int main(int argc, char *argv[]) {
unsigned int power2[12] = {1,2,4,8,16,32,64,128,256,512,1024,2048};
int numMSG;
int dim;
int MASTER;
/******************* Number of MSGs **********/
numMSG = atoi(argv[1]);
dim = atoi(argv[2]);
/******************* MPI Init ****************** */
unsigned int rank, size;
int MPI_ERR;
MPI_Status status;
MPI_Init(&argc, &argv);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &size);
MASTER = size -1 ;
/************************** MASTER ******************************/
if (rank == MASTER){
int *msg = (int *)calloc(MSGSIZE, sizeof(int));
int i=0;
int numberOfPrints = (size-1) * (numMSG*2);
int dest = 0;
int next;
int doneCounter = 0;
int printCounter = 0;
int treeDone = 0;
makeMessage(msg, MASTER, ZERO, TREE, 0);
MPI_Send(&msg[0], MSGSIZE, MPI_INT, ZERO, DATA_TAG, MPI_COMM_WORLD);
while(1){
MPI_Recv(&msg[0], MSGSIZE, MPI_INT, ZERO, DATA_TAG, MPI_COMM_WORLD, &status);
if(msg[TYPE] == TREEDONE){
treeDone += 1;
if(treeDone < 2 ){
makeMessage(msg, MASTER, size-2, INITDONE, 0);
MPI_Send(&msg[0], MSGSIZE, MPI_INT, size-2, DATA_TAG, MPI_COMM_WORLD);
}
}
if(msg[TYPE] == PRINT){
printCounter += 1;
print(msg, msg[SOURCEORIGINAL], dim, power2);
}
if(msg[TYPE] == DONE){
doneCounter += 1;
}
if(doneCounter == (size-1) && printCounter == ( (numMSG * 2) * (size-1) )){
break;
}
}
// Send a STOP messages using Reduce Forward to Zero
dest = ZERO;
makeMessage(msg, MASTER, dest, REDUCEFORWARD, 0);
next = compute_next_dest(rank, dest, power2);
MPI_Send(&msg[0], MSGSIZE, MPI_INT, dest, DATA_TAG, MPI_COMM_WORLD);
// printf("%d\n", printCounter);
}
/************************** Workers ******************************/
else{
int i;
int x;
int dest;
int *msg = (int *)calloc(MSGSIZE, sizeof(int));
int *children = (int *)calloc(dim, sizeof(int));
int childCounter = 0;
int msgno = 0;
unsigned int next;
int flag;
int msgCounter = 0;
int makeMsgCount = 0;
int ackCounter = 0;
int stopCount = 0;
int parent = 0;
int parentSent = 0;
int parentCounter = 0;
int initDone = 0;
int forBcastCounter = 0;
int bcastCouter = 0;
int bSent = 0;
int reduceForwardCount = 0;
int reduceBackwardCount = 0;
int reduceFlag = 1;
while(1){
MPI_Iprobe(MPI_ANY_SOURCE, MPI_ANY_TAG, MPI_COMM_WORLD, &flag, &status);
if(flag == 1){
MPI_Recv(&msg[0] , MSGSIZE, MPI_INT, status.MPI_SOURCE, MPI_ANY_TAG, MPI_COMM_WORLD, &status);
if(rank == ZERO){
if(msg[TYPE] == PRINT){
MPI_Send(&msg[0] , MSGSIZE, MPI_INT, MASTER, DATA_TAG, MPI_COMM_WORLD);
}
else if(msg[TYPE] == DONE){
MPI_Send(&msg[0] , MSGSIZE, MPI_INT, MASTER, DATA_TAG, MPI_COMM_WORLD);
}
else if(msg[TYPE] == TREEDONE){
msg[SOURCE] = ZERO;
msg[DEST] = MASTER;
MPI_Send(&msg[0] , MSGSIZE, MPI_INT, MASTER, DATA_TAG, MPI_COMM_WORLD);
}
}
dest = msg[DEST];
/* If the message is for you */
if(dest == rank){
/* If you got an ACK, increament your counter */
if(msg[TYPE] == ACK){
ackCounter += 1;
/* Send for Printing */
msg[SOURCE] = rank;
msg[DEST] = ZERO;
msg[TYPEORIGINAL] = ACK ; //msg[TYPE];
msg[TYPE] = PRINT;
next = compute_next_dest(rank, ZERO, power2);
if(rank == ZERO){
MPI_Send(&msg[0] , MSGSIZE, MPI_INT, MASTER, DATA_TAG, MPI_COMM_WORLD);
}else{
MPI_Send(&msg[0] , MSGSIZE, MPI_INT, next, DATA_TAG, MPI_COMM_WORLD);
}
}
/* If you got a message destined to you, send an ACK */
else if(msg[TYPE] == MSG){
/* Send for printing */
msg[SOURCE] = rank;
msg[DEST] = ZERO;
msg[TYPEORIGINAL] = MSG ;
msg[TYPE] = PRINT;
next = compute_next_dest(rank, ZERO, power2);
if(rank == ZERO){
MPI_Send(&msg[0] , MSGSIZE, MPI_INT, MASTER, DATA_TAG, MPI_COMM_WORLD);
}else{
MPI_Send(&msg[0] , MSGSIZE, MPI_INT, next, DATA_TAG, MPI_COMM_WORLD);
}
/* Send Ack */
dest = msg[SOURCEORIGINAL];
msgno = msg[MSGNO];
makeMessage(msg, rank, dest, ACK, msgno);
next = compute_next_dest(rank, dest, power2);
msg[HOPCNT] = 1;
msg[SOURCEORIGINAL] = rank;
MPI_Send(&msg[0] , MSGSIZE, MPI_INT, next, DATA_TAG, MPI_COMM_WORLD);
}
else if(msg[TYPE] == TREE && msg[SOURCE] == ZERO){
parent = findParent(rank, dim);
makeMessage(msg, rank, parent, PARENT, msgno);
next = compute_next_dest(rank, parent, power2);
MPI_Send(&msg[0] , MSGSIZE, MPI_INT, next, DATA_TAG, MPI_COMM_WORLD);
// printf("Rank %d sending to %d \n", rank, msg[DEST]);
}
else if(msg[TYPE] == TREE && msg[SOURCE] == MASTER){
int layer = (int)(pow(2,dim)) - (((int)(pow(2,dim))) / 2);
for(i = 0 ; i < layer ; i++){
makeMessage(msg, ZERO, i, TREE, 0);
next = compute_next_dest(ZERO, i, power2);
MPI_Send(&msg[0] , MSGSIZE, MPI_INT, next, DATA_TAG, MPI_COMM_WORLD);
}
}
else if(msg[TYPE] == PARENT){
children[childCounter] = msg[SOURCE];
childCounter += 1;
if(rank != size-2 && parentSent == 0){
parent = findParent(rank, dim);
makeMessage(msg, rank, parent, PARENT, msgno);
next = compute_next_dest(rank, parent, power2);
MPI_Send(&msg[0] , MSGSIZE, MPI_INT, next, DATA_TAG, MPI_COMM_WORLD);
parentSent = 1;
}
if(rank == size-2){
parentCounter += 1;
if(parentCounter == ((int)pow(2,(dim-1)) - 1)){
dest = ZERO;
makeMessage(msg, rank, dest, TREEDONE, msgno);
next = compute_next_dest(rank, dest, power2);
MPI_Send(&msg[0] , MSGSIZE, MPI_INT, next, DATA_TAG, MPI_COMM_WORLD);
}
}
}
else if(msg[TYPE] == STOP){
if(childCounter > 0){
for(i = 0 ; i < childCounter ; i++){
dest = children[i];
makeMessage(msg, rank, dest, STOP, msgno);
MPI_Send(&msg[0] , MSGSIZE, MPI_INT, dest, DATA_TAG, MPI_COMM_WORLD);
}
}
break;
}
else if(msg[TYPE] == INITDONE){
if(childCounter > 0){
for(i = 0 ; i < childCounter ; i++){
dest = children[i];
makeMessage(msg, rank, dest, INITDONE, msgno);
MPI_Send(&msg[0] , MSGSIZE, MPI_INT, dest, DATA_TAG, MPI_COMM_WORLD);
}
}
initDone = 1;
}
/* Only 11..1 gets this */
else if(msg[TYPE] == FORBCAST){
if(childCounter > 0){
for(i = 0 ; i < childCounter ; i++){
dest = children[i];
makeMessage(msg, rank, dest, BCAST, msgno);
MPI_Send(&msg[0] , MSGSIZE, MPI_INT, dest, DATA_TAG, MPI_COMM_WORLD);
bcastCouter += 1;
// printf("Rank %d send bcast type %d to %d - bcounter %d \n", rank, msg[TYPE], msg[DEST], bcastCouter);
}
// bcastCouter += childCounter;
}
/* FORBCASTACK immediately to Zero */
makeMessage(msg, rank, ZERO, FORBCASTACK, msgno);
MPI_Send(&msg[0] , MSGSIZE, MPI_INT, ZERO, DATA_TAG, MPI_COMM_WORLD);
// printf("Rank %d send forbcastack type %d to %d - bcounter %d \n", rank, msg[TYPE], msg[DEST], bcastCouter);
}
else if(msg[TYPE] == BCAST){
if(childCounter > 0){
for(i = 0 ; i < childCounter ; i++){
dest = children[i];
makeMessage(msg, rank, dest, BCAST, msgno);
MPI_Send(&msg[0] , MSGSIZE, MPI_INT, dest, DATA_TAG, MPI_COMM_WORLD);
bcastCouter += 1;
// printf("Rank %d send bcast type %d to %d - bcounter %d \n", rank, msg[TYPE], msg[DEST], bcastCouter);
}
}
else{
/* BCASTACK to parent only when you received as many acks as you sent bcasts */
makeMessage(msg, rank, parent, BCASTACK, msgno);
MPI_Send(&msg[0] , MSGSIZE, MPI_INT, parent, DATA_TAG, MPI_COMM_WORLD);
// printf("BCAST ACK from %d to %d \n", rank, msg[DEST]);
}
}
/* Only Zero receives this */
else if(msg[TYPE] == FORBCASTACK){
// printf("Rank %d got Type %d from %d \n",rank, msg[TYPE], msg[SOURCE] );
forBcastCounter -= 1;
}
else if(msg[TYPE] == BCASTACK){
bcastCouter -= 1;
if(bcastCouter == 0 && rank != size-2){
/* BCASTACK to parent only when you received as many acks as you sent bcasts */
makeMessage(msg, rank, parent, BCASTACK, msgno);
MPI_Send(&msg[0] , MSGSIZE, MPI_INT, parent, DATA_TAG, MPI_COMM_WORLD);
// printf("BCAST ACK from %d to %d \n", rank, msg[DEST]);
}
}
else if(msg[TYPE] == REDUCEFORWARD){
// if(msg[SOURCEORIGINAL] == rank){
// reduceFlag = 0;
// printf("Returned to Source : %d - Loop Finished!\n", rank);
// }
// else{
int msgNumber = msg[MSGNO];
int source = msg[SOURCEORIGINAL];
dest = myNext(rank, dim);
makeMessage(msg, rank, dest, REDUCEFORWARD, msgNumber);
msg[SOURCEORIGINAL] = source;
MPI_Send(&msg[0], MSGSIZE, MPI_INT, dest, DATA_TAG, MPI_COMM_WORLD);
reduceFlag = 0;
printf("HOP : Rank %d myNext : %d - Message Number : %d - TERMINATED !\n", rank, dest, msg[MSGNO]);
break;
// }
}
else if(msg[TYPE] == REDUCEBACKWARD){
if(msg[SOURCEORIGINAL] == rank){
reduceFlag = 0;
printf("Returned to Source : %d - Loop Finished!\n", rank);
}
else{
int msgNumber = msg[MSGNO];
int source = msg[SOURCEORIGINAL];
dest = myPrev(rank, dim);
makeMessage(msg, rank, dest, REDUCEBACKWARD, msgNumber);
msg[SOURCEORIGINAL] = source;
MPI_Send(&msg[0], MSGSIZE, MPI_INT, dest, DATA_TAG, MPI_COMM_WORLD);
reduceFlag = 0;
printf("HOP : Rank %d myPrev : %d - Message Number : %d \n", rank, dest, msg[MSGNO]);
}
}
}
/* If you are not destination */
else{
/* MSG or ACK */
if(msg[TYPE] == MSG || msg[TYPE] == ACK){
/* add to number of hobs */
msg[HOPCNT] += 1;
int nHubs = msg[HOPCNT];
msg[(HOPCNT + nHubs)] = rank;
next = compute_next_dest(rank, dest, power2);
MPI_Send(&msg[0] , MSGSIZE, MPI_INT, next, DATA_TAG, MPI_COMM_WORLD);
}
/* Pass it along : print, done, stop, parent, parentfound */
else{
next = compute_next_dest(rank, dest, power2);
MPI_Send(&msg[0] , MSGSIZE, MPI_INT, next, DATA_TAG, MPI_COMM_WORLD);
}
}
/* if you are done send a DONE message to Zero (forBcastCounter == 0 && bcastCouter == 0 && reduceFlag == 0) */
if(ackCounter == numMSG ){
// printf("Rank %d is DoNe !!!! \n", rank);
dest = ZERO;
makeMessage(msg, rank, dest, DONE, msgno);
MPI_Send(&msg[0] , MSGSIZE, MPI_INT, dest, DATA_TAG, MPI_COMM_WORLD);
ackCounter += 1;
// printf("Rank %d TYPE %d to %d\n", rank, msg[TYPE], msg[DEST]);
}
}
/* Flag = 0 */
else{
if(initDone == 1){
/* Still have to make your messages */
if(makeMsgCount < numMSG){
dest = randomGenerator(size-1) ;
makeMessage(msg, rank, dest, MSG, msgno);
next = compute_next_dest(rank, dest, power2);
msg[HOPCNT] = 1;
msg[SOURCEORIGINAL] = rank;
MPI_Send(&msg[0] , MSGSIZE, MPI_INT, next, DATA_TAG, MPI_COMM_WORLD);
makeMsgCount += 1;
}
}
}
}
}
MPI_Finalize();
return 0;
}
bool Slave::checkMessages() {
//if (engine.decisionLevel() <= engine.assumptions.size()) return false;
double t = wallClockTime();
if (t <= next_check && engine.decisionLevel() > 0) return false;
real_time += t;
// cpu_time += cpuTime();
int received;
//fprintf(stderr, "%d: CheckMessages! \n", thread_no);
while (true) {
MPI_Iprobe(0, MPI_ANY_TAG, MPI_COMM_WORLD, &received, &s);
if (!received) break;
switch (s.MPI_TAG) {
case INTERRUPT_TAG:
MPI_Recv(NULL, 0, MPI_INT, 0, INTERRUPT_TAG, MPI_COMM_WORLD, &s);
if (PAR_DEBUG) fprintf(stderr, "%d: Interrupted! %f\n", thread_no, wallClockTime());
real_time -= wallClockTime();
// cpu_time -= cpuTime();
return true;
case STEAL_TAG:
//fprintf(stderr, "%d asked to split job\n", thread_no);
splitJob();
break;
case LEARNTS_TAG:
//fprintf(stderr, "%d receives learnts\n", thread_no);
receiveLearnts();
break;
case SOLUTION_TAG:
receiveSolution();
break;
case INT_SOLUTION_TAG:
assert(false && "This version does not support forwarding solution! ");
break;
case NEW_EXPORT_LIMIT:
// Only 1 Integer: New length
int newLength;
//fprintf(stderr, "Receiving new export limit! \n");
MPI_Recv(&newLength, 1, MPI_INT, 0, NEW_EXPORT_LIMIT, MPI_COMM_WORLD, &s);
so.maxClSz = newLength;
//fprintf(stderr, "Done, set to %d \n", newLength);
break;
case NEW_STATE_TAG:
int newState;
MPI_Recv(&newState, 1, MPI_INT, 0, NEW_STATE_TAG, MPI_COMM_WORLD, &s);
if(_state != newState){
if(_state == RUNNING_GREEDY)
engine.stop_init_phase = true;
_state = newState;
}
MPI_Bsend(&_state, 1, MPI_INT, 0, SLAVE_NEW_STATE_TAG , MPI_COMM_WORLD);
break;
default:
fprintf(stderr, "assert false!\n");
assert(false);
}
}
if ((++checks%int(report_freq/check_freq) == 0)
|| unitFound){
unitFound = false;
sendReport();
}
t = wallClockTime();
if(engine.decisionLevel() > 0)
next_check = t + check_freq;
real_time -= t;
if(engine.decisionLevel() == 0){
for(int i = 0 ; i < storedClauses.size() ; i++){
assert(storedClauses[i].size() > 0);
assert(storedClauses[i][0] == tmp_header[i]);
// TODO:
vec<int> tmp;
for(int j = 0 ; j < storedClauses[i].size() ; j++)
tmp.push(storedClauses[i][j]);
sat.convertToClause(tmp);
sat.addLearnt();
}
storedClauses.clear();
tmp_header.clear();
}
return false;
}
void
comm_receive(const switchboard_t* sb)
{
#ifdef WITH_MPI
MPI_Status status;
status.MPI_ERROR = MPI_SUCCESS;
int flag = 0;
MPI_Iprobe(MPI_ANY_SOURCE, MPI_ANY_TAG, MPI_COMM_WORLD, &flag, &status);
if (status.MPI_ERROR != MPI_SUCCESS)
mpi_fatal(sb, &status, "Got error while probing for messages");
while(flag) { /* only execute the body if `flag` is true */
int size;
MPI_Get_count(&status, MPI_BYTE, &size);
void* xa = xmalloc(size);
switch(status.MPI_TAG) {
case TAG_ROW_SPARSE: {
sparse_row_t* new_row = sparse_row_alloc_placed(xa, size);
assert(NULL != new_row);
MPI_Recv(new_row, size, MPI_BYTE,
status.MPI_SOURCE, status.MPI_TAG, MPI_COMM_WORLD,
&status);
if (status.MPI_ERROR != MPI_SUCCESS)
mpi_fatal(sb, &status, "Got error while receiving sparse row");
# ifndef NDEBUG
int recv_size;
MPI_Get_count(&status, MPI_BYTE, &recv_size);
assert(size == recv_size);
# endif
assert(is_local(sb, new_row->starting_column_));
vpu_t* vpu = vpu_for_column(sb, new_row->starting_column_);
assert(vpu != NULL);
vpu_recv_row(vpu, new_row, ROW_SPARSE);
break;
};
case TAG_ROW_DENSE: {
dense_row_t* new_row = dense_row_alloc_placed(xa, size);
assert(NULL != new_row);
MPI_Recv(new_row, size, MPI_BYTE,
status.MPI_SOURCE, status.MPI_TAG, MPI_COMM_WORLD,
&status);
if (status.MPI_ERROR != MPI_SUCCESS)
mpi_fatal(sb, &status, "Got error while receiving dense row");
# ifndef NDEBUG
int recv_size;
MPI_Get_count(&status, MPI_BYTE, &recv_size);
assert(size == recv_size);
# endif
assert(is_local(sb, new_row->starting_column_));
vpu_t* vpu = vpu_for_column(sb, new_row->starting_column_);
assert(NULL != vpu);
vpu_recv_row(vpu, new_row, ROW_DENSE);
break;
};
case TAG_END: {
// "end" message received; no new rows will be coming.
