffer);
newW.push_back(awalker);
--dn;
}
} else if(toLeft>0) { //send to node-1
int dn=toLeft;
num_deleted+=dn;
OOMPI_Packed sendBuffer(dn*wRef.byteSize(),OOMPI_COMM_WORLD);
while(dn) {
W[last]->putMessage(sendBuffer); --dn; --last;
}
OOMPI_COMM_WORLD[MyContext-1].Send(sendBuffer);
}
if(toRight<0) { // send to node+1
int dn=-toRight;
num_deleted+=dn;
OOMPI_Packed sendBuffer(dn*wRef.byteSize(),OOMPI_COMM_WORLD);
while(dn) {
W[last]->putMessage(sendBuffer); --dn; --last;
}
OOMPI_COMM_WORLD[MyContext+1].Send(sendBuffer);
} else if(toRight>0) { //recv from node+1
OOMPI_Packed recvBuffer(toRight*wRef.byteSize(),OOMPI_COMM_WORLD);
OOMPI_COMM_WORLD[MyContext+1].Recv(recvBuffer);
int dn=toRight;
while(dn) {
Walker_t *awalker= new Walker_t(wRef);
awalker->getMessage(recvBuffer);
newW.push_back(awalker);
--dn;
}
}
while(num_deleted>0) {
W.pop_back(); --num_deleted;
}
if(newW.size()) W.insert(W.end(),newW.begin(),newW.end());
}
/** swap Walkers with Irecv/Send
*
* The algorithm ensures that the load per node can differ only by one walker.
* The communication is one-dimensional.
*/
void GlobalWalkerControl::swapWalkersAsync(MCWalkerConfiguration& W) {
NumSwaps++;
OffSet[0]=0;
for(int i=0; i<NumContexts; i++) OffSet[i+1]=OffSet[i]+NumPerNode[i];
FairPartition(Cur_pop,NumContexts,FairOffSet);
int toLeft=FairOffSet[MyContext]-OffSet[MyContext];
int toRight=FairOffSet[MyContext+1]-OffSet[MyContext+1];
Walker_t& wRef(*W[0]);
int num_deleted=0;
int last=NumPerNode[MyContext]-1;
OOMPI_Packed* recvLeftBuffer=0;
OOMPI_Packed* recvRightBuffer=0;
OOMPI_Request recvLeftRequest;
OOMPI_Request recvRightRequest;
//initiate Irecv
if(toLeft<0) { // recv from node-1
int dn=-toLeft;
recvLeftBuffer = new OOMPI_Packed(dn*wRef.byteSize(),OOMPI_COMM_WORLD);
recvLeftRequest = OOMPI_COMM_WORLD[MyContext-1].Irecv(*recvLeftBuffer,MPI_ANY_TAG);
}
if(toRight>0) { //recv from node+1
recvRightBuffer = new OOMPI_Packed(toRight*wRef.byteSize(),OOMPI_COMM_WORLD);
recvRightRequest = OOMPI_COMM_WORLD[MyContext+1].Irecv(*recvRightBuffer,MPI_ANY_TAG);
}
//send the data
if(toLeft>0) { //send to node-1
int dn=toLeft;
num_deleted+=dn;
OOMPI_Packed sendBuffer(dn*wRef.byteSize(),OOMPI_COMM_WORLD);
while(dn) {
W[last]->putMessage(sendBuffer); --dn; --last;
}
OOMPI_COMM_WORLD[MyContext-1].Send(sendBuffer);
}
if(toRight<0) { // send to node+1
int dn=-toRight;
num_deleted+=dn;
OOMPI_Packed sendBuffer(dn*wRef.byteSize(),OOMPI_COMM_WORLD);
while(dn) {
W[last]->putMessage(sendBuffer); --dn; --last;
}
OOMPI_COMM_WORLD[MyContext+1].Send(sendBuffer);
}
while(num_deleted>0) {
W.pop_back(); --num_deleted;
}
if(recvLeftBuffer) {
recvLeftRequest.Wait();
int dn=-toLeft;
while(dn) {
Walker_t *awalker= new Walker_t(wRef);
awalker->getMessage(*recvLeftBuffer);
W.push_back(awalker);
--dn;
}
delete recvLeftBuffer;
recvLeftBuffer=0;
}
if(recvRightBuffer) {
recvRightRequest.Wait();
int dn=toRight;
while(dn) {
Walker_t *awalker= new Walker_t(wRef);
awalker->getMessage(*recvRightBuffer);
W.push_back(awalker);
--dn;
}
delete recvRightBuffer;
recvRightBuffer=0;
}
}