// But some other rows could have arrived or could
// already be in the `inbox`, so we need to make another
// pass anyway. All this boils down to: set a flag,
// make another iteration, and end the loop next time.
coord_t column;
MPI_Recv(&column, 1, MPI_LONG,
status.MPI_SOURCE, status.MPI_TAG, MPI_COMM_WORLD,
&status);
if (status.MPI_ERROR != MPI_SUCCESS)
mpi_fatal(sb, &status, "Got error while receiving sparse END tag");
# ifndef NDEBUG
int recv_size;
MPI_Get_count(&status, MPI_BYTE, &recv_size);
assert(size == recv_size);
# endif
assert(column >= 0 && is_local(sb, column));
vpu_t* vpu = vpu_for_column(sb, column);
assert(NULL != vpu);
vpu_end_phase(vpu);
break;
};
}; // switch(status.MPI_TAG)
MPI_Iprobe(MPI_ANY_SOURCE, MPI_ANY_TAG, MPI_COMM_WORLD, &flag, &status);
if (status.MPI_ERROR != MPI_SUCCESS)
mpi_fatal(sb, &status, "Got error while probing for messages");
}; // while(MPI_Iprobe)
#endif
}
int main(int argc, char** argv) {
int ROOT=_ROOT;
int ACK_COUNT=0;
int MSG_SIZE = _MSG_SIZE;
int i,j,k,source,my_rank,num_nodes,my_state,tmpmsg;
MPI_Status status;
// initialize mpi stuff
MPI_Init(&argc, &argv);
MPI_Comm_size(MPI_COMM_WORLD,&num_nodes);
MPI_Comm_rank(MPI_COMM_WORLD,&my_rank);
int msg[MSG_SIZE];
for (i=0;i<MSG_SIZE;i++) msg[i] = -1; //build msg
int done = 0;
int flag = 0;
if (ROOT == my_rank) {
my_state = R0;
while (!done) {
tmpmsg = get_random_msg();
for (i=0;i<MSG_SIZE;i++) msg[i] = tmpmsg; //build msg
// send msg to nodes
for (j=1;j<num_nodes;j++)
MPI_Send(&msg,MSG_SIZE,MPI_INT,j,0,MPI_COMM_WORLD); // blocking send, not ideal for efficiency
// check for ACK
flag=0;
MPI_Iprobe(MPI_ANY_SOURCE,MPI_ANY_TAG,MPI_COMM_WORLD,&flag,&status); // a non-blocking check for ACK message from nodes in FINAL state
if (1 == flag) {
source = status.MPI_SOURCE;
MPI_Recv(&msg,MSG_SIZE,MPI_INT,source,0,MPI_COMM_WORLD,&status);
if (num_nodes-1 == ++ACK_COUNT)
++done;
}
}
} else {
my_state = Q0;
int done = 0;
while (!done) {
MPI_Recv(&msg,MSG_SIZE,MPI_INT,ROOT,0,MPI_COMM_WORLD,&status);
//printf("Node %d received MSG=%d from Node %d\n",my_rank,msg,ROOT);
// react based on msg
switch (msg[0]) { // presumably, the first element of the msg array is the same as all other elements
case A:
if (Q0 == my_state) {
my_state = next_state_proc(my_state,msg[0]);
printf("Node %d now in state %d\n",my_rank,my_state);
}
break;
case B:
if (Q1 == my_state) {
my_state = next_state_proc(my_state,msg[0]);
printf("Node %d now in state %d\n",my_rank,my_state);
}
break;
case C:
if (Q2 == my_state) {
my_state = next_state_proc(my_state,msg[0]);
printf("Node %d now in FINAL state %d (shutting down...)\n",my_rank,my_state);
for (i=0;i<MSG_SIZE;i++) msg[i] = ACK; //build msg
MPI_Send(&msg,MSG_SIZE,MPI_INT,ROOT,0,MPI_COMM_WORLD); // blocking send, not ideal for efficiency
++done;
}
break;
}
}
}
MPI_Finalize();
exit(EXIT_SUCCESS);
}
void Broker::checkReceivedRanking(int* flag, MPI_Status* status){
MPI_Iprobe(MPI_ANY_SOURCE, MPI_ANY_TAG, MPI_COMM_WORLD, flag, status);
}
int main(int argc, char *argv[]) {
unsigned int power2[12] = {1,2,4,8,16,32,64,128,256,512,1024,2048};
int numMSG;
int dim;
int MASTER;
/******************* Number of MSGs **********/
numMSG = atoi(argv[1]);
dim = atoi(argv[2]);
/******************* MPI Init ****************** */
unsigned int rank, size;
int MPI_ERR;
MPI_Status status;
MPI_Init(&argc, &argv);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &size);
MASTER = size -1 ;
/************************** MASTER ******************************/
if (rank == MASTER){
int *msg = (int *)calloc(MSGSIZE, sizeof(int));
int i=0;
int numberOfPrints = (size-1) * (numMSG*2);
int dest = 0;
int doneCounter = 0;
int printCounter = 0;
int treeDone = 0;
makeMessage(msg, MASTER, ZERO, TREE, 0);
MPI_Send(&msg[0], MSGSIZE, MPI_INT, ZERO, DATA_TAG, MPI_COMM_WORLD);
while(1){
MPI_Recv(&msg[0], MSGSIZE, MPI_INT, ZERO, DATA_TAG, MPI_COMM_WORLD, &status);
if(msg[TYPE] == TREEDONE){
treeDone = 1;
}
if(msg[TYPE] == PRINT){
printCounter += 1;
print(msg, msg[SOURCEORIGINAL], dim, power2);
}
if(msg[TYPE] == DONE){
doneCounter += 1;
}
if(treeDone == 1 && doneCounter == (size-1) && printCounter == ( (numMSG * 2) * (size-1) )){
break;
}
}
// Send a STOP messages to One
dest = size-2;
makeMessage(msg, MASTER, dest, STOP, 0);
MPI_Send(&msg[0], MSGSIZE, MPI_INT, dest, DATA_TAG, MPI_COMM_WORLD);
// printf("%d\n", printCounter);
}
/************************** Workers ******************************/
else{
int i;
int x;
int dest;
int *msg = (int *)calloc(MSGSIZE, sizeof(int));
int *children = (int *)calloc(dim, sizeof(int));
int childCounter = 0;
int msgno = 0;
unsigned int next;
int flag;
int msgCounter = 0;
int makeMsgCount = 0;
int ackCounter = 0;
int stopCount = 0;
int parent = 0;
int parentSent = 0;
int parentCounter = 0;
while(1){
MPI_Iprobe(MPI_ANY_SOURCE, MPI_ANY_TAG, MPI_COMM_WORLD, &flag, &status);
if(flag == 1){
MPI_Recv(&msg[0] , MSGSIZE, MPI_INT, status.MPI_SOURCE, MPI_ANY_TAG, MPI_COMM_WORLD, &status);
if(rank == ZERO){
if(msg[TYPE] == PRINT){
MPI_Send(&msg[0] , MSGSIZE, MPI_INT, MASTER, DATA_TAG, MPI_COMM_WORLD);
}
else if(msg[TYPE] == DONE){
MPI_Send(&msg[0] , MSGSIZE, MPI_INT, MASTER, DATA_TAG, MPI_COMM_WORLD);
}
// else if(msg[TYPE] == STOP){
// stopCount += 1;
// msg[SOURCE] = ZERO;
// msg[DEST] = (size-1) - stopCount;
// next = compute_next_dest(ZERO, msg[DEST], power2);
// MPI_Send(&msg[0] , MSGSIZE, MPI_INT, next, DATA_TAG, MPI_COMM_WORLD);
// }
else if(msg[TYPE] == TREEDONE){
msg[SOURCE] = ZERO;
msg[DEST] = MASTER;
MPI_Send(&msg[0] , MSGSIZE, MPI_INT, MASTER, DATA_TAG, MPI_COMM_WORLD);
}
}
dest = msg[DEST];
/* If the message is for you */
if(dest == rank){
/* you are done so STOP */
// if(msg[TYPE] == STOP && msg[SOURCE] == ZERO){
// break;
// }
/* If you got an ACK, increament your counter */
if(msg[TYPE] == ACK){
ackCounter += 1;
/* Send for Printing */
msg[SOURCE] = rank;
msg[DEST] = ZERO;
msg[TYPEORIGINAL] = ACK ; //msg[TYPE];
msg[TYPE] = PRINT;
next = compute_next_dest(rank, ZERO, power2);
if(rank == ZERO){
MPI_Send(&msg[0] , MSGSIZE, MPI_INT, MASTER, DATA_TAG, MPI_COMM_WORLD);
}else{
MPI_Send(&msg[0] , MSGSIZE, MPI_INT, next, DATA_TAG, MPI_COMM_WORLD);
}
}
/* If you got a message destined to you, send an ACK */
else if(msg[TYPE] == MSG){
/* Send for printing */
msg[SOURCE] = rank;
msg[DEST] = ZERO;
msg[TYPEORIGINAL] = MSG ;
msg[TYPE] = PRINT;
next = compute_next_dest(rank, ZERO, power2);
if(rank == ZERO){
MPI_Send(&msg[0] , MSGSIZE, MPI_INT, MASTER, DATA_TAG, MPI_COMM_WORLD);
}else{
MPI_Send(&msg[0] , MSGSIZE, MPI_INT, next, DATA_TAG, MPI_COMM_WORLD);
}
/* Send Ack */
dest = msg[SOURCEORIGINAL];
msgno = msg[MSGNO];
makeMessage(msg, rank, dest, ACK, msgno);
next = compute_next_dest(rank, dest, power2);
msg[HOPCNT] = 1;
msg[SOURCEORIGINAL] = rank;
MPI_Send(&msg[0] , MSGSIZE, MPI_INT, next, DATA_TAG, MPI_COMM_WORLD);
}
else if(msg[TYPE] == TREE && msg[SOURCE] == ZERO){
parent = findParent(rank, dim);
makeMessage(msg, rank, parent, PARENT, msgno);
next = compute_next_dest(rank, parent, power2);
MPI_Send(&msg[0] , MSGSIZE, MPI_INT, next, DATA_TAG, MPI_COMM_WORLD);
// printf("Rank %d sending to %d \n", rank, msg[DEST]);
}
else if(msg[TYPE] == TREE && msg[SOURCE] == MASTER){
int layer = (int)(pow(2,dim)) - (((int)(pow(2,dim))) / 2);
for(i = 0 ; i < layer ; i++){
makeMessage(msg, ZERO, i, TREE, 0);
next = compute_next_dest(ZERO, i, power2);
MPI_Send(&msg[0] , MSGSIZE, MPI_INT, next, DATA_TAG, MPI_COMM_WORLD);
}
}
else if(msg[TYPE] == PARENT){
children[childCounter] = msg[SOURCE];
childCounter += 1;
if(rank != size-2 && parentSent == 0){
parent = findParent(rank, dim);
makeMessage(msg, rank, parent, PARENT, msgno);
next = compute_next_dest(rank, parent, power2);
MPI_Send(&msg[0] , MSGSIZE, MPI_INT, next, DATA_TAG, MPI_COMM_WORLD);
parentSent = 1;
}
if(rank == size-2){
parentCounter += 1;
if(parentCounter == ((int)pow(2,(dim-1)) - 1)){
dest = ZERO;
makeMessage(msg, rank, dest, TREEDONE, msgno);
next = compute_next_dest(rank, dest, power2);
MPI_Send(&msg[0] , MSGSIZE, MPI_INT, next, DATA_TAG, MPI_COMM_WORLD);
}
}
}
else if(msg[TYPE] == STOP){
if(childCounter > 0){
for(i = 0 ; i < childCounter ; i++){
dest = children[i];
makeMessage(msg, rank, dest, STOP, msgno);
MPI_Send(&msg[0] , MSGSIZE, MPI_INT, dest, DATA_TAG, MPI_COMM_WORLD);
}
}
break;
}
}
/* If you are not destination */
else{
/* MSG or ACK */
if(msg[TYPE] == MSG || msg[TYPE] == ACK){
/* add to number of hobs */
msg[HOPCNT] += 1;
int nHubs = msg[HOPCNT];
msg[(HOPCNT + nHubs)] = rank;
next = compute_next_dest(rank, dest, power2);
MPI_Send(&msg[0] , MSGSIZE, MPI_INT, next, DATA_TAG, MPI_COMM_WORLD);
}
/* Pass it along : print, done, stop, parent, parentfound */
else{
next = compute_next_dest(rank, dest, power2);
MPI_Send(&msg[0] , MSGSIZE, MPI_INT, next, DATA_TAG, MPI_COMM_WORLD);
}
}
/* if you are done send a DONE message to Zero */
if(ackCounter == numMSG){
dest = ZERO;
makeMessage(msg, rank, dest, DONE, msgno);
MPI_Send(&msg[0] , MSGSIZE, MPI_INT, dest, DATA_TAG, MPI_COMM_WORLD);
ackCounter += 1;
// printf("Rank %d TYPE %d to %d\n", rank, msg[TYPE], msg[DEST]);
}
}
/* Flag = 0 */
else{
/* Still have to make your messages */
if(makeMsgCount < numMSG){
dest = randomGenerator(size-1) ;
makeMessage(msg, rank, dest, MSG, msgno);
next = compute_next_dest(rank, dest, power2);
msg[HOPCNT] = 1;
msg[SOURCEORIGINAL] = rank;
MPI_Send(&msg[0] , MSGSIZE, MPI_INT, next, DATA_TAG, MPI_COMM_WORLD);
makeMsgCount += 1;
}
}
}
}
MPI_Finalize();
return 0;
}
FC_FUNC(mpi_iprobe, MPI_IPROBE)(int * source, int * tag, int * comm,
int * flag, int *status, int * ierr)
{
*ierr = MPI_Iprobe(*source, *tag, *comm, flag, mpi_c_status(status));
}
int main(int argc, char **argv) {
int procid, nproc, i;
MPI_Win llist_win;
llist_ptr_t head_ptr, tail_ptr;
MPI_Init(&argc, &argv);
MPI_Comm_rank(MPI_COMM_WORLD, &procid);
MPI_Comm_size(MPI_COMM_WORLD, &nproc);
#ifdef TEST_MPI3_ROUTINES
MPIX_Win_create_dynamic(MPI_INFO_NULL, MPI_COMM_WORLD, &llist_win);
/* Process 0 creates the head node */
if (procid == 0)
head_ptr.disp = alloc_elem(-1, llist_win);
/* Broadcast the head pointer to everyone */
head_ptr.rank = 0;
MPI_Bcast(&head_ptr.disp, 1, MPI_AINT, 0, MPI_COMM_WORLD);
tail_ptr = head_ptr;
/* All processes concurrently append NUM_ELEMS elements to the list */
for (i = 0; i < NUM_ELEMS; i++) {
llist_ptr_t new_elem_ptr;
int success;
/* Create a new list element and register it with the window */
new_elem_ptr.rank = procid;
new_elem_ptr.disp = alloc_elem(procid, llist_win);
/* Append the new node to the list. This might take multiple attempts if
others have already appended and our tail pointer is stale. */
do {
llist_ptr_t next_tail_ptr = nil;
MPI_Win_lock(MPI_LOCK_SHARED, tail_ptr.rank, MPI_MODE_NOCHECK, llist_win);
MPIX_Compare_and_swap((void*) &new_elem_ptr.rank, (void*) &nil.rank,
(void*) &next_tail_ptr.rank, MPI_INT, tail_ptr.rank,
(MPI_Aint) &(((llist_elem_t*)tail_ptr.disp)->next.rank), llist_win);
MPI_Win_unlock(tail_ptr.rank, llist_win);
success = (next_tail_ptr.rank == nil.rank);
if (success) {
int i, flag;
MPI_Aint result;
MPI_Win_lock(MPI_LOCK_SHARED, tail_ptr.rank, MPI_MODE_NOCHECK, llist_win);
MPIX_Fetch_and_op(&new_elem_ptr.disp, &result, MPI_AINT, tail_ptr.rank,
(MPI_Aint) &(((llist_elem_t*)tail_ptr.disp)->next.disp),
MPI_REPLACE, llist_win);
/* Note: accumulate is faster, since we don't need the result. Replacing with
Fetch_and_op to create a more complete test case. */
/*
MPI_Accumulate(&new_elem_ptr.disp, 1, MPI_AINT, tail_ptr.rank,
(MPI_Aint) &(((llist_elem_t*)tail_ptr.disp)->next.disp), 1,
MPI_AINT, MPI_REPLACE, llist_win);
*/
MPI_Win_unlock(tail_ptr.rank, llist_win);
tail_ptr = new_elem_ptr;
/* For implementations that use pt-to-pt messaging, force progress for other threads'
RMA operations. */
for (i = 0; i < NPROBE; i++)
MPI_Iprobe(MPI_ANY_SOURCE, MPI_ANY_TAG, MPI_COMM_WORLD, &flag, MPI_STATUS_IGNORE);
} else {
/* Tail pointer is stale, fetch the displacement. May take multiple tries
if it is being updated. */
do {
MPI_Aint junk = 0;
MPI_Win_lock(MPI_LOCK_SHARED, tail_ptr.rank, MPI_MODE_NOCHECK, llist_win);
MPIX_Fetch_and_op(NULL, &next_tail_ptr.disp, MPI_AINT, tail_ptr.rank,
(MPI_Aint) &(((llist_elem_t*)tail_ptr.disp)->next.disp),
MPIX_NO_OP, llist_win);
MPI_Win_unlock(tail_ptr.rank, llist_win);
} while (next_tail_ptr.disp == nil.disp);
tail_ptr = next_tail_ptr;
}
} while (!success);
}
MPI_Barrier(MPI_COMM_WORLD);
/* Traverse the list and verify that all processes inserted exactly the correct
number of elements. */
if (procid == 0) {
int have_root = 0;
int errors = 0;
int *counts, count = 0;
counts = (int*) malloc(sizeof(int) * nproc);
assert(counts != NULL);
for (i = 0; i < nproc; i++)
counts[i] = 0;
tail_ptr = head_ptr;
/* Walk the list and tally up the number of elements inserted by each rank */
while (tail_ptr.disp != nil.disp) {
llist_elem_t elem;
MPI_Win_lock(MPI_LOCK_SHARED, tail_ptr.rank, MPI_MODE_NOCHECK, llist_win);
MPI_Get(&elem, sizeof(llist_elem_t), MPI_BYTE,
tail_ptr.rank, tail_ptr.disp, sizeof(llist_elem_t), MPI_BYTE, llist_win);
MPI_Win_unlock(tail_ptr.rank, llist_win);
tail_ptr = elem.next;
/* This is not the root */
if (have_root) {
assert(elem.value >= 0 && elem.value < nproc);
counts[elem.value]++;
count++;
if (verbose) {
int last_elem = tail_ptr.disp == nil.disp;
printf("%2d%s", elem.value, last_elem ? "" : " -> ");
if (count % ELEM_PER_ROW == 0 && !last_elem)
printf("\n");
}
}
/* This is the root */
else {
assert(elem.value == -1);
have_root = 1;
}
}
if (verbose)
printf("\n\n");
/* Verify the counts we collected */
for (i = 0; i < nproc; i++) {
int expected = NUM_ELEMS;
if (counts[i] != expected) {
printf("Error: Rank %d inserted %d elements, expected %d\n", i, counts[i], expected);
errors++;
}
}
printf("%s\n", errors == 0 ? " No Errors" : "FAIL");
free(counts);
}
MPI_Win_free(&llist_win);
/* Free all the elements in the list */
for ( ; my_elems_count > 0; my_elems_count--)
MPI_Free_mem(my_elems[my_elems_count-1]);
#else /* ! TEST_MPI3_ROUTINES */
if (procid == 0)
printf(" No Errors\n");
#endif
MPI_Finalize();
return 0;
}
/* data1, odata1 and odata2 are packed in the format (for communication):
data[0] = is_max, no of is
data[1] = size of is[0]
...
data[is_max] = size of is[is_max-1]
data[is_max + 1] = data(is[0])
...
data[is_max+1+sum(size of is[k]), k=0,...,i-1] = data(is[i])
...
data2 is packed in the format (for creating output is[]):
data[0] = is_max, no of is
data[1] = size of is[0]
...
data[is_max] = size of is[is_max-1]
data[is_max + 1] = data(is[0])
...
data[is_max + 1 + Mbs*i) = data(is[i])
...
*/
static PetscErrorCode MatIncreaseOverlap_MPISBAIJ_Once(Mat C,PetscInt is_max,IS is[])
{
Mat_MPISBAIJ *c = (Mat_MPISBAIJ*)C->data;
PetscErrorCode ierr;
PetscMPIInt size,rank,tag1,tag2,*len_s,nrqr,nrqs,*id_r1,*len_r1,flag,len,*iwork;
const PetscInt *idx_i;
PetscInt idx,isz,col,*n,*data1,**data1_start,*data2,*data2_i,*data,*data_i;
PetscInt Mbs,i,j,k,*odata1,*odata2;
PetscInt proc_id,**odata2_ptr,*ctable=0,*btable,len_max,len_est;
PetscInt proc_end=0,len_unused,nodata2;
PetscInt ois_max; /* max no of is[] in each of processor */
char *t_p;
MPI_Comm comm;
MPI_Request *s_waits1,*s_waits2,r_req;
MPI_Status *s_status,r_status;
PetscBT *table; /* mark indices of this processor's is[] */
PetscBT table_i;
PetscBT otable; /* mark indices of other processors' is[] */
PetscInt bs=C->rmap->bs,Bn = c->B->cmap->n,Bnbs = Bn/bs,*Bowners;
IS garray_local,garray_gl;
PetscFunctionBegin;
ierr = PetscObjectGetComm((PetscObject)C,&comm);
CHKERRQ(ierr);
size = c->size;
rank = c->rank;
Mbs = c->Mbs;
ierr = PetscObjectGetNewTag((PetscObject)C,&tag1);
CHKERRQ(ierr);
ierr = PetscObjectGetNewTag((PetscObject)C,&tag2);
CHKERRQ(ierr);
/* create tables used in
step 1: table[i] - mark c->garray of proc [i]
step 3: table[i] - mark indices of is[i] when whose=MINE
table[0] - mark incideces of is[] when whose=OTHER */
len = PetscMax(is_max, size);
CHKERRQ(ierr);
ierr = PetscMalloc2(len,&table,(Mbs/PETSC_BITS_PER_BYTE+1)*len,&t_p);
CHKERRQ(ierr);
for (i=0; i<len; i++) {
table[i] = t_p + (Mbs/PETSC_BITS_PER_BYTE+1)*i;
}
ierr = MPIU_Allreduce(&is_max,&ois_max,1,MPIU_INT,MPI_MAX,comm);
CHKERRQ(ierr);
/* 1. Send this processor's is[] to other processors */
/*---------------------------------------------------*/
/* allocate spaces */
ierr = PetscMalloc1(is_max,&n);
CHKERRQ(ierr);
len = 0;
for (i=0; i<is_max; i++) {
ierr = ISGetLocalSize(is[i],&n[i]);
CHKERRQ(ierr);
len += n[i];
}
if (!len) {
is_max = 0;
} else {
len += 1 + is_max; /* max length of data1 for one processor */
}
ierr = PetscMalloc1(size*len+1,&data1);
CHKERRQ(ierr);
ierr = PetscMalloc1(size,&data1_start);
CHKERRQ(ierr);
for (i=0; i<size; i++) data1_start[i] = data1 + i*len;
ierr = PetscMalloc4(size,&len_s,size,&btable,size,&iwork,size+1,&Bowners);
CHKERRQ(ierr);
/* gather c->garray from all processors */
ierr = ISCreateGeneral(comm,Bnbs,c->garray,PETSC_COPY_VALUES,&garray_local);
CHKERRQ(ierr);
ierr = ISAllGather(garray_local, &garray_gl);
CHKERRQ(ierr);
ierr = ISDestroy(&garray_local);
CHKERRQ(ierr);
ierr = MPI_Allgather(&Bnbs,1,MPIU_INT,Bowners+1,1,MPIU_INT,comm);
CHKERRQ(ierr);
Bowners[0] = 0;
for (i=0; i<size; i++) Bowners[i+1] += Bowners[i];
if (is_max) {
/* hash table ctable which maps c->row to proc_id) */
ierr = PetscMalloc1(Mbs,&ctable);
CHKERRQ(ierr);
for (proc_id=0,j=0; proc_id<size; proc_id++) {
for (; j<C->rmap->range[proc_id+1]/bs; j++) ctable[j] = proc_id;
}
/* hash tables marking c->garray */
ierr = ISGetIndices(garray_gl,&idx_i);
CHKERRQ(ierr);
for (i=0; i<size; i++) {
table_i = table[i];
ierr = PetscBTMemzero(Mbs,table_i);
CHKERRQ(ierr);
for (j = Bowners[i]; j<Bowners[i+1]; j++) { /* go through B cols of proc[i]*/
ierr = PetscBTSet(table_i,idx_i[j]);
CHKERRQ(ierr);
}
}
ierr = ISRestoreIndices(garray_gl,&idx_i);
CHKERRQ(ierr);
} /* if (is_max) */
ierr = ISDestroy(&garray_gl);
CHKERRQ(ierr);
/* evaluate communication - mesg to who, length, and buffer space */
for (i=0; i<size; i++) len_s[i] = 0;
/* header of data1 */
for (proc_id=0; proc_id<size; proc_id++) {
iwork[proc_id] = 0;
*data1_start[proc_id] = is_max;
data1_start[proc_id]++;
for (j=0; j<is_max; j++) {
if (proc_id == rank) {
*data1_start[proc_id] = n[j];
} else {
*data1_start[proc_id] = 0;
}
data1_start[proc_id]++;
}
}
for (i=0; i<is_max; i++) {
ierr = ISGetIndices(is[i],&idx_i);
CHKERRQ(ierr);
for (j=0; j<n[i]; j++) {
idx = idx_i[j];
*data1_start[rank] = idx;
data1_start[rank]++; /* for local proccessing */
proc_end = ctable[idx];
for (proc_id=0; proc_id<=proc_end; proc_id++) { /* for others to process */
if (proc_id == rank) continue; /* done before this loop */
if (proc_id < proc_end && !PetscBTLookup(table[proc_id],idx)) continue; /* no need for sending idx to [proc_id] */
*data1_start[proc_id] = idx;
data1_start[proc_id]++;
len_s[proc_id]++;
}
}
/* update header data */
for (proc_id=0; proc_id<size; proc_id++) {
if (proc_id== rank) continue;
*(data1 + proc_id*len + 1 + i) = len_s[proc_id] - iwork[proc_id];
iwork[proc_id] = len_s[proc_id];
}
ierr = ISRestoreIndices(is[i],&idx_i);
CHKERRQ(ierr);
}
nrqs = 0;
nrqr = 0;
for (i=0; i<size; i++) {
data1_start[i] = data1 + i*len;
if (len_s[i]) {
nrqs++;
len_s[i] += 1 + is_max; /* add no. of header msg */
}
}
for (i=0; i<is_max; i++) {
ierr = ISDestroy(&is[i]);
CHKERRQ(ierr);
}
ierr = PetscFree(n);
CHKERRQ(ierr);
ierr = PetscFree(ctable);
CHKERRQ(ierr);
/* Determine the number of messages to expect, their lengths, from from-ids */
ierr = PetscGatherNumberOfMessages(comm,NULL,len_s,&nrqr);
CHKERRQ(ierr);
ierr = PetscGatherMessageLengths(comm,nrqs,nrqr,len_s,&id_r1,&len_r1);
CHKERRQ(ierr);
/* Now post the sends */
ierr = PetscMalloc2(size,&s_waits1,size,&s_waits2);
CHKERRQ(ierr);
k = 0;
for (proc_id=0; proc_id<size; proc_id++) { /* send data1 to processor [proc_id] */
if (len_s[proc_id]) {
ierr = MPI_Isend(data1_start[proc_id],len_s[proc_id],MPIU_INT,proc_id,tag1,comm,s_waits1+k);
CHKERRQ(ierr);
k++;
}
}
/* 2. Receive other's is[] and process. Then send back */
/*-----------------------------------------------------*/
len = 0;
for (i=0; i<nrqr; i++) {
if (len_r1[i] > len) len = len_r1[i];
}
ierr = PetscFree(len_r1);
CHKERRQ(ierr);
ierr = PetscFree(id_r1);
CHKERRQ(ierr);
for (proc_id=0; proc_id<size; proc_id++) len_s[proc_id] = iwork[proc_id] = 0;
ierr = PetscMalloc1(len+1,&odata1);
CHKERRQ(ierr);
ierr = PetscMalloc1(size,&odata2_ptr);
CHKERRQ(ierr);
ierr = PetscBTCreate(Mbs,&otable);
CHKERRQ(ierr);
len_max = ois_max*(Mbs+1); /* max space storing all is[] for each receive */
len_est = 2*len_max; /* estimated space of storing is[] for all receiving messages */
ierr = PetscMalloc1(len_est+1,&odata2);
CHKERRQ(ierr);
nodata2 = 0; /* nodata2+1: num of PetscMalloc(,&odata2_ptr[]) called */
odata2_ptr[nodata2] = odata2;
len_unused = len_est; /* unused space in the array odata2_ptr[nodata2]-- needs to be >= len_max */
k = 0;
while (k < nrqr) {
/* Receive messages */
ierr = MPI_Iprobe(MPI_ANY_SOURCE,tag1,comm,&flag,&r_status);
CHKERRQ(ierr);
if (flag) {
ierr = MPI_Get_count(&r_status,MPIU_INT,&len);
CHKERRQ(ierr);
proc_id = r_status.MPI_SOURCE;
ierr = MPI_Irecv(odata1,len,MPIU_INT,proc_id,r_status.MPI_TAG,comm,&r_req);
CHKERRQ(ierr);
ierr = MPI_Wait(&r_req,&r_status);
CHKERRQ(ierr);
/* Process messages */
/* make sure there is enough unused space in odata2 array */
if (len_unused < len_max) { /* allocate more space for odata2 */
ierr = PetscMalloc1(len_est+1,&odata2);
CHKERRQ(ierr);
odata2_ptr[++nodata2] = odata2;
len_unused = len_est;
}
ierr = MatIncreaseOverlap_MPISBAIJ_Local(C,odata1,OTHER,odata2,&otable);
CHKERRQ(ierr);
len = 1 + odata2[0];
for (i=0; i<odata2[0]; i++) len += odata2[1 + i];
/* Send messages back */
ierr = MPI_Isend(odata2,len,MPIU_INT,proc_id,tag2,comm,s_waits2+k);
CHKERRQ(ierr);
k++;
odata2 += len;
len_unused -= len;
len_s[proc_id] = len; /* num of messages sending back to [proc_id] by this proc */
}
}
ierr = PetscFree(odata1);
CHKERRQ(ierr);
ierr = PetscBTDestroy(&otable);
CHKERRQ(ierr);
/* 3. Do local work on this processor's is[] */
/*-------------------------------------------*/
/* make sure there is enough unused space in odata2(=data) array */
len_max = is_max*(Mbs+1); /* max space storing all is[] for this processor */
if (len_unused < len_max) { /* allocate more space for odata2 */
ierr = PetscMalloc1(len_est+1,&odata2);
CHKERRQ(ierr);
odata2_ptr[++nodata2] = odata2;
}
data = odata2;
ierr = MatIncreaseOverlap_MPISBAIJ_Local(C,data1_start[rank],MINE,data,table);
CHKERRQ(ierr);
ierr = PetscFree(data1_start);
CHKERRQ(ierr);
/* 4. Receive work done on other processors, then merge */
/*------------------------------------------------------*/
/* get max number of messages that this processor expects to recv */
ierr = MPIU_Allreduce(len_s,iwork,size,MPI_INT,MPI_MAX,comm);
CHKERRQ(ierr);
ierr = PetscMalloc1(iwork[rank]+1,&data2);
CHKERRQ(ierr);
ierr = PetscFree4(len_s,btable,iwork,Bowners);
CHKERRQ(ierr);
k = 0;
while (k < nrqs) {
/* Receive messages */
ierr = MPI_Iprobe(MPI_ANY_SOURCE,tag2,comm,&flag,&r_status);
CHKERRQ(ierr);
if (flag) {
ierr = MPI_Get_count(&r_status,MPIU_INT,&len);
CHKERRQ(ierr);
proc_id = r_status.MPI_SOURCE;
ierr = MPI_Irecv(data2,len,MPIU_INT,proc_id,r_status.MPI_TAG,comm,&r_req);
CHKERRQ(ierr);
ierr = MPI_Wait(&r_req,&r_status);
CHKERRQ(ierr);
if (len > 1+is_max) { /* Add data2 into data */
data2_i = data2 + 1 + is_max;
for (i=0; i<is_max; i++) {
table_i = table[i];
data_i = data + 1 + is_max + Mbs*i;
isz = data[1+i];
for (j=0; j<data2[1+i]; j++) {
col = data2_i[j];
if (!PetscBTLookupSet(table_i,col)) data_i[isz++] = col;
}
data[1+i] = isz;
if (i < is_max - 1) data2_i += data2[1+i];
}
}
k++;
}
}
ierr = PetscFree(data2);
CHKERRQ(ierr);
ierr = PetscFree2(table,t_p);
CHKERRQ(ierr);
/* phase 1 sends are complete */
ierr = PetscMalloc1(size,&s_status);
CHKERRQ(ierr);
if (nrqs) {
ierr = MPI_Waitall(nrqs,s_waits1,s_status);
CHKERRQ(ierr);
}
ierr = PetscFree(data1);
CHKERRQ(ierr);
/* phase 2 sends are complete */
if (nrqr) {
ierr = MPI_Waitall(nrqr,s_waits2,s_status);
CHKERRQ(ierr);
}
ierr = PetscFree2(s_waits1,s_waits2);
CHKERRQ(ierr);
ierr = PetscFree(s_status);
CHKERRQ(ierr);
/* 5. Create new is[] */
/*--------------------*/
for (i=0; i<is_max; i++) {
data_i = data + 1 + is_max + Mbs*i;
ierr = ISCreateGeneral(PETSC_COMM_SELF,data[1+i],data_i,PETSC_COPY_VALUES,is+i);
CHKERRQ(ierr);
}
for (k=0; k<=nodata2; k++) {
ierr = PetscFree(odata2_ptr[k]);
CHKERRQ(ierr);
}
ierr = PetscFree(odata2_ptr);
CHKERRQ(ierr);
PetscFunctionReturn(0);
}
int main(int argc, char *argv[]) {
unsigned int power2[12] = {1,2,4,8,16,32,64,128,256,512,1024,2048};
int numMSG;
int dim;
int MASTER;
/******************* Number of MSGs **********/
numMSG = atoi(argv[1]);
dim = atoi(argv[2]);
/******************* MPI Init ****************** */
unsigned int rank, size;
int MPI_ERR;
MPI_Status status;
MPI_Init(&argc, &argv);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &size);
MASTER = size -1 ;
/************************** MASTER ******************************/
if (rank == MASTER){
}
/************************** Workers ******************************/
else{
int i;
int x;
int dest;
int *msg = (int *)calloc(MSGSIZE, sizeof(int));
int msgno = 0;
unsigned int next;
int flag;
int msgCounter = 0;
int makeMsgCount = 0;
while(1){
MPI_Iprobe(MPI_ANY_SOURCE, MPI_ANY_TAG, MPI_COMM_WORLD, &flag, &status);
if(flag == 1){
MPI_Recv(&msg[0] , MSGSIZE, MPI_INT, status.MPI_SOURCE, DATA_TAG, MPI_COMM_WORLD, &status);
dest = msg[1];
next = compute_next_dest(rank, dest, power2);
if(dest == rank){
print(msg, rank, dim, power2);
}else{
/* add to number of hobs */
int nHubs = msg[HOPCNT];
msg[HOPCNT] += 1;
msg[(HOPCNT + nHubs + 1)] = rank;
MPI_Send(&msg[0] , MSGSIZE, MPI_INT, next, DATA_TAG, MPI_COMM_WORLD);
}
}
/* Flag = 0 */
else{
if(makeMsgCount < numMSG){
dest = randomGenerator(size) ;
makeMessage(msg, rank, dest, MSG, msgno);
next = compute_next_dest(rank, dest, power2);
MPI_Send(&msg[0] , MSGSIZE, MPI_INT, next, DATA_TAG, MPI_COMM_WORLD);
makeMsgCount += 1;
// printf("Rank %d sending \n", rank);
}
}
}
}
MPI_Finalize();
return 0;
}
/*
* === FUNCTION ======================================================================
* Name: main
* Description:
* =====================================================================================
*/
int
main ( int argc, char *argv[] )
{
// initialize MPI
MPI_Init (&argc, &argv);
// we have to remember the number of PEs
int numpes;
MPI_Comm_size (MPI_COMM_WORLD, &numpes);
//for this we need 3 PEs
assert(numpes == 3);
// which rank does this process have?
int myid;
MPI_Comm_rank (MPI_COMM_WORLD, &myid);
switch ( myid ) {
case 0 :
{
printf("0: I have A....sending it!\n");
char *msg = "A";
MPI_Send(msg, strlen(msg) + 1, MPI_CHAR, 1, 0, MPI_COMM_WORLD);
MPI_Send(msg, strlen(msg) + 1, MPI_CHAR, 2, 0, MPI_COMM_WORLD);
break;
}
case 1 :
{
char *msg = "B";
MPI_Status status;
MPI_Probe(MPI_ANY_SOURCE, MPI_ANY_TAG, MPI_COMM_WORLD, &status);
int msglen;
MPI_Get_count(&status, MPI_CHAR, &msglen);
assert(msglen > 0);
char *recMsg = malloc(msglen * sizeof(char));
MPI_Recv (recMsg, msglen, MPI_CHAR, status.MPI_SOURCE, status.MPI_TAG, MPI_COMM_WORLD, &status);
printf("1: I have B....\n");
printf("1: Received an %s so sending my B\n", recMsg);
MPI_Send(msg, strlen(msg) + 1, MPI_CHAR, 2, 0, MPI_COMM_WORLD);
break;
}
case 2 :
{
int hasMail;
int receivedNotAll = 2;
MPI_Status status;
while(receivedNotAll){
MPI_Iprobe(MPI_ANY_SOURCE, MPI_ANY_TAG, MPI_COMM_WORLD, &hasMail, &status);
if (hasMail) {
int msglen;
MPI_Get_count(&status, MPI_CHAR, &msglen);
assert(msglen > 0);
char *msg = malloc(msglen * sizeof(char));
if (!msg) {
fprintf(stderr, "Could not allocate memory for %d bytes in message\n", msglen);
exit(-2);
}
MPI_Recv (msg, msglen, MPI_CHAR, status.MPI_SOURCE, status.MPI_TAG, MPI_COMM_WORLD, &status);
printf("2: Received a %s\n", msg);
free(msg);
receivedNotAll--;
}
}
break;
}
default:
break;
} /* ----- end switch ----- */
MPI_Finalize();
return EXIT_SUCCESS;
} /* ---------- end of function main ---------- */
void Master::solve() {
lastCubeFinished = wallClockTime();
num_threads = so.num_threads;
job_start_time.growTo(num_threads, DONT_DISTURB);
job_start_time_backup.growTo(num_threads, DONT_DISTURB);
cur_job.growTo(num_threads, NULL);
lhead.growTo(num_threads, 0);
last_send_learnts.growTo(num_threads, 0);
global_learnts.reserve(10000000);
long maxCycleTime = 0;
stoppedInit = false;
if(engine.opt_var){
bestObjReceived = engine.opt_type ? engine.opt_var->getMin() : engine.opt_var->getMax();
}
if(so.purePortfolio){
for(int i = 0 ; i < num_threads ; i++)
job_queue.push(new SClause());
}
else
job_queue.push(new SClause());
for(int i = 0 ; i < num_threads ; i++){
if(so.greedyInit && (i % 3 < 2)){
slaveStates.push_back(RUNNING_GREEDY);
slavesRunningGreedy++;
setState(i, RUNNING_GREEDY);
}
else{
slaveStates.push_back(NORMAL_SEARCH);
setState(i, NORMAL_SEARCH);
}
}
MPI_Buffer_attach(malloc(MPI_BUFFER_SIZE), MPI_BUFFER_SIZE);
// Search:
int lastPrinted = time(NULL);
int tStart = lastPrinted;
long lastSleep = clock();
bool stealJobsNow = true;
while (status == RES_UNK && time(NULL) < so.time_out) {
//fprintf(stderr, "Trying to send jobs...\n");
while (num_free_slaves > 0 && job_queue.size() > 0) {
if(!sendJob())
break;
}
/**************************************************************************
* Ask all jobs to finish the first phase, if this has not happened yet.
* */
if(!stoppedInit && time(NULL) > engine.half_time){
if(PAR_DEBUG) fprintf(stderr, "Asking remaining init-workers to continue with normal search...\n");
for(int i = 0 ; i < num_threads ; i++){
if(slaveStates[i] == RUNNING_GREEDY){
setState(i, NORMAL_SEARCH);
}
}
stoppedInit = true;
}
int received;
MPI_Iprobe(MPI_ANY_SOURCE, MPI_ANY_TAG, MPI_COMM_WORLD, &received, &s);
if (received) {
double t;
profile_start();
// printf("Profiling started...\n");
switch (s.MPI_TAG) {
case REPORT_TAG:
// printf("It was a report!\n");
receiveReport();
// printf("ReceiveReport done!\n");
profile_end("receive report", 0);
continue;
case SPLIT_TAG:
// printf("Split tag received!\n");
receiveJobs();
profile_end("receive jobs", 0);
continue;
case REPORT_SOLUTION_TAG:
if(PAR_DEBUG) fprintf(stderr, "Received solution! \n");
receiveSolution();
continue;
case REPORT_OPTIMUM_TAG:
receiveOptObj();
continue;
case SOLUTION_PHASE_TAG:
receivePhase();
continue;
case SLAVE_NEW_STATE_TAG:
int dummy;
MPI_Recv(&dummy, 1, MPI_INT, s.MPI_SOURCE, SLAVE_NEW_STATE_TAG, MPI_COMM_WORLD, &s);
if(PAR_DEBUG) fprintf(stderr, "Setting state of slave %d to %d\n", s.MPI_SOURCE-1, dummy);
if(slaveStates[s.MPI_SOURCE-1] == RUNNING_GREEDY && dummy == NORMAL_SEARCH)
slavesRunningGreedy--;
slaveStates[s.MPI_SOURCE-1] = dummy;
continue;
continue;
default:
assert(false);
}
}
if (job_queue.size() < 2*num_threads-2 && !so.purePortfolio ) {
// Steal jobs if
// - normal mode
// - greedy-init, and at least one job finished now...
if(!so.greedyInit || slavesRunningGreedy < num_threads){
stealJobs();
//continue;
}
}
long now = clock();
maxCycleTime = std::max(maxCycleTime, now - lastSleep);
usleep(500);
lastSleep = clock();
}
if (PAR_DEBUG){
fprintf(stderr, "Waiting for slaves to terminate...\n");
fprintf(stderr, "Max cycle time: %d (%lf)\n", maxCycleTime, (double)maxCycleTime/CLOCKS_PER_SEC);
fprintf(stderr, "End of problem called\n");
}
MPI_Request r;
for (int i = 0; i < num_threads; i++) {
MPI_Isend(NULL, 0, MPI_INT, i+1, INTERRUPT_TAG, MPI_COMM_WORLD, &r);
MPI_Isend(NULL, 0, MPI_INT, i+1, FINISH_TAG, MPI_COMM_WORLD, &r);
}
while (num_free_slaves != num_threads) {
MPI_Probe(MPI_ANY_SOURCE, MPI_ANY_TAG, MPI_COMM_WORLD, &s);
int thread_no = s.MPI_SOURCE-1;
MPI_Get_count(&s, MPI_INT, &message_length);
message = (int*) malloc(message_length*sizeof(int));
MPI_Recv(message, message_length, MPI_INT, s.MPI_SOURCE, s.MPI_TAG, MPI_COMM_WORLD, &s);
if (s.MPI_TAG == REPORT_TAG) {
if (message[0] != RES_SEA) {
assert(job_start_time[thread_no] != NOT_WORKING);
num_free_slaves++;
job_start_time[thread_no] = NOT_WORKING;
if (PAR_DEBUG) fprintf(stderr, "%d is free, %f\n", thread_no, wallClockTime());
}
}
free(message);
}
collectStats();
if(PAR_DEBUG) fprintf(stderr, "Master terminating with obj in %d and %d, bestResult= %d\n", engine.opt_var->getMin(), engine.opt_var->getMax(), bestObjReceived);
if(so.verbosity > 0)
printStats();
}
double farmer(int numprocs) {
MPI_Status status;
int i, flag, source, w_id;
double result, incoming[5], *derp;
int w_out[numprocs-1];
// Set up stack
stack *stack = new_stack();
double data[5] = {A, B, F(A), F(B), (F(A)+F(B)) * (B-A)/2};
push(data, stack);
// Set up queue
queue *queue = new_queue();
for (i=1; i<numprocs; i++) {
push_queue(i, queue);
w_out[i-1] = 0;
}
while (1) {
if (!is_empty(stack)) {
derp = pop(stack);
w_id = pop_queue(queue);
w_out[w_id] = 1;
MPI_Send(derp, 5, MPI_DOUBLE, w_id, TAG_DO_TASK, MPI_COMM_WORLD);
tasks_per_process[w_id]++;
}
// peek for messages
MPI_Iprobe(MPI_ANY_SOURCE, MPI_ANY_TAG, MPI_COMM_WORLD, &flag, &status);
// if there is a message
if (flag) {
switch(status.MPI_TAG) {
case TAG_ADD_TASK:
// receive data and push onto stack
source = status.MPI_SOURCE;
MPI_Recv(&incoming, 5, MPI_DOUBLE, source, TAG_ADD_TASK, MPI_COMM_WORLD, &status);
push(incoming, stack);
if (w_out[source]) {
push_queue(source, queue);
w_out[source] = 0;
}
break;
case TAG_RESULT:
source = status.MPI_SOURCE;
MPI_Recv(&incoming, 5, MPI_DOUBLE, source, TAG_RESULT, MPI_COMM_WORLD, &status);
result += incoming[4];
if (w_out[source]) {
push_queue(source, queue);
w_out[source] = 0;
}
break;
}
}
// ready to finish?
if (workers_available(queue) == numprocs-1 && is_empty(stack)) { break; }
}
// kill and free
for (i=1; i<numprocs; i++) {
MPI_Send(&data, 5, MPI_DOUBLE, i, TAG_KILL, MPI_COMM_WORLD);
}
free_stack(stack);
free_queue(queue);
return result;
}
void nomad_fun(const rksvm_problem *prob, const rksvm_parameter *param, Scheduler *scheduler, double *Q)
{
int l = prob->l;
int global_l = prob->global_l;
int thread_count = param->thread_count;
int nr_ranks = param->nr_ranks;
int current_rank = mpi_get_rank();
int *nr_send = scheduler->nr_send;
int *nr_recv = scheduler->nr_recv;
//atomic variables
atomic<int> count_setup_threads;
count_setup_threads = 0;
atomic<int> computed_data_nodes;//record the number of data_nodes that have been utilized.
computed_data_nodes = 0;
atomic<int> sended_count;//record the number of data_nodes that have been sended.
sended_count = 0;
atomic<int> recvd_count;////record the number of data_nodes that have been received.
recvd_count = 0;
// two auxiliary atomic flags for both sending and receiving
atomic<bool> flag_send_ready;
flag_send_ready = false;
atomic<bool> flag_receive_ready;
flag_receive_ready = false;
//build several job queues and one sending queue
con_queue *job_queues = callocator<con_queue>().allocate(thread_count);
for(int i=0;i<thread_count;i++)
callocator<con_queue>().construct(job_queues + i);
con_queue send_queue;
//initilize job queues
int interval = (int)ceil((double)prob->l/thread_count);
int thread_id = 0;
for(int i=0;i<l;i++)
{
data_node *copy_x = nullptr;
copy_x = scheduler->pop();
if((i!=0)&&(i%interval==0))
thread_id++;
job_queues[thread_id].push(copy_x);
}
//the first function
auto QMatrix = [&](struct data_node *copy_x)->void{//{{{
int i = 0;
int global_index = copy_x->global_index;
for(i=0;i<l;i++)
{
rksvm_node *s = prob->x[i];
rksvm_node *t = copy_x->x;
Q[global_index + i*global_l] = k_function(s,t,*param);
}
return;
};//}}}
//the second function
auto computer_fun = [&](int thread_id)->void{///{{{
count_setup_threads++;
while(count_setup_threads < thread_count)
{
std::this_thread::yield();
}
while(true)
{
if(computed_data_nodes == global_l)
break;
data_node *copy_x = nullptr;
bool success = job_queues[thread_id].try_pop(copy_x);
if(success)
{
if(copy_x->first_time)
{
QMatrix(copy_x);
computed_data_nodes++;
if(nr_ranks==1)
{
int lth = copy_x->length;
callocator<rksvm_node>().deallocate(copy_x->x, lth);
callocator<data_node>().destroy(copy_x);
callocator<data_node>().deallocate(copy_x,1);
}
else
{
copy_x->first_time = false;
send_queue.push(copy_x);
flag_send_ready = true;
}
}
else
{
QMatrix(copy_x);
computed_data_nodes++;
copy_x->current_rank = current_rank;
int next_rank = cyclic_loading_rank(copy_x->current_rank, nr_ranks);
if(next_rank==copy_x->initial_rank)
{
int lth = copy_x->length;
callocator<rksvm_node>().deallocate(copy_x->x, lth);
callocator<data_node>().destroy(copy_x);
callocator<data_node>().deallocate(copy_x,1);
}
else
{
send_queue.push(copy_x);
}
}
}
}
return;
};///}}}
//the third function
auto sender_fun = [&]()->void{///{{{
while(flag_send_ready == false)
{
std::this_thread::yield();
}
int lth;
int msg_bytenum;
while(true)
{
if(sended_count == nr_send[current_rank])
break;
data_node *copy_x = nullptr;
bool success = send_queue.try_pop(copy_x);
if(success)
{
int next_rank = cyclic_loading_rank(copy_x->current_rank, nr_ranks);
if(next_rank == copy_x->initial_rank)
{
lth = copy_x->length;
callocator<rksvm_node>().deallocate(copy_x->x, lth);
callocator<data_node>().destroy(copy_x);
callocator<data_node>().deallocate(copy_x,1);
}
else
{
lth = copy_x->length;
msg_bytenum = sizeof(bool)+4*sizeof(int)+lth*sizeof(rksvm_node);
char *send_message = sallocator<char>().allocate(msg_bytenum);
*(reinterpret_cast<bool *>(send_message)) = copy_x->first_time;
*(reinterpret_cast<int *>(send_message + sizeof(bool))) = copy_x->length;
*(reinterpret_cast<int *>(send_message + sizeof(bool) + sizeof(int))) = copy_x->initial_rank;
*(reinterpret_cast<int *>(send_message + sizeof(bool) + 2*sizeof(int))) = copy_x->current_rank;
*(reinterpret_cast<int *>(send_message + sizeof(bool) + 3*sizeof(int))) = copy_x->global_index;
rksvm_node *dest = reinterpret_cast<rksvm_node *>(send_message + sizeof(bool) + 4*sizeof(int));
std::copy(copy_x->x, copy_x->x + lth, dest);
flag_receive_ready = true;
MPI_Ssend(send_message, msg_bytenum, MPI_CHAR, next_rank, 1, MPI_COMM_WORLD);
//destroying
callocator<rksvm_node>().deallocate(copy_x->x, lth);
callocator<data_node>().destroy(copy_x);
callocator<data_node>().deallocate(copy_x,1);
//record the sended count
sended_count++;
sallocator<char>().deallocate(send_message, msg_bytenum);
}
}
}
return;
};///}}}
//the fourth function
auto receiver_fun = [&]()->void{///{{{
while(flag_receive_ready == false)
{
std::this_thread::yield();
}
int flag = 0;
int src_rank;
int lth;
MPI_Status status;
while(true)
{
if(recvd_count == nr_recv[mpi_get_rank()])
break;
MPI_Iprobe(MPI_ANY_SOURCE, 1, MPI_COMM_WORLD, &flag, &status);
if(flag == 0)
{
std::this_thread::yield();
}
else
{
src_rank = status.MPI_SOURCE;
int msg_size = 0;
MPI_Get_count(&status, MPI_CHAR, &msg_size);
char *recv_message = sallocator<char>().allocate(msg_size);
MPI_Recv(recv_message, msg_size, MPI_CHAR, src_rank, 1, MPI_COMM_WORLD, &status);
//recovering
data_node *copy_x = callocator<data_node>().allocate(1);
copy_x->first_time = *(reinterpret_cast<bool *>(recv_message));
copy_x->length = *(reinterpret_cast<int *>(recv_message + sizeof(bool)));
copy_x->initial_rank = *(reinterpret_cast<int *>(recv_message + sizeof(bool) + sizeof(int)));
copy_x->current_rank = *(reinterpret_cast<int *>(recv_message + sizeof(bool) + 2*sizeof(int)));
copy_x->global_index = *(reinterpret_cast<int *>(recv_message + sizeof(bool) + 3*sizeof(int)));
rksvm_node *dest = reinterpret_cast<rksvm_node *>(recv_message + sizeof(bool) + 4*sizeof(int));
//please notice that the approach to recover cp_x->x
lth = copy_x->length;
copy_x->x = callocator<rksvm_node>().allocate(lth);
memcpy(copy_x->x, dest, (size_t)sizeof(rksvm_node)*lth);
sallocator<char>().deallocate(recv_message, msg_size);
//push an item to the job_queue who has the smallest number of items.
//In doing so, the dynamic loading balancing can be achieved.
int smallest_items_thread_id = 0;
auto smallest_items = job_queues[0].unsafe_size();
for(int i=1;i<thread_count;i++)
{
auto tmp = job_queues[i].unsafe_size();
if(tmp < smallest_items)
{
smallest_items_thread_id = i;
smallest_items = tmp;
}
}
job_queues[smallest_items_thread_id].push(copy_x);
recvd_count++;
}
}
return;
};///}}}
//notice that tht above functions are important to our program
//create some functional threads
std::vector<std::thread> computers;
std::thread *sender = nullptr;
std::thread *receiver = nullptr;
for (int i=0; i < thread_count; i++){
computers.push_back(std::thread(computer_fun, i));
}
if(nr_ranks>1)
{
sender = new std::thread(sender_fun);
receiver = new std::thread(receiver_fun);
}
//wait until data loading and initialization
//the main thread is used to test the results
while(count_setup_threads < thread_count){
std::this_thread::yield();
}
if(current_rank==0)
{
printf("Start to compute kernel matrix!\n");
fflush(stdout);
}
//test the time used to compute Q
tbb::tick_count start_time = tbb::tick_count::now();
while(true)
{
if(nr_ranks==1)
{
if(computed_data_nodes == global_l)
break;
}
else
{
if((computed_data_nodes==global_l)&&
(sended_count==nr_send[current_rank])&&
(recvd_count==nr_recv[current_rank]))
break;
}
}
MPI_Barrier(MPI_COMM_WORLD);//sychronization
double elapsed_seconds = (tbb::tick_count::now() - start_time).seconds();
if(current_rank==0)
{
printf("Computing Q has done!, the elapsed time is %f secs\n", elapsed_seconds);
fflush(stdout);
}
callocator<con_queue>().deallocate(job_queues, thread_count);
for(auto &th: computers)
th.join();
if(nr_ranks > 1)
{
sender->join();
receiver->join();
delete sender;
delete receiver;
}
return;
}
static PetscErrorCode PetscCommBuildTwoSidedFReq_Ibarrier(MPI_Comm comm,PetscMPIInt count,MPI_Datatype dtype,PetscMPIInt nto,const PetscMPIInt *toranks,const void *todata,
PetscMPIInt *nfrom,PetscMPIInt **fromranks,void *fromdata,PetscMPIInt ntags,MPI_Request **toreqs,MPI_Request **fromreqs,
PetscErrorCode (*send)(MPI_Comm,const PetscMPIInt[],PetscMPIInt,PetscMPIInt,void*,MPI_Request[],void*),
PetscErrorCode (*recv)(MPI_Comm,const PetscMPIInt[],PetscMPIInt,void*,MPI_Request[],void*),void *ctx)
{
PetscErrorCode ierr;
PetscMPIInt nrecvs,tag,*tags,done,i;
MPI_Aint lb,unitbytes;
char *tdata;
MPI_Request *sendreqs,*usendreqs,*req,barrier;
PetscSegBuffer segrank,segdata,segreq;
PetscBool barrier_started;
PetscFunctionBegin;
ierr = PetscCommDuplicate(comm,&comm,&tag);CHKERRQ(ierr);
ierr = PetscMalloc1(ntags,&tags);CHKERRQ(ierr);
for (i=0; i<ntags; i++) {
ierr = PetscCommGetNewTag(comm,&tags[i]);CHKERRQ(ierr);
}
ierr = MPI_Type_get_extent(dtype,&lb,&unitbytes);CHKERRQ(ierr);
if (lb != 0) SETERRQ1(comm,PETSC_ERR_SUP,"Datatype with nonzero lower bound %ld\n",(long)lb);
tdata = (char*)todata;
ierr = PetscMalloc1(nto,&sendreqs);CHKERRQ(ierr);
ierr = PetscMalloc1(nto*ntags,&usendreqs);CHKERRQ(ierr);
/* Post synchronous sends */
for (i=0; i<nto; i++) {
ierr = MPI_Issend((void*)(tdata+count*unitbytes*i),count,dtype,toranks[i],tag,comm,sendreqs+i);CHKERRQ(ierr);
}
/* Post actual payloads. These are typically larger messages. Hopefully sending these later does not slow down the
* synchronous messages above. */
for (i=0; i<nto; i++) {
PetscMPIInt k;
for (k=0; k<ntags; k++) usendreqs[i*ntags+k] = MPI_REQUEST_NULL;
ierr = (*send)(comm,tags,i,toranks[i],tdata+count*unitbytes*i,usendreqs+i*ntags,ctx);CHKERRQ(ierr);
}
ierr = PetscSegBufferCreate(sizeof(PetscMPIInt),4,&segrank);CHKERRQ(ierr);
ierr = PetscSegBufferCreate(unitbytes,4*count,&segdata);CHKERRQ(ierr);
ierr = PetscSegBufferCreate(sizeof(MPI_Request),4,&segreq);CHKERRQ(ierr);
nrecvs = 0;
barrier = MPI_REQUEST_NULL;
/* MPICH-3.2 sometimes does not create a request in some "optimized" cases. This is arguably a standard violation,
* but we need to work around it. */
barrier_started = PETSC_FALSE;
for (done=0; !done; ) {
PetscMPIInt flag;
MPI_Status status;
ierr = MPI_Iprobe(MPI_ANY_SOURCE,tag,comm,&flag,&status);CHKERRQ(ierr);
if (flag) { /* incoming message */
PetscMPIInt *recvrank,k;
void *buf;
ierr = PetscSegBufferGet(segrank,1,&recvrank);CHKERRQ(ierr);
ierr = PetscSegBufferGet(segdata,count,&buf);CHKERRQ(ierr);
*recvrank = status.MPI_SOURCE;
ierr = MPI_Recv(buf,count,dtype,status.MPI_SOURCE,tag,comm,MPI_STATUS_IGNORE);CHKERRQ(ierr);
ierr = PetscSegBufferGet(segreq,ntags,&req);CHKERRQ(ierr);
for (k=0; k<ntags; k++) req[k] = MPI_REQUEST_NULL;
ierr = (*recv)(comm,tags,status.MPI_SOURCE,buf,req,ctx);CHKERRQ(ierr);
nrecvs++;
}
if (!barrier_started) {
PetscMPIInt sent,nsends;
ierr = PetscMPIIntCast(nto,&nsends);CHKERRQ(ierr);
ierr = MPI_Testall(nsends,sendreqs,&sent,MPI_STATUSES_IGNORE);CHKERRQ(ierr);
if (sent) {
#if defined(PETSC_HAVE_MPI_IBARRIER)
ierr = MPI_Ibarrier(comm,&barrier);CHKERRQ(ierr);
#elif defined(PETSC_HAVE_MPIX_IBARRIER)
ierr = MPIX_Ibarrier(comm,&barrier);CHKERRQ(ierr);
#endif
barrier_started = PETSC_TRUE;
}
} else {
ierr = MPI_Test(&barrier,&done,MPI_STATUS_IGNORE);CHKERRQ(ierr);
}
}
*nfrom = nrecvs;
ierr = PetscSegBufferExtractAlloc(segrank,fromranks);CHKERRQ(ierr);
ierr = PetscSegBufferDestroy(&segrank);CHKERRQ(ierr);
ierr = PetscSegBufferExtractAlloc(segdata,fromdata);CHKERRQ(ierr);
ierr = PetscSegBufferDestroy(&segdata);CHKERRQ(ierr);
*toreqs = usendreqs;
ierr = PetscSegBufferExtractAlloc(segreq,fromreqs);CHKERRQ(ierr);
ierr = PetscSegBufferDestroy(&segreq);CHKERRQ(ierr);
ierr = PetscFree(sendreqs);CHKERRQ(ierr);
ierr = PetscFree(tags);CHKERRQ(ierr);
ierr = PetscCommDestroy(&comm);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
int main(int argc, char *argv[])
{
initTracer();
int i, p;
int rank, size;
scene *curscene;
int res;
TGAFILE* tga = NULL;
vector *scanline;
MPI_Status status;
MPI_Init(&argc, &argv);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &size);
if (rank == 0) {
curscene = sceneLoad();
tga = openTGA(curscene);
} else
curscene = calloc(1, sizeof(scene));
double start, end;
start = MPI_Wtime();
MPI_Bcast((void*)curscene, sizeof(scene) / sizeof(float), MPI_FLOAT, 0, MPI_COMM_WORLD);
res = sceneResolution(curscene);
scanline = calloc(res, sizeof(vector));
if (rank == 0) {
int lastline = 0;
int waiting = res;
for (lastline = 0; lastline < size - 1; lastline++) {
MPI_Send(&lastline, 1, MPI_INT, lastline + 1, 0, MPI_COMM_WORLD);
}
int flag;
while(1) {
if (waiting == 0)
break;
MPI_Iprobe(MPI_ANY_SOURCE, MPI_ANY_TAG, MPI_COMM_WORLD, &flag, &status);
if (flag) {
MPI_Recv(scanline, (sizeof(vector) / sizeof(float)) * res, MPI_FLOAT, status.MPI_SOURCE, status.MPI_TAG, MPI_COMM_WORLD, &status);
if (lastline < res) {
MPI_Send(&lastline, 1, MPI_INT, status.MPI_SOURCE, 0, MPI_COMM_WORLD);
lastline++;
}
for (i = 0; i < res; i++)
writeTGAColor(tga, curscene, status.MPI_TAG, i, scanline[i].x, scanline[i].y, scanline[i].z);
waiting--;
}
}
flag = -1;
for (p = 1; p < size; p++) {
MPI_Send(&flag, 1, MPI_INT, p, 0, MPI_COMM_WORLD);
}
} else {
int line;
while(1) {
MPI_Recv(&line, 1, MPI_INT, 0, 0, MPI_COMM_WORLD, &status);
if (line < 0) {
break;
}
printf("Process %d scanning horizontal line %d.\n", rank, line);
for (i = 0; i < res; i++)
scanline[i] = pixelraytrace(curscene, line, i);
MPI_Send(scanline, (sizeof(vector) / sizeof(float)) * res, MPI_FLOAT, 0, line, MPI_COMM_WORLD);
}
}
free(curscene);
end = MPI_Wtime();
if (rank == 0)
printf("Entire process took %.4f seconds.\n", end - start);
MPI_Finalize();
return 0;
}
static PetscErrorCode PetscCommBuildTwoSided_Ibarrier(MPI_Comm comm,PetscMPIInt count,MPI_Datatype dtype,PetscMPIInt nto,const PetscMPIInt *toranks,const void *todata,PetscMPIInt *nfrom,PetscMPIInt **fromranks,void *fromdata)
{
PetscErrorCode ierr;
PetscMPIInt nrecvs,tag,done,i;
MPI_Aint lb,unitbytes;
char *tdata;
MPI_Request *sendreqs,barrier;
PetscSegBuffer segrank,segdata;
PetscBool barrier_started;
PetscFunctionBegin;
ierr = PetscCommDuplicate(comm,&comm,&tag);CHKERRQ(ierr);
ierr = MPI_Type_get_extent(dtype,&lb,&unitbytes);CHKERRQ(ierr);
if (lb != 0) SETERRQ1(comm,PETSC_ERR_SUP,"Datatype with nonzero lower bound %ld\n",(long)lb);
tdata = (char*)todata;
ierr = PetscMalloc1(nto,&sendreqs);CHKERRQ(ierr);
for (i=0; i<nto; i++) {
ierr = MPI_Issend((void*)(tdata+count*unitbytes*i),count,dtype,toranks[i],tag,comm,sendreqs+i);CHKERRQ(ierr);
}
ierr = PetscSegBufferCreate(sizeof(PetscMPIInt),4,&segrank);CHKERRQ(ierr);
ierr = PetscSegBufferCreate(unitbytes,4*count,&segdata);CHKERRQ(ierr);
nrecvs = 0;
barrier = MPI_REQUEST_NULL;
/* MPICH-3.2 sometimes does not create a request in some "optimized" cases. This is arguably a standard violation,
* but we need to work around it. */
barrier_started = PETSC_FALSE;
for (done=0; !done; ) {
PetscMPIInt flag;
MPI_Status status;
ierr = MPI_Iprobe(MPI_ANY_SOURCE,tag,comm,&flag,&status);CHKERRQ(ierr);
if (flag) { /* incoming message */
PetscMPIInt *recvrank;
void *buf;
ierr = PetscSegBufferGet(segrank,1,&recvrank);CHKERRQ(ierr);
ierr = PetscSegBufferGet(segdata,count,&buf);CHKERRQ(ierr);
*recvrank = status.MPI_SOURCE;
ierr = MPI_Recv(buf,count,dtype,status.MPI_SOURCE,tag,comm,MPI_STATUS_IGNORE);CHKERRQ(ierr);
nrecvs++;
}
if (!barrier_started) {
PetscMPIInt sent,nsends;
ierr = PetscMPIIntCast(nto,&nsends);CHKERRQ(ierr);
ierr = MPI_Testall(nsends,sendreqs,&sent,MPI_STATUSES_IGNORE);CHKERRQ(ierr);
if (sent) {
#if defined(PETSC_HAVE_MPI_IBARRIER)
ierr = MPI_Ibarrier(comm,&barrier);CHKERRQ(ierr);
#elif defined(PETSC_HAVE_MPIX_IBARRIER)
ierr = MPIX_Ibarrier(comm,&barrier);CHKERRQ(ierr);
#endif
barrier_started = PETSC_TRUE;
ierr = PetscFree(sendreqs);CHKERRQ(ierr);
}
} else {
ierr = MPI_Test(&barrier,&done,MPI_STATUS_IGNORE);CHKERRQ(ierr);
}
}
*nfrom = nrecvs;
ierr = PetscSegBufferExtractAlloc(segrank,fromranks);CHKERRQ(ierr);
ierr = PetscSegBufferDestroy(&segrank);CHKERRQ(ierr);
ierr = PetscSegBufferExtractAlloc(segdata,fromdata);CHKERRQ(ierr);
ierr = PetscSegBufferDestroy(&segdata);CHKERRQ(ierr);
ierr = PetscCommDestroy(&comm);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
void declareBindings (void)
{
/* === Point-to-point === */
void* buf;
int count;
MPI_Datatype datatype;
int dest;
int tag;
MPI_Comm comm;
MPI_Send (buf, count, datatype, dest, tag, comm); // L12
int source;
MPI_Status status;
MPI_Recv (buf, count, datatype, source, tag, comm, &status); // L15
MPI_Get_count (&status, datatype, &count);
MPI_Bsend (buf, count, datatype, dest, tag, comm);
MPI_Ssend (buf, count, datatype, dest, tag, comm);
MPI_Rsend (buf, count, datatype, dest, tag, comm);
void* buffer;
int size;
MPI_Buffer_attach (buffer, size); // L22
MPI_Buffer_detach (buffer, &size);
MPI_Request request;
MPI_Isend (buf, count, datatype, dest, tag, comm, &request); // L25
MPI_Ibsend (buf, count, datatype, dest, tag, comm, &request);
MPI_Issend (buf, count, datatype, dest, tag, comm, &request);
MPI_Irsend (buf, count, datatype, dest, tag, comm, &request);
MPI_Irecv (buf, count, datatype, source, tag, comm, &request);
MPI_Wait (&request, &status);
int flag;
MPI_Test (&request, &flag, &status); // L32
MPI_Request_free (&request);
MPI_Request* array_of_requests;
int index;
MPI_Waitany (count, array_of_requests, &index, &status); // L36
MPI_Testany (count, array_of_requests, &index, &flag, &status);
MPI_Status* array_of_statuses;
MPI_Waitall (count, array_of_requests, array_of_statuses); // L39
MPI_Testall (count, array_of_requests, &flag, array_of_statuses);
int incount;
int outcount;
int* array_of_indices;
MPI_Waitsome (incount, array_of_requests, &outcount, array_of_indices,
array_of_statuses); // L44--45
MPI_Testsome (incount, array_of_requests, &outcount, array_of_indices,
array_of_statuses); // L46--47
MPI_Iprobe (source, tag, comm, &flag, &status); // L48
MPI_Probe (source, tag, comm, &status);
MPI_Cancel (&request);
MPI_Test_cancelled (&status, &flag);
MPI_Send_init (buf, count, datatype, dest, tag, comm, &request);
MPI_Bsend_init (buf, count, datatype, dest, tag, comm, &request);
MPI_Ssend_init (buf, count, datatype, dest, tag, comm, &request);
MPI_Rsend_init (buf, count, datatype, dest, tag, comm, &request);
MPI_Recv_init (buf, count, datatype, source, tag, comm, &request);
MPI_Start (&request);
MPI_Startall (count, array_of_requests);
void* sendbuf;
int sendcount;
MPI_Datatype sendtype;
int sendtag;
void* recvbuf;
int recvcount;
MPI_Datatype recvtype;
MPI_Datatype recvtag;
MPI_Sendrecv (sendbuf, sendcount, sendtype, dest, sendtag,
recvbuf, recvcount, recvtype, source, recvtag,
comm, &status); // L67--69
MPI_Sendrecv_replace (buf, count, datatype, dest, sendtag, source, recvtag,
comm, &status); // L70--71
MPI_Datatype oldtype;
MPI_Datatype newtype;
MPI_Type_contiguous (count, oldtype, &newtype); // L74
int blocklength;
{
int stride;
MPI_Type_vector (count, blocklength, stride, oldtype, &newtype); // L78
}
{
MPI_Aint stride;
MPI_Type_hvector (count, blocklength, stride, oldtype, &newtype); // L82
}
int* array_of_blocklengths;
{
int* array_of_displacements;
MPI_Type_indexed (count, array_of_blocklengths, array_of_displacements,
oldtype, &newtype); // L87--88
}
{
MPI_Aint* array_of_displacements;
MPI_Type_hindexed (count, array_of_blocklengths, array_of_displacements,
oldtype, &newtype); // L92--93
MPI_Datatype* array_of_types;
MPI_Type_struct (count, array_of_blocklengths, array_of_displacements,
array_of_types, &newtype); // L95--96
}
void* location;
MPI_Aint address;
MPI_Address (location, &address); // L100
MPI_Aint extent;
MPI_Type_extent (datatype, &extent); // L102
MPI_Type_size (datatype, &size);
MPI_Aint displacement;
MPI_Type_lb (datatype, &displacement); // L105
MPI_Type_ub (datatype, &displacement);
MPI_Type_commit (&datatype);
MPI_Type_free (&datatype);
MPI_Get_elements (&status, datatype, &count);
void* inbuf;
void* outbuf;
int outsize;
int position;
MPI_Pack (inbuf, incount, datatype, outbuf, outsize, &position, comm); // L114
int insize;
MPI_Unpack (inbuf, insize, &position, outbuf, outcount, datatype,
comm); // L116--117
MPI_Pack_size (incount, datatype, comm, &size);
/* === Collectives === */
MPI_Barrier (comm); // L121
int root;
MPI_Bcast (buffer, count, datatype, root, comm); // L123
MPI_Gather (sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype,
root, comm); // L124--125
int* recvcounts;
int* displs;
MPI_Gatherv (sendbuf, sendcount, sendtype,
recvbuf, recvcounts, displs, recvtype,
root, comm); // L128--130
MPI_Scatter (sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype,
root, comm); // L131--132
int* sendcounts;
MPI_Scatterv (sendbuf, sendcounts, displs, sendtype,
recvbuf, recvcount, recvtype, root, comm); // L134--135
MPI_Allgather (sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype,
comm); // L136--137
MPI_Allgatherv (sendbuf, sendcount, sendtype,
recvbuf, recvcounts, displs, recvtype,
comm); // L138--140
MPI_Alltoall (sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype,
comm); // L141--142
int* sdispls;
int* rdispls;
MPI_Alltoallv (sendbuf, sendcounts, sdispls, sendtype,
recvbuf, recvcounts, rdispls, recvtype,
comm); // L145--147
MPI_Op op;
MPI_Reduce (sendbuf, recvbuf, count, datatype, op, root, comm); // L149
#if 0
MPI_User_function function;
int commute;
MPI_Op_create (function, commute, &op); // L153
#endif
MPI_Op_free (&op); // L155
MPI_Allreduce (sendbuf, recvbuf, count, datatype, op, comm);
MPI_Reduce_scatter (sendbuf, recvbuf, recvcounts, datatype, op, comm);
MPI_Scan (sendbuf, recvbuf, count, datatype, op, comm);
/* === Groups, contexts, and communicators === */
MPI_Group group;
MPI_Group_size (group, &size); // L162
int rank;
MPI_Group_rank (group, &rank); // L164
MPI_Group group1;
int n;
int* ranks1;
MPI_Group group2;
int* ranks2;
MPI_Group_translate_ranks (group1, n, ranks1, group2, ranks2); // L170
int result;
MPI_Group_compare (group1, group2, &result); // L172
MPI_Group newgroup;
MPI_Group_union (group1, group2, &newgroup); // L174
MPI_Group_intersection (group1, group2, &newgroup);
MPI_Group_difference (group1, group2, &newgroup);
int* ranks;
MPI_Group_incl (group, n, ranks, &newgroup); // L178
MPI_Group_excl (group, n, ranks, &newgroup);
extern int ranges[][3];
MPI_Group_range_incl (group, n, ranges, &newgroup); // L181
MPI_Group_range_excl (group, n, ranges, &newgroup);
MPI_Group_free (&group);
MPI_Comm_size (comm, &size);
MPI_Comm_rank (comm, &rank);
MPI_Comm comm1;
MPI_Comm comm2;
MPI_Comm_compare (comm1, comm2, &result);
MPI_Comm newcomm;
MPI_Comm_dup (comm, &newcomm);
MPI_Comm_create (comm, group, &newcomm);
int color;
int key;
MPI_Comm_split (comm, color, key, &newcomm); // L194
MPI_Comm_free (&comm);
MPI_Comm_test_inter (comm, &flag);
MPI_Comm_remote_size (comm, &size);
MPI_Comm_remote_group (comm, &group);
MPI_Comm local_comm;
int local_leader;
MPI_Comm peer_comm;
int remote_leader;
MPI_Comm newintercomm;
MPI_Intercomm_create (local_comm, local_leader, peer_comm, remote_leader, tag,
&newintercomm); // L204--205
MPI_Comm intercomm;
MPI_Comm newintracomm;
int high;
MPI_Intercomm_merge (intercomm, high, &newintracomm); // L209
int keyval;
#if 0
MPI_Copy_function copy_fn;
MPI_Delete_function delete_fn;
void* extra_state;
MPI_Keyval_create (copy_fn, delete_fn, &keyval, extra_state); // L215
#endif
MPI_Keyval_free (&keyval); // L217
void* attribute_val;
MPI_Attr_put (comm, keyval, attribute_val); // L219
MPI_Attr_get (comm, keyval, attribute_val, &flag);
MPI_Attr_delete (comm, keyval);
/* === Environmental inquiry === */
char* name;
int resultlen;
MPI_Get_processor_name (name, &resultlen); // L226
MPI_Errhandler errhandler;
#if 0
MPI_Handler_function function;
MPI_Errhandler_create (function, &errhandler); // L230
#endif
MPI_Errhandler_set (comm, errhandler); // L232
MPI_Errhandler_get (comm, &errhandler);
MPI_Errhandler_free (&errhandler);
int errorcode;
char* string;
MPI_Error_string (errorcode, string, &resultlen); // L237
int errorclass;
MPI_Error_class (errorcode, &errorclass); // L239
MPI_Wtime ();
MPI_Wtick ();
int argc;
char** argv;
MPI_Init (&argc, &argv); // L244
MPI_Finalize ();
MPI_Initialized (&flag);
MPI_Abort (comm, errorcode);
}
int main(int argc, char *argv[]) {
int rank, nprocs, x, y;
Window win; // initialization for a window
GC gc; // graphics context
Display *display = NULL;
Colormap screenColourmap;
XColor colours[15];
unsigned int width = X_RESN, height = Y_RESN; /* window size */
clock_t start, end, elapsed;
int pixelToDraw[3];
MPI_Init(&argc, &argv);
MPI_Comm_size(MPI_COMM_WORLD, &nprocs);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
ComplexNumber c;
ComplexNumber z;
int depth = 1000;
int cSize = 4;
double cX = -2;
double cY = -2;
double cMinR = cX;
double cMinI = cY;
double cMaxR = cMinR + cSize;
double cMaxI = cMinI + cSize;
double increaseX = (cMaxR - cMinR) / X_RESN;
double increaseY = (cMaxI - cMinI) / Y_RESN;
int drawable[X_RESN][Y_RESN];
if(rank==0)
{
display = x11setup(&win, &gc, width, height);
screenColourmap = DefaultColormap(display, DefaultScreen(display));
setupColours(display, colours, screenColourmap);
int probeFlag = 0;
MPI_Status status;
int count = 0;
int start = 1;
for (int i = 1; i < nprocs; i++) {
MPI_Send(&start, 1, MPI_INT, i, 98, MPI_COMM_WORLD);
}
for (int y = 0; y < Y_RESN; y++) {
for (int x = 0; x < X_RESN; x++) {
while (!probeFlag) {
MPI_Iprobe(MPI_ANY_SOURCE, 99, MPI_COMM_WORLD, &probeFlag, &status);
if (probeFlag) {
int coords[2] = {x, y};
int slave = status.MPI_SOURCE;
MPI_Recv(&pixelToDraw, 3, MPI_INT, MPI_ANY_SOURCE, 99, MPI_COMM_WORLD, MPI_STATUS_IGNORE);
MPI_Send(&coords, 2, MPI_INT, slave, 99, MPI_COMM_WORLD);
drawable[pixelToDraw[0]][pixelToDraw[1]] = pixelToDraw[2];
count++;
} else {
MPI_Iprobe(MPI_ANY_SOURCE, 98, MPI_COMM_WORLD, &probeFlag, &status);
if (probeFlag) {
int slave = status.MPI_SOURCE;
int coords[2] = {x, y};
MPI_Recv(&start, 1, MPI_INT, MPI_ANY_SOURCE, 98, MPI_COMM_WORLD, MPI_STATUS_IGNORE);
MPI_Send(&coords, 2, MPI_INT, slave, 99, MPI_COMM_WORLD);
}
}
}
probeFlag = 0;
}
}
MPI_Recv(&pixelToDraw, 3, MPI_INT, MPI_ANY_SOURCE, 99, MPI_COMM_WORLD, MPI_STATUS_IGNORE);
drawable[pixelToDraw[0]][pixelToDraw[1]] = pixelToDraw[2];
for (int y = 0; y < Y_RESN; y++) {
for (int x = 0; x < X_RESN; x++) {
drawPoint(display, win, gc, colours, x, y, depth, drawable[x][y]);
}
}
} else {
int flag = 1;
int coords[2];
MPI_Recv(&start, 1, MPI_INT, 0, 98, MPI_COMM_WORLD, MPI_STATUS_IGNORE);
MPI_Send(&start, 1, MPI_INT, 0, 98, MPI_COMM_WORLD);
while (flag) {
MPI_Recv(&coords, 2, MPI_INT, 0, 99, MPI_COMM_WORLD, MPI_STATUS_IGNORE);
c.r = coords[0] * increaseX - 2;
c.i = coords[1] * increaseY - 2;
resetComplexNumber(&z);
int limit = seriesDiverges(depth, &z, &c);
pixelToDraw[0] = coords[0];
pixelToDraw[1] = coords[1];
pixelToDraw[2] = limit;
MPI_Send(&pixelToDraw, 3, MPI_INT, 0, 99, MPI_COMM_WORLD);
}
}
// main loop
int running = 1; // loop variable
start = clock();
while(running) {
// checks to see if there have been any events,
// will exit the main loop if any key is pressed
if(rank==0) {
if(XPending(display)) {
XEvent ev;
XNextEvent(display, &ev);
switch(ev.type) {
case KeyPress:
running = 0;
break;
}
}
}
end = clock();
elapsed = end - start;
// only update the display if > 1 millisecond has passed since the last update
if(elapsed / (CLOCKS_PER_SEC/1000) > 1 && rank==0) {
//XClearWindow(display, win);
start = end;
XFlush(display);
}
}
if(rank==0 && display) {
XCloseDisplay(display); // close the display window
printf("Display closed\n");
} else {
MPI_Finalize();
return 0;
}
MPI_Finalize();
return 0;
}
void manager_initialize(WSET *ws, int NUM_GROUP, int stages,
int numbprocs, char filename[50])
{
int i,slv,eqnid=0,eqnpack[NUM_GROUP],new_eqnpack[NUM_GROUP],
quotient,remain,eqnsize,flag,quit_collect=0,cnt_recv;
MPI_Status status;
/* Initialization phase: sending eqn. index to workers */
if(NUM_GROUP<=numbprocs-1) /* worker with id<=NUM_GROUP will get one eqn. */
{
if(v>3) printf("NUM_GROUP<=numbprocs-1\n");
for(slv=1;slv<=NUM_GROUP;slv++)
{
eqnid = slv;
MPI_Send(&eqnid,1,MPI_INT,slv,EQN_TAG,MPI_COMM_WORLD);
if(v>3 ) printf("manager sent eqn. index %d to %d\n", eqnid, slv);
}
}
else /* NUM_GROUP>numbprocs-1 */
{
if(v>3) printf("NUM_GROUP>numbprocs-1\n");
quotient = NUM_GROUP/(numbprocs-1);
remain = (int)fmod(NUM_GROUP,numbprocs-1);
if(v>3) printf("quotient=%d,remain=%d\n",quotient,remain);
eqnid = 1;
for(slv=1;slv<=remain;slv++)
{
for(i=1;i<=quotient+1;i++)
{
MPI_Send(&eqnid,1,MPI_INT,slv,EQN_TAG,MPI_COMM_WORLD);
eqnid++;
}
}
for(slv=remain+1;slv<=numbprocs-1;slv++)
{
for(i=1;i<=quotient;i++)
{
MPI_Send(&eqnid,1,MPI_INT,slv,EQN_TAG,MPI_COMM_WORLD);
eqnid++;
}
}
if(v>3) printf("the last eqnid=%d\n", eqnid-1);
assert((eqnid-1)<=NUM_GROUP);
}
/* send 0 to everyone */
if(v>3) printf("manager sent 0 to everyone.\n");
eqnid = 0;
for(slv=1;slv<=numbprocs-1;slv++)
MPI_Send(&eqnid,1,MPI_INT,slv,EQN_TAG,MPI_COMM_WORLD);
/* initial collect */
cnt_recv = 0;
while(!quit_collect)
{
MPI_Iprobe(MPI_ANY_SOURCE,EQN_TAG,MPI_COMM_WORLD,&flag,&status);
while(flag) /* while flag -- pending recv */
{
slv = status.MPI_SOURCE; /* which slave sent this JOB back */
MPI_Get_count(&status, MPI_INT, &eqnsize);
MPI_Recv(new_eqnpack,eqnsize,MPI_INT,slv,EQN_TAG,MPI_COMM_WORLD,&status);
if(v>3)
{
printf("manager recv returned eqn index from node %d with eqnsize=%d\n",
slv, eqnsize); fflush;
for(i=0;i<eqnsize;i++)
printf("%d\n", new_eqnpack[i]);
fflush;
}
cnt_recv++;
if(cnt_recv==numbprocs-1) /* all workers have sent back */
{ quit_collect=1; break;}
MPI_Iprobe(MPI_ANY_SOURCE,EQN_TAG,MPI_COMM_WORLD,&flag,&status);
}
}
printf("Initialization phase is DONE!\n"); fflush;
/* End of Initialization phase */
/* mainloop initialization */
mainloop_initialize(ws, NUM_GROUP, stages, numbprocs, filename);
}
void newflow_inbox(file_object_vector &inbox_file_objects, SeqGraph *resident_sgraph, SplitBuckets *buckets,
uintptr_t &total_bucket_nodes)
{
#ifdef VELOUR_TBB
tbb::tick_count::interval_t time_inboxload, time_relatedcomponents, time_insertnodes;
tbb::tick_count::interval_t time_removetips, time_concatenate;
tbb::tick_count::interval_t time_slicing, time_splitting, time_cleanup;
#endif
if (!inbox_file_objects.empty()) { // no inbox for the first partition
// incrementally load components from the inbox files
for (file_object_vector::iterator itr=inbox_file_objects.begin() ; itr != inbox_file_objects.end() ; ++itr) {
if (itr->filetype != BUCKET) {
fprintf(stderr, "ERROR: Cannot use %s file %s as flow input. Exiting...\n",
FILE_TYPES[itr->filetype], itr->filename);
exit(EXIT_FAILURE);
}
#ifdef VELOUR_MPI
unsigned inbox_partition = itr->fileindex;
#endif // VELOUR_MPI
printf(" flowing inbox: %s\n", itr->filename); fflush(stdout);
uintptr_t round = 1;
size_t file_offset = 0;
#ifdef VELOUR_MPI
int producer_finished = 0;
size_t safe_length = 0;
do {
//if (safe_length < p__MPI_SAFE_LENGTH_ARRAY[inbox_partition])
// printf("DBG: Partition %u read new safe_length %lli\n", g__PARTITION_INDEX, safe_length); fflush(stdout);
size_t new_safe_length = * static_cast<volatile size_t *>(&p__MPI_SAFE_LENGTH_ARRAY[inbox_partition]);
if (new_safe_length == safe_length) {
MPI_Status mpi_status;
MPI_Iprobe(PART_TO_RANK(inbox_partition), 42, MPI_COMM_WORLD, &producer_finished, &mpi_status);
// TODO: do MPI_Recv() on the message()
continue;
} else {
safe_length = new_safe_length;
}
#endif // VELOUR_MPI
int filedes = open(itr->filename, O_RDONLY);
if (filedes == -1) {
fprintf(stderr, "ERROR: failed to open file: %s\n", itr->filename);
perror("REASON: ");
exit(EXIT_FAILURE);
}
struct stat file_stat;
if (fstat(filedes, &file_stat) != 0) {
fprintf(stderr,"ERROR: failed to stat file: %s\n", itr->filename);
exit(EXIT_FAILURE);
}
size_t file_length = file_stat.st_size;
//itr->length = file_stat.st_size;
#ifdef VELOUR_MPI
file_length = min(file_length, safe_length);
#endif
if (file_length == 0) { // otherwise, mmap will fail
if (close(filedes) == -1) {
fprintf(stderr, "ERROR: failed to close file: %s\n", itr->filename);
perror("REASON: ");
exit(EXIT_FAILURE);
}
#ifdef VELOUR_MPI
MPI_Status mpi_status;
MPI_Iprobe(PART_TO_RANK(inbox_partition), 42, MPI_COMM_WORLD, &producer_finished, &mpi_status);
if (producer_finished && safe_length == * static_cast<volatile size_t *>(&p__MPI_SAFE_LENGTH_ARRAY[inbox_partition])) {
// TODO: do MPI_Recv() on the message()
//printf("DBG: Partition %u producer finished break\n", g__PARTITION_INDEX); fflush(stdout);
break; // break out of do loop to start next file
} else {
//printf("DBG: Partition %u not finished continue\n", g__PARTITION_INDEX); fflush(stdout);
continue; // continue do loop
}
#else
continue; // continue to start next file
#endif // VELOUR_MPI
}
char *file_mmap = static_cast<char *>( mmap( 0, file_length, PROT_READ, MAP_PRIVATE, filedes, 0) );
if (file_mmap == reinterpret_cast<char *>(-1)) {
fprintf(stderr, "ERROR: failed to mmap file: %s\n", itr->filename);
perror("REASON: ");
exit(EXIT_FAILURE);
}
flow_nodelist_t flowlist;
#ifdef VELOUR_TBB
flowlist.reserve(1000000); // XXX: constant, should relate to value below
#endif
while (file_offset < file_length) {
uintptr_t round_inbox_nodes = 0;
uintptr_t round_import_nodes = 0;
#ifdef VELOUR_TBB
tbb::tick_count time0 = tbb::tick_count::now();
#endif
SeqNode *inbox_chain = NULL;
uintptr_t inbox_chain_length = 0;
uintptr_t min_inbox_chain_length = max(100000, resident_sgraph->node_count >> 5); // about 3%
//
// CHOOSE WHICH MODE: PERFORMANCE or MINIMUM PEAK FOOTPRINT MODE FOR PUBLICATIONS
//
uintptr_t max_inbox_chain_length;
if (g__MINIMIZE_FOOTPRINT) {
max_inbox_chain_length = min_inbox_chain_length;
} else {
max_inbox_chain_length = max(min_inbox_chain_length, (g__NODE_ALLOCATORS->GetMaxSafeAllocation() / 100)); // XXX: max average seqnode size constant
}
while (file_offset < file_length && inbox_chain_length < max_inbox_chain_length) {
size_t component_size;
uintptr_t component_length = sg_load_mmap_stream_component((file_mmap+file_offset), &component_size, &inbox_chain);
file_offset += component_size;
inbox_chain_length += component_length;
}
total_bucket_nodes += inbox_chain_length;
round_inbox_nodes += inbox_chain_length;
p__peakSeqNodes = max(p__peakSeqNodes, resident_sgraph->node_count + inbox_chain_length);
p__peakSeqLiveMemory = max(p__peakSeqLiveMemory, g__SEQNODE_ALLOCATOR->GetSeqLiveMemory());
#ifdef VELOUR_TBB
tbb::tick_count time1 = tbb::tick_count::now();
#endif
#ifdef VELOUR_TBB
uintptr_t nodes_memoized = sg_parallel_chain_memoize_related_components(inbox_chain, resident_sgraph, &flowlist);
//uintptr_t nodes_memoized = sg_serial_chain_memoize_related_components(inbox_chain, resident_sgraph, &flowlist);
#else
uintptr_t nodes_memoized = sg_serial_chain_memoize_related_components(inbox_chain, resident_sgraph, &flowlist);
#endif
round_import_nodes += nodes_memoized;
#ifdef VELOUR_TBB
tbb::tick_count time2 = tbb::tick_count::now();
#endif
while (inbox_chain != NULL) {
SeqNode *node = inbox_chain;
inbox_chain = inbox_chain->head_next;
node->head_next = NULL; // XXX
flowlist.push_back(node);
resident_sgraph->insertNodeAndUpdateColors(node);
}
#ifdef VELOUR_TBB
tbb::tick_count time3 = tbb::tick_count::now();
time_inboxload += time1 - time0;
time_relatedcomponents += time2 - time1;
time_insertnodes += time3 - time2;
#endif
//printf("(flow) inbox component nodes: %"PRIuPTR"\n", round_inbox_nodes); // XXX: remove me
//printf("(flow) inbox component related nodes: %"PRIuPTR"\n", round_import_nodes); fflush(stdout); // XXX: remove me
#ifdef VERIFY
resident_sgraph->verify(true);
#endif
#ifdef VELOUR_TBB
tbb::tick_count time4 = tbb::tick_count::now();
#endif
#ifdef VELOUR_TBB
//TODO sg_parallel_nodelist_remove_tips(resident_sgraph, true, &flowlist);
sg_nodelist_remove_tips(resident_sgraph, true, &flowlist);
#else
sg_nodelist_remove_tips(resident_sgraph, true, &flowlist);
#endif
#ifdef VELOUR_TBB
tbb::tick_count time5 = tbb::tick_count::now();
#endif
#ifdef VERIFY
resident_sgraph->verify(true);
#endif
//sg_concatenate(resident_sgraph, true);
sg_nodelist_concatenate(resident_sgraph, true, &flowlist);
#ifdef VELOUR_TBB
tbb::tick_count time6 = tbb::tick_count::now();
#endif
#ifdef VERIFY
resident_sgraph->verify(true);
#endif
// FIXME: execution time is counted with concatenation
//sg_nodelist_pop_bubbles(resident_sgraph, true, &flowlist);
#ifdef VELOUR_TBB
//tbb::tick_count time6 = tbb::tick_count::now();
#endif
#ifdef VERIFY
//resident_sgraph->verify(true);
#endif
// emit sub-components that are no longer relevant
//slice2_graph(resident_sgraph, g__PARTITION_INDEX);
if (g__SLICING) slice2_nodelist(resident_sgraph, g__PARTITION_INDEX, &flowlist);
#ifdef VELOUR_TBB
tbb::tick_count time7 = tbb::tick_count::now();
#endif
// emit components that are no longer relevant to the working graph
//buckets->split(resident_sgraph);
buckets->split_nodelist(resident_sgraph, &flowlist);
#ifdef VELOUR_TBB
tbb::tick_count time8 = tbb::tick_count::now();
#endif
// lastly, reset claim tid on nodes for future iterations
for (flow_nodelist_t::iterator it = flowlist.begin(); it != flowlist.end() ; ++it) {
SeqNode *node = *it;
//assert( node->claim_tid == 1 ); // don't assert this as not true for grown nodes and for inbox nodes
node->claim_tid = 0;
}
flowlist.clear();
#ifdef VELOUR_TBB
tbb::tick_count time9 = tbb::tick_count::now();
time_removetips += time5 - time4;
time_concatenate += time6 - time5;
time_slicing += time7 - time6;
time_splitting += time8 - time7;
time_cleanup += time9 - time8;
#endif
// initiate garbage collection if needed
if (g__NODE_ALLOCATORS->flag_gc_needed_) { g__NODE_ALLOCATORS->GarbageCollect(); }
++ round;
}
if (munmap(file_mmap, file_length) == -1) {
fprintf(stderr, "ERROR: failed to munmap file: %s\n", itr->filename);
perror("REASON: ");
exit(EXIT_FAILURE);
}
if (close(filedes) == -1) {
fprintf(stderr, "ERROR: failed to close file: %s\n", itr->filename);
perror("REASON: ");
exit(EXIT_FAILURE);
}
#ifdef VELOUR_MPI
checkdone:
MPI_Status mpi_status;
MPI_Iprobe(PART_TO_RANK(inbox_partition), 42, MPI_COMM_WORLD, &producer_finished, &mpi_status);
// TODO: do MPI_Recv() on the message()
} while (!producer_finished || safe_length != * static_cast<volatile size_t *>(&p__MPI_SAFE_LENGTH_ARRAY[inbox_partition]));
#endif // VELOUR_MPI
}
/*if( g__FULL_STATISTICS ) {
sg_stat_components(resident_sgraph, stdout);
}*/
// last, split the resident graph to outboxes // FIXME: should this be empty!?
printf("FLOW: %zu nodes (FIXME maybe dead) in resident graph at end of flowing.... slicing and splitting resident graph.\n",
resident_sgraph->node_count); fflush(stdout); // XXX: HACK
if (g__SLICING) { slice2_graph(resident_sgraph, g__PARTITION_INDEX); } // XXX: HACK HACK HACK
buckets->split(resident_sgraph);
//assert( resident_sgraph->node_count == 0 );
}
#ifdef VELOUR_TBB
tbb::tick_count::interval_t time_total = time_inboxload + time_relatedcomponents + time_insertnodes +
time_removetips + time_concatenate + time_slicing + time_splitting + time_cleanup;
printf(" flow time: %lfs inbox: %02.1lf%% related: %02.1lf%% insert: %02.1lf%%"
" remove: %02.1lf%% concat: %02.1lf%% slice: %02.1lf%% split: %02.1lf%%"
" clean: %02.1lf%%\n",
time_total.seconds(),
100.0 * (time_inboxload.seconds() / time_total.seconds()),
100.0 * (time_relatedcomponents.seconds() / time_total.seconds()),
100.0 * (time_insertnodes.seconds() / time_total.seconds()),
100.0 * (time_removetips.seconds() / time_total.seconds()),
100.0 * (time_concatenate.seconds() / time_total.seconds()),
100.0 * (time_slicing.seconds() / time_total.seconds()),
100.0 * (time_splitting.seconds() / time_total.seconds()),
100.0 * (time_cleanup.seconds() / time_total.seconds()) );
fflush(stdout);
#endif
}
void slave_initialize(int id, int NUM_GROUP, char filename[50])
{
int flag,*buffer,*newbuffer,count,i,eqn_1,dim_1,solution_1,eqn_2,
cnt_eqns,eqnid,eqn_pack[NUM_GROUP],quit_flag=0,
fail,k,n;
char outfile[80];
MPI_Status status;
if(v>2) printf("Node %d knows the filename: %s\n",id, filename);fflush;
fail = read_named_target_without_solutions ((int) strlen(filename),filename);
fail = copy_target_system_to_container();
cnt_eqns = 0;
while(!quit_flag)
{
MPI_Iprobe(0,MPI_ANY_TAG,MPI_COMM_WORLD,&flag,&status);
while(flag) /* while flag -- pending recv */
{
if(status.MPI_TAG==EQN_TAG)
{
MPI_Recv(&eqnid,1,MPI_INT,0,EQN_TAG,MPI_COMM_WORLD,&status);
if(eqnid!=0)
{
eqn_pack[cnt_eqns] = eqnid;
cnt_eqns++;
if(v>2) printf("node %d solving eqn. %d; cnt_eqns=%d \n",
id, eqnid, cnt_eqns); fflush;
/* solving eqn. */
sprintf(outfile, "%s_%d", filename, eqnid );
n = (int) strlen(outfile);
fail = hypersurface_witness_set(eqnid,n,outfile);
/* end of solving eqn. */
}
else
{
if(v>2) printf("node %d recv end signal and send back: \n", id);
/* send the index of the eqns for which the worker has computed back */
eqn_pack[cnt_eqns] = 0; /* the last number is 0 */
cnt_eqns++;
if(v>2)
{
printf("cnt_eqns=%d\n", cnt_eqns);
for(i=0;i<cnt_eqns;i++)
printf("%d\n", eqn_pack[i]);
}
MPI_Send(eqn_pack,cnt_eqns,MPI_INT,0,EQN_TAG,MPI_COMM_WORLD);
/* clean the system and solution container */
fail = solcon_clear_standard_solutions( );
if(fail>0) printf("fail to clear solution container.\n");
fail = syscon_clear_standard_system( );
if(fail>0) printf("fail to clear system container.\n");
/* end of cleaning the system and solution container */
quit_flag =1;
}
fflush;
} /* status.MPI_TAG */
MPI_Iprobe(0,MPI_ANY_TAG,MPI_COMM_WORLD,&flag,&status);
} /* flag */
} /* quit_flag */
}
int wc_mpi_iprobe(int src_id, int tag, int *flag, wc_mpistatus_t *status)
{
int rc = MPI_Iprobe(src_id, tag, MPI_COMM_WORLD, flag, status);
WC_HANDLE_MPI_ERROR(MPI_Iprobe, rc);
return (rc == MPI_SUCCESS) ? 0 : -1;
}
void send_collect(WSET *ws, int num_group, int stages, int cnt_stage,
int numbprocs, int cnt_step, IDLE_ELEMENT **ie,
LISTELEMENT **listpointer, int *TaskCount, int NUM_GROUP,
int *cnt_index, int update_flag, int n1, int n2, int cd)
{
int slv,flght,flag,*buffer,*package,count,cnt_count,
idle_procs[numbprocs],idle_pos,cnt_idle,
r,c,s_max,s_min,
deg,n=cd,m[2],send[3],label_sol,fail,i=0,j=0;
double sol1[2*n1+5], sol2[2*n2+5], ps[2*cd+5];
WSET *ws_p;
JOB *temp_j;
MPI_Status status;
if(v>3) printf("inside of send_collect\n"); fflush;
temp_j = (JOB *)((*listpointer)->dataitem);
count = temp_j->num_eqns_1 + temp_j->num_eqns_2 + 8;
if(cnt_step==2 && update_flag)
{
package = (int*) malloc(count*sizeof(int));
pack(count, temp_j, package);
}
cnt_idle = num_idle(*ie);
/* Initial JOB distribution */
while(*ie != NULL) /* there are idle processors */
{
if(*listpointer!=NULL) /* JOB queue is not empty */
{
/* pop one JOB from the queue */
*listpointer = removeitem (*listpointer);
slv = (*ie)->data;
*ie = removeslv(*ie);
/* obtain start solution & send */
if(v>3) printf("sent a job & path to node %d.\n",slv);
if(cnt_step==1)
{
if(v>3)
fail = get_next_start_product
(&i,&j,1,temp_j->num_vars_1,temp_j->num_vars_2,
temp_j->dim_1,temp_j->dim_2,
temp_j->deg_1,temp_j->deg_2,cd,sol1,sol2,ps);
else
fail = get_next_start_product
(&i,&j,0,temp_j->num_vars_1,temp_j->num_vars_2,
temp_j->dim_1,temp_j->dim_2,
temp_j->deg_1,temp_j->deg_2,cd,sol1,sol2,ps);
m[0] = n; m[1] = 1;
send[0] = slv; send[1] = m[1]; send[2] = m[0];
MPI_Send(send,3,MPI_INT,slv,SEND_SMUL,MPI_COMM_WORLD);
MPI_Send(ps,2*n+5,MPI_DOUBLE,slv,SEND_SSOL,MPI_COMM_WORLD);
}
if(cnt_step==2)
{
fail = solcon_read_next_solution(n,&m[1],ps);
if(fail>0) printf("solcon_read_next_solution fail!\n");
m[0] = n;
send[0] = slv; send[1] = m[1]; send[2] = m[0];
MPI_Send(send,3,MPI_INT,slv,SEND_SMUL,MPI_COMM_WORLD);
MPI_Send(ps,2*n+5,MPI_DOUBLE,slv,SEND_SSOL,MPI_COMM_WORLD);
}
/* end of obtaining start solution & sending */
*(TaskCount+slv-1)=*(TaskCount+slv-1)+1;
}
else
break;
}
flght = cnt_idle - num_idle(*ie);
flag = 0;
while(flght>0 || *listpointer!=NULL) /* JOB queue loop */
{
if(flght>0)
{
MPI_Iprobe(MPI_ANY_SOURCE,MPI_ANY_TAG,MPI_COMM_WORLD,&flag,&status);
while(flag) /* while flag -- pending recv */
{
if(v>3) printf("manager starting recv... \n");
slv = status.MPI_SOURCE; /* which slave sent this JOB back */
/* recv end solution */
MPI_Recv(ps,2*n+7,MPI_DOUBLE,MPI_ANY_SOURCE,SEND_TSOL,
MPI_COMM_WORLD,&status);
m[1] = (int) ps[2*n+5];
label_sol = (int) ps[2*n+6] - 1;
fail = solcon_write_next_solution_to_defined_output_file
(&label_sol,n,m[1],ps);
/* end of recv. end solution */
/* update idle processor list */
*ie = addslv(*ie, slv);
/* update corresponding cell when cnt_step==2 && update_flag is TRUE */
if(cnt_step==2 && update_flag)
{
c = cnt_stage;
r = *cnt_index;
ws_p = ws+r*(stages+1)+c;
if(ws_p->count==0) /* first returned JOB for current witness set */
{
printf("update num_eqns, source, dim, deg\n");
ws_p->num_eqns = package[0];
ws_p->source = (int*) malloc(package[0]*sizeof(int));
deg = 1;
for(cnt_count=0;cnt_count<package[0];cnt_count++)
{
ws_p->source[cnt_count] = package[cnt_count+1];
printf("ws[%d][%d].source[%d]=%d\n",r,c,cnt_count,ws_p->source[cnt_count]);
deg = (ws+(package[cnt_count+1]-1)*(stages+1))->deg*deg;
}
ws_p->dim = package[package[0]+1];
/*ws_p->sols = (int*) malloc(deg*sizeof(int));*/
ws_p->deg = deg;
}
/*ws_p->sols[ws_p->count]=ws_p->count+1; */
ws_p->count++;
if(ws_p->count==ws_p->deg) /* this witness set is complete */
{
if(ws_p->num_eqns==NUM_GROUP)
{
printf("\nrecord [%d][%d]\n", r,c);
printf("ALL DONE! aha........\n");
print_ws(ws_p);
}
} /* if: count==deg */
} /* cnt_step == 2 && update_flag */
if(*listpointer!=NULL) /* JOB queue is not empty */
{
/* pop one JOB from the queue */
*listpointer = removeitem (*listpointer);
slv = (*ie)->data;
*ie = removeslv(*ie);
if(v>3) printf("sending a job & path to node %d.\n",slv);
/* obtain start solution & send */
if(cnt_step==1)
{
if(v>3)
fail = get_next_start_product
(&i,&j,1,temp_j->num_vars_1,temp_j->num_vars_2,
temp_j->dim_1,temp_j->dim_2,
temp_j->deg_1,temp_j->deg_2,cd,sol1,sol2,ps);
else
fail = get_next_start_product
(&i,&j,0,temp_j->num_vars_1,temp_j->num_vars_2,
temp_j->dim_1,temp_j->dim_2,
temp_j->deg_1,temp_j->deg_2,cd,sol1,sol2,ps);
m[0] = n; m[1] = 1;
send[0] = slv; send[1] = m[1]; send[2] = m[0];
MPI_Send(send,3,MPI_INT,slv,SEND_SMUL,MPI_COMM_WORLD);
MPI_Send(ps,2*n+5,MPI_DOUBLE,slv,SEND_SSOL,MPI_COMM_WORLD);
}
if(cnt_step==2)
{
fail = solcon_read_next_solution(n,&m[1],ps);
m[0] = n;
send[0] = slv; send[1] = m[1]; send[2] = m[0];
MPI_Send(send,3,MPI_INT,slv,SEND_SMUL,MPI_COMM_WORLD);
MPI_Send(ps,2*n+5,MPI_DOUBLE,slv,SEND_SSOL,MPI_COMM_WORLD);
}
/* end of obtaining start solution & sending */
*(TaskCount+slv-1)=*(TaskCount+slv-1)+1;
}
else /* JOB queue is empty */
{
flght=flght-1; /* one in-flight task less */
if(v>3)
{printf ("Job queue empty!\n");
printf("flght=%d\n",flght);
}
}
MPI_Iprobe(MPI_ANY_SOURCE,MPI_ANY_TAG,MPI_COMM_WORLD,&flag,&status);
} /* while flag */
} /* if flght */
} /* while flght listpointer */
/* send termination to all workers */
if(v>3) printf("\nmanager sent termination\n");
for(slv=1;slv<=numbprocs-1;slv++)
{
count = -1;
MPI_Send(&count,1,MPI_INT,slv,COUNT_TAG,MPI_COMM_WORLD);
}
}
int main(int argc, char **argv) {
int rank, size, i;
int root = 0;
int hits = 0; // index used for 'hits'
int total = 1; // index used for 'total'
int msg_waiting = 0;
double results[2] = {0};
MPI_Init(&argc, &argv);
MPI_Comm comm = MPI_COMM_WORLD;
MPI_Comm_rank(comm, &rank);
MPI_Comm_size(comm, &size);
MPI_Status status;
MPI_Request request;
// is root process
if(rank == root) {
double area;
double total_hits = 0;
double total_pokes = 0;
while (1) {
// check each slave process for results (non-blocking)
for (i = 1; i < size; i++) {
MPI_Iprobe(i, 0, comm, &msg_waiting, &status);
// if slave process is sending results
if (msg_waiting) {
MPI_Recv(&results, 2, MPI_DOUBLE, i, 0, comm, &status);
total_hits += results[hits];
total_pokes += results[total];
}
}
if (total_pokes >= 15000000000) {
area = (total_hits / total_pokes) * 4;
printf("Area=%.12lf\n", area);
// send terminating message to each slave process
for (i = 1; i < size; i++) {
MPI_Isend(&area, 1, MPI_DOUBLE, i, 0, comm, &request);
}
break;
}
}
// is slave process
} else {
int cpu_count = get_cpu_count();
double shared_results[cpu_count * 2];
double l_hits = 0;
double l_total = 0;
pthread_t threads[cpu_count];
t_data thread_data[cpu_count];
for (i = 0; i < cpu_count; i++) {
thread_data[i].id = i;
thread_data[i].rank = rank;
thread_data[i].results = shared_results;
pthread_create(&threads[i], NULL, &throw_darts, &thread_data[i]);
}
// periodically reads results from shared memory; sends to root process
while(1) {
sleep(3);
// first checks for termination flag from root process
MPI_Iprobe(root, 0, comm, &msg_waiting, &status);
if (msg_waiting) {
// terminate threads
for (i = 0; i < cpu_count; i++) {
pthread_cancel(threads[i]);
}
break;
} else {
results[hits] = 0;
results[total] = 0;
for (i = 0; i < cpu_count; i++) {
results[hits] += shared_results[i * 2];
results[total] += shared_results[i * 2 + 1];
}
results[hits] -= l_hits;
results[total] -= l_total;
l_hits += results[hits];
l_total += results[total];
// send results to root process
MPI_Isend(&results, 2, MPI_DOUBLE, root, 0, comm, &request);
}
}
}
MPI_Finalize();
return 0;
}
int
main( int argc, char **argv)
{
int rank, size, i, recv_flag, ret, passed;
MPI_Status Status;
char message[17];
MPI_Init(&argc, &argv);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &size);
if (rank == 0) {
Test_Init("barrier", rank);
/* Receive the startup messages from each of the
other clients */
for (i = 0; i < size - 1; i++) {
MPI_Recv(message, 17, MPI_CHAR, MPI_ANY_SOURCE, 2000,
MPI_COMM_WORLD, &Status);
}
/* Now use Iprobe to make sure no more messages arive for a
while */
passed = 1;
for (i = 0; i < WAIT_TIMES; i++){
recv_flag = 0;
MPI_Iprobe(MPI_ANY_SOURCE, 2000, MPI_COMM_WORLD,
&recv_flag, &Status);
if (recv_flag)
passed = 0;
}
if (passed)
Test_Passed("Barrier Test 1");
else
Test_Failed("Barrier Test 1");
/* Now go into the barrier myself */
MPI_Barrier(MPI_COMM_WORLD);
/* And get everyones message who came out */
for (i = 0; i < size - 1; i++) {
MPI_Recv(message, 13, MPI_CHAR, MPI_ANY_SOURCE, 2000,
MPI_COMM_WORLD, &Status);
}
/* Now use Iprobe to make sure no more messages arive for a
while */
passed = 1;
for (i = 0; i < WAIT_TIMES; i++){
recv_flag = 0;
MPI_Iprobe(MPI_ANY_SOURCE, 2000, MPI_COMM_WORLD,
&recv_flag, &Status);
if (recv_flag)
passed = 0;
}
if (passed)
Test_Passed("Barrier Test 2");
else
Test_Failed("Barrier Test 2");
Test_Waitforall( );
ret = Summarize_Test_Results();
Test_Finalize();
MPI_Finalize();
return ret;
} else {
MPI_Send("Entering Barrier", 17, MPI_CHAR, 0, 2000, MPI_COMM_WORLD);
MPI_Barrier(MPI_COMM_WORLD);
MPI_Send("Past Barrier", 13, MPI_CHAR, 0, 2000, MPI_COMM_WORLD);
Test_Waitforall( );
MPI_Finalize();
return 0;
}
}
void frog_spin(int rank, int size, int nfrog, long * rng_state, MPI_Comm comm)
{
list frogs = NULL;
int i;
int start;
int end;
int chunk;
int msg;
chunk = nfrog/(size - 3);
start = chunk * (rank - 3);
end = start + chunk;
if ( rank == size - 1 ) {
end = nfrog;
}
/* Build initial list of frogs */
for ( i = start; i < end; i++ ) {
frog f = make_frog(0,0);
frogHop(0, 0, &f->x, &f->y, rng_state); /* random starting position */
push(f, &frogs);
}
i = 0;
/* Loop as for cell list, looking for interrupts. To ensure we
* don't get deadlocks, lots of magic happens in update_frog_list */
do {
MPI_Status status;
int flag;
msg = -1;
update_frog_list(&frogs, rng_state, comm);
MPI_Iprobe(0, CONTROL_TAG, comm, &flag, &status);
if ( flag ) {
MPI_Recv(&msg, 1, MPI_INT, status.MPI_SOURCE,
status.MPI_TAG, comm, &status);
}
if ( msg == YEAR_END ) {
int nfrogs[2] = {0,0};
list it = frogs;
/* Calculate global number of frogs (and diseased
* frogs) */
while ( it ) {
nfrogs[0]++;
nfrogs[1] += ((frog)(it->data))->diseased;
it = it->next;
}
/* Should do this by building a sub-communicator and doing
* a reduction, but that's hard work */
if ( rank == 3 ) {
int tmp[2];
int j;
for ( j = 4; j < size; j++ ) {
MPI_Recv(&tmp, 2, MPI_INT, j, FROG_REDUCTION, comm, &status);
nfrogs[0] += tmp[0];
nfrogs[1] += tmp[1];
}
printf("There are %d frogs (%d infected) in year %d\n",
nfrogs[0], nfrogs[1], i++);
} else {
MPI_Send(&nfrogs, 2, MPI_INT, 3, FROG_REDUCTION, comm);
}
}
} while ( msg != SIMULATION_END )
;
delete_list(&frogs, &delete_frog);
}
static void test_pair (void)
{
int prev, next, count, tag, index, i, outcount, indices[2];
int rank, size, flag, ierr, reqcount;
double send_buf[TEST_SIZE], recv_buf[TEST_SIZE];
double buffered_send_buf[TEST_SIZE * 2 + MPI_BSEND_OVERHEAD]; /* factor of two is based on guessing - only dynamic allocation would be safe */
void *buffer;
MPI_Status statuses[2];
MPI_Status status;
MPI_Request requests[2];
MPI_Comm dupcom, intercom;
#ifdef V_T
struct _VT_FuncFrameHandle {
char *name;
int func;
int frame;
};
typedef struct _VT_FuncFrameHandle VT_FuncFrameHandle_t;
VT_FuncFrameHandle_t normal_sends,
buffered_sends,
buffered_persistent_sends,
ready_sends,
sync_sends,
nblock_sends,
nblock_rsends,
nblock_ssends,
pers_sends,
pers_rsends,
pers_ssends,
sendrecv,
sendrecv_repl,
intercomm;
int classid;
VT_classdef( "Application:test_pair", &classid );
#define VT_REGION_DEF( _name, _nameframe, _class ) \
(_nameframe).name=_name; \
VT_funcdef( (_nameframe).name, _class, &((_nameframe).func) );
#define VT_BEGIN_REGION( _nameframe ) \
LOCDEF(); \
VT_begin( (_nameframe).func )
#define VT_END_REGION( _nameframe ) \
LOCDEF(); VT_end( (_nameframe).func )
#else
#define VT_REGION_DEF( _name, _nameframe, _class )
#define VT_BEGIN_REGION( _nameframe )
#define VT_END_REGION( _nameframe )
#endif
ierr = MPI_Comm_rank(MPI_COMM_WORLD, &rank);
ierr = MPI_Comm_size(MPI_COMM_WORLD, &size);
if ( size < 2 ) {
if ( rank == 0 ) {
printf("Program needs to be run on at least 2 processes.\n");
}
ierr = MPI_Abort( MPI_COMM_WORLD, 66 );
}
ierr = MPI_Comm_dup(MPI_COMM_WORLD, &dupcom);
if ( rank >= 2 ) {
/* printf( "%d Calling finalize.\n", rank ); */
ierr = MPI_Finalize( );
exit(0);
}
next = rank + 1;
if (next >= 2)
next = 0;
prev = rank - 1;
if (prev < 0)
prev = 1;
VT_REGION_DEF( "Normal_Sends", normal_sends, classid );
VT_REGION_DEF( "Buffered_Sends", buffered_sends, classid );
VT_REGION_DEF( "Buffered_Persistent_Sends", buffered_persistent_sends, classid );
VT_REGION_DEF( "Ready_Sends", ready_sends, classid );
VT_REGION_DEF( "Sync_Sends", sync_sends, classid );
VT_REGION_DEF( "nblock_Sends", nblock_sends, classid );
VT_REGION_DEF( "nblock_RSends", nblock_rsends, classid );
VT_REGION_DEF( "nblock_SSends", nblock_ssends, classid );
VT_REGION_DEF( "Pers_Sends", pers_sends, classid );
VT_REGION_DEF( "Pers_RSends", pers_rsends, classid );
VT_REGION_DEF( "Pers_SSends", pers_ssends, classid );
VT_REGION_DEF( "SendRecv", sendrecv, classid );
VT_REGION_DEF( "SendRevc_Repl", sendrecv_repl, classid );
VT_REGION_DEF( "InterComm", intercomm, classid );
/*
* Normal sends
*/
VT_BEGIN_REGION( normal_sends );
if (rank == 0)
printf ("Send\n");
tag = 0x100;
count = TEST_SIZE / 5;
clear_test_data(recv_buf,TEST_SIZE);
if (rank == 0) {
init_test_data(send_buf,TEST_SIZE,0);
LOCDEF();
MPI_Send(send_buf, count, MPI_DOUBLE, next, tag, MPI_COMM_WORLD);
MPI_Recv(recv_buf, TEST_SIZE, MPI_DOUBLE, MPI_ANY_SOURCE,
MPI_ANY_TAG, MPI_COMM_WORLD, &status);
msg_check(recv_buf, prev, tag, count, &status, TEST_SIZE, "send and recv");
}
else {
LOCDEF();
MPI_Recv(recv_buf, TEST_SIZE, MPI_DOUBLE,MPI_ANY_SOURCE, MPI_ANY_TAG,
MPI_COMM_WORLD, &status);
msg_check( recv_buf, prev, tag, count, &status, TEST_SIZE,"send and recv");
init_test_data(recv_buf,TEST_SIZE,1);
MPI_Send(recv_buf, count, MPI_DOUBLE, next, tag, MPI_COMM_WORLD);
}
VT_END_REGION( normal_sends );
/*
* Buffered sends
*/
VT_BEGIN_REGION( buffered_sends );
if (rank == 0)
printf ("Buffered Send\n");
tag = 138;
count = TEST_SIZE / 5;
clear_test_data(recv_buf,TEST_SIZE);
if (rank == 0) {
init_test_data(send_buf,TEST_SIZE,0);
LOCDEF();
MPI_Buffer_attach(buffered_send_buf, sizeof(buffered_send_buf));
MPI_Bsend(send_buf, count, MPI_DOUBLE, next, tag, MPI_COMM_WORLD);
MPI_Buffer_detach(&buffer, &size);
if(buffer != buffered_send_buf || size != sizeof(buffered_send_buf)) {
printf ("[%d] Unexpected buffer returned by MPI_Buffer_detach(): %p/%d != %p/%d\n", rank, buffer, size, buffered_send_buf, (int)sizeof(buffered_send_buf));
MPI_Abort(MPI_COMM_WORLD, 201);
}
MPI_Recv(recv_buf, TEST_SIZE, MPI_DOUBLE, MPI_ANY_SOURCE,
MPI_ANY_TAG, MPI_COMM_WORLD, &status);
msg_check(recv_buf, prev, tag, count, &status, TEST_SIZE, "send and recv");
}
else {
LOCDEF();
MPI_Recv(recv_buf, TEST_SIZE, MPI_DOUBLE,MPI_ANY_SOURCE, MPI_ANY_TAG,
MPI_COMM_WORLD, &status);
msg_check( recv_buf, prev, tag, count, &status, TEST_SIZE,"send and recv");
init_test_data(recv_buf,TEST_SIZE,1);
MPI_Send(recv_buf, count, MPI_DOUBLE, next, tag, MPI_COMM_WORLD);
}
VT_END_REGION( buffered_sends );
/*
* Buffered sends
*/
VT_BEGIN_REGION( buffered_persistent_sends );
if (rank == 0)
printf ("Buffered Persistent Send\n");
tag = 238;
count = TEST_SIZE / 5;
clear_test_data(recv_buf,TEST_SIZE);
if (rank == 0) {
init_test_data(send_buf,TEST_SIZE,0);
LOCDEF();
MPI_Buffer_attach(buffered_send_buf, sizeof(buffered_send_buf));
MPI_Bsend_init(send_buf, count, MPI_DOUBLE, next, tag, MPI_COMM_WORLD, requests);
MPI_Start(requests);
MPI_Wait(requests, statuses);
MPI_Request_free(requests);
MPI_Buffer_detach(&buffer, &size);
if(buffer != buffered_send_buf || size != sizeof(buffered_send_buf)) {
printf ("[%d] Unexpected buffer returned by MPI_Buffer_detach(): %p/%d != %p/%d\n", rank, buffer, size, buffered_send_buf, (int)sizeof(buffered_send_buf));
MPI_Abort(MPI_COMM_WORLD, 201);
}
MPI_Recv(recv_buf, TEST_SIZE, MPI_DOUBLE, MPI_ANY_SOURCE,
MPI_ANY_TAG, MPI_COMM_WORLD, &status);
msg_check(recv_buf, prev, tag, count, &status, TEST_SIZE, "send and recv");
}
else {
LOCDEF();
MPI_Recv(recv_buf, TEST_SIZE, MPI_DOUBLE,MPI_ANY_SOURCE, MPI_ANY_TAG,
MPI_COMM_WORLD, &status);
msg_check( recv_buf, prev, tag, count, &status, TEST_SIZE,"send and recv");
init_test_data(recv_buf,TEST_SIZE,1);
MPI_Send(recv_buf, count, MPI_DOUBLE, next, tag, MPI_COMM_WORLD);
}
VT_END_REGION( buffered_persistent_sends );
/*
* Ready sends. Note that we must insure that the receive is posted
* before the rsend; this requires using Irecv.
*/
VT_BEGIN_REGION( ready_sends );
if (rank == 0)
printf ("Rsend\n");
tag = 1456;
count = TEST_SIZE / 3;
clear_test_data(recv_buf,TEST_SIZE);
if (rank == 0) {
init_test_data(send_buf,TEST_SIZE,0);
MPI_Recv(MPI_BOTTOM, 0, MPI_INT, next, tag, MPI_COMM_WORLD, &status);
MPI_Rsend(send_buf, count, MPI_DOUBLE, next, tag, MPI_COMM_WORLD);
MPI_Probe(MPI_ANY_SOURCE, tag, MPI_COMM_WORLD, &status);
if (status.MPI_SOURCE != prev)
printf ("Incorrect src, expected %d, got %d\n",prev, status.MPI_SOURCE);
if (status.MPI_TAG != tag)
printf ("Incorrect tag, expected %d, got %d\n",tag, status.MPI_TAG);
MPI_Get_count(&status, MPI_DOUBLE, &i);
if (i != count)
printf ("Incorrect count, expected %d, got %d\n",count,i);
MPI_Recv(recv_buf, TEST_SIZE, MPI_DOUBLE, MPI_ANY_SOURCE, MPI_ANY_TAG,
MPI_COMM_WORLD, &status);
msg_check( recv_buf, prev, tag, count, &status, TEST_SIZE,
"rsend and recv");
}
else {
MPI_Irecv(recv_buf, TEST_SIZE, MPI_DOUBLE, MPI_ANY_SOURCE, MPI_ANY_TAG,
MPI_COMM_WORLD, requests);
MPI_Send( MPI_BOTTOM, 0, MPI_INT, next, tag, MPI_COMM_WORLD);
MPI_Wait(requests, &status);
msg_check( recv_buf, prev, tag, count, &status, TEST_SIZE,
"rsend and recv");
init_test_data(recv_buf,TEST_SIZE,1);
MPI_Send(recv_buf, count, MPI_DOUBLE, next, tag, MPI_COMM_WORLD);
}
VT_END_REGION( ready_sends );
/*
* Synchronous sends
*/
VT_BEGIN_REGION( sync_sends );
if (rank == 0)
printf ("Ssend\n");
tag = 1789;
count = TEST_SIZE / 3;
clear_test_data(recv_buf,TEST_SIZE);
if (rank == 0) {
init_test_data(send_buf,TEST_SIZE,0);
MPI_Iprobe(MPI_ANY_SOURCE, tag, MPI_COMM_WORLD, &flag, &status);
if (flag)
printf ("Iprobe succeeded! source %d, tag %d\n",status.MPI_SOURCE,
status.MPI_TAG);
MPI_Ssend(send_buf, count, MPI_DOUBLE, next, tag, MPI_COMM_WORLD);
while (!flag)
MPI_Iprobe(MPI_ANY_SOURCE, tag, MPI_COMM_WORLD, &flag, &status);
if (status.MPI_SOURCE != prev)
printf ("Incorrect src, expected %d, got %d\n",prev, status.MPI_SOURCE);
if (status.MPI_TAG != tag)
printf ("Incorrect tag, expected %d, got %d\n",tag, status.MPI_TAG);
MPI_Get_count(&status, MPI_DOUBLE, &i);
if (i != count)
printf ("Incorrect count, expected %d, got %d\n",count,i);
MPI_Recv(recv_buf, TEST_SIZE, MPI_DOUBLE, MPI_ANY_SOURCE, MPI_ANY_TAG,
MPI_COMM_WORLD, &status);
msg_check( recv_buf, prev, tag, count, &status, TEST_SIZE, "ssend and recv");
}
else {
MPI_Recv(recv_buf, TEST_SIZE, MPI_DOUBLE, MPI_ANY_SOURCE, MPI_ANY_TAG,
MPI_COMM_WORLD, &status);
msg_check( recv_buf, prev, tag, count, &status, TEST_SIZE, "ssend and recv"); init_test_data(recv_buf,TEST_SIZE,1);
MPI_Ssend(recv_buf, count, MPI_DOUBLE, next, tag, MPI_COMM_WORLD);
}
VT_END_REGION( sync_sends );
/*
* Nonblocking normal sends
*/
VT_BEGIN_REGION( nblock_sends );
if (rank == 0)
printf ("Isend\n");
tag = 2123;
count = TEST_SIZE / 5;
clear_test_data(recv_buf,TEST_SIZE);
if (rank == 0) {
MPI_Irecv(recv_buf, TEST_SIZE, MPI_DOUBLE,MPI_ANY_SOURCE, MPI_ANY_TAG,
MPI_COMM_WORLD, requests);
init_test_data(send_buf,TEST_SIZE,0);
MPI_Isend(send_buf, count, MPI_DOUBLE, next, tag, MPI_COMM_WORLD,
(requests+1));
MPI_Waitall(2, requests, statuses);
rq_check( requests, 2, "isend and irecv" );
msg_check(recv_buf,prev,tag,count,statuses, TEST_SIZE,"isend and irecv");
}
else {
MPI_Recv(recv_buf, TEST_SIZE, MPI_DOUBLE,MPI_ANY_SOURCE, MPI_ANY_TAG,
MPI_COMM_WORLD, &status);
msg_check(recv_buf,prev,tag,count,&status, TEST_SIZE,"isend and irecv"); init_test_data(recv_buf,TEST_SIZE,1);
MPI_Isend(recv_buf, count, MPI_DOUBLE, next, tag,MPI_COMM_WORLD,
(requests));
MPI_Wait((requests), &status);
rq_check(requests, 1, "isend (and recv)");
}
VT_END_REGION( nblock_sends );
/*
* Nonblocking ready sends
*/
VT_BEGIN_REGION( nblock_rsends );
if (rank == 0)
printf ("Irsend\n");
tag = 2456;
count = TEST_SIZE / 3;
clear_test_data(recv_buf,TEST_SIZE);
if (rank == 0) {
MPI_Irecv(recv_buf, TEST_SIZE, MPI_DOUBLE, MPI_ANY_SOURCE, MPI_ANY_TAG,
MPI_COMM_WORLD, requests);
init_test_data(send_buf,TEST_SIZE,0);
MPI_Sendrecv( MPI_BOTTOM, 0, MPI_INT, next, 0,
MPI_BOTTOM, 0, MPI_INT, next, 0,
dupcom, &status);
MPI_Irsend(send_buf, count, MPI_DOUBLE, next, tag, MPI_COMM_WORLD,
(requests+1));
reqcount = 0;
while (reqcount != 2) {
MPI_Waitany( 2, requests, &index, statuses);
if( index == 0 ) {
memcpy( &status, statuses, sizeof(status) );
}
reqcount++;
}
rq_check( requests, 1, "irsend and irecv");
msg_check(recv_buf,prev,tag,count,&status, TEST_SIZE,"irsend and irecv");
}
else {
MPI_Irecv(recv_buf, TEST_SIZE, MPI_DOUBLE, MPI_ANY_SOURCE, MPI_ANY_TAG,
MPI_COMM_WORLD, requests);
MPI_Sendrecv( MPI_BOTTOM, 0, MPI_INT, next, 0,
MPI_BOTTOM, 0, MPI_INT, next, 0,
dupcom, &status);
flag = 0;
while (!flag)
MPI_Test(requests, &flag, &status);
rq_check( requests, 1, "irsend and irecv (test)");
msg_check( recv_buf, prev, tag, count, &status, TEST_SIZE,
"irsend and irecv"); init_test_data(recv_buf,TEST_SIZE,1);
MPI_Irsend(recv_buf, count, MPI_DOUBLE, next, tag,
MPI_COMM_WORLD, requests);
MPI_Waitall(1, requests, statuses);
rq_check( requests, 1, "irsend and irecv");
}
VT_END_REGION( nblock_rsends );
/*
* Nonblocking synchronous sends
*/
VT_BEGIN_REGION( nblock_ssends );
if (rank == 0)
printf ("Issend\n");
tag = 2789;
count = TEST_SIZE / 3;
clear_test_data(recv_buf,TEST_SIZE);
if (rank == 0) {
MPI_Irecv(recv_buf, TEST_SIZE, MPI_DOUBLE, MPI_ANY_SOURCE, MPI_ANY_TAG,
MPI_COMM_WORLD, requests );
init_test_data(send_buf,TEST_SIZE,0);
MPI_Issend(send_buf, count, MPI_DOUBLE, next, tag, MPI_COMM_WORLD,
(requests+1));
flag = 0;
while (!flag)
MPI_Testall(2, requests, &flag, statuses);
rq_check( requests, 2, "issend and irecv (testall)");
msg_check( recv_buf, prev, tag, count, statuses, TEST_SIZE,
"issend and recv");
}
else {
MPI_Recv(recv_buf, TEST_SIZE, MPI_DOUBLE, MPI_ANY_SOURCE, MPI_ANY_TAG,
MPI_COMM_WORLD, &status);
msg_check( recv_buf, prev, tag, count, &status, TEST_SIZE,
"issend and recv"); init_test_data(recv_buf,TEST_SIZE,1);
MPI_Issend(recv_buf, count, MPI_DOUBLE, next, tag, MPI_COMM_WORLD,requests);
flag = 0;
while (!flag)
MPI_Testany(1, requests, &index, &flag, statuses);
rq_check( requests, 1, "issend and recv (testany)");
}
VT_END_REGION( nblock_ssends );
/*
* Persistent normal sends
*/
VT_BEGIN_REGION( pers_sends );
if (rank == 0)
printf ("Send_init\n");
tag = 3123;
count = TEST_SIZE / 5;
clear_test_data(recv_buf,TEST_SIZE);
MPI_Send_init(send_buf, count, MPI_DOUBLE, next, tag,
MPI_COMM_WORLD, requests);
MPI_Recv_init(recv_buf, TEST_SIZE, MPI_DOUBLE,MPI_ANY_SOURCE, MPI_ANY_TAG,
MPI_COMM_WORLD, (requests+1));
if (rank == 0) {
init_test_data(send_buf,TEST_SIZE,0);
MPI_Startall(2, requests);
MPI_Waitall(2, requests, statuses);
msg_check( recv_buf, prev, tag, count, (statuses+1),
TEST_SIZE, "persistent send/recv");
}
else {
MPI_Start((requests+1));
MPI_Wait((requests+1), &status);
msg_check( recv_buf, prev, tag, count, &status, TEST_SIZE,
"persistent send/recv");
init_test_data(send_buf,TEST_SIZE,1);
MPI_Start(requests);
MPI_Wait(requests, &status);
}
MPI_Request_free(requests);
MPI_Request_free((requests+1));
VT_END_REGION( pers_sends );
/*
* Persistent ready sends
*/
VT_BEGIN_REGION( pers_rsends );
if (rank == 0)
printf ("Rsend_init\n");
tag = 3456;
count = TEST_SIZE / 3;
clear_test_data(recv_buf,TEST_SIZE);
MPI_Rsend_init(send_buf, count, MPI_DOUBLE, next, tag,
MPI_COMM_WORLD, requests);
MPI_Recv_init(recv_buf, TEST_SIZE, MPI_DOUBLE, MPI_ANY_SOURCE,
MPI_ANY_TAG, MPI_COMM_WORLD, (requests+1));
if (rank == 0) {
init_test_data(send_buf,TEST_SIZE,0); MPI_Barrier( MPI_COMM_WORLD );
MPI_Startall(2, requests);
reqcount = 0;
while (reqcount != 2) {
MPI_Waitsome(2, requests, &outcount, indices, statuses);
for (i=0; i<outcount; i++) {
if (indices[i] == 1) {
msg_check( recv_buf, prev, tag, count, (statuses+i),
TEST_SIZE, "waitsome");
}
reqcount++;
}
}
}
else {
MPI_Start((requests+1)); MPI_Barrier( MPI_COMM_WORLD );
flag = 0;
while (!flag)
MPI_Test((requests+1), &flag, &status);
msg_check( recv_buf, prev, tag, count, &status, TEST_SIZE, "test");
init_test_data(send_buf,TEST_SIZE,1);
MPI_Start(requests);
MPI_Wait(requests, &status);
}
MPI_Request_free(requests);
MPI_Request_free((requests+1));
VT_END_REGION( pers_rsends );
/*
* Persistent synchronous sends
*/
VT_BEGIN_REGION( pers_ssends );
if (rank == 0)
printf ("Ssend_init\n");
tag = 3789;
count = TEST_SIZE / 3;
clear_test_data(recv_buf,TEST_SIZE);
MPI_Ssend_init(send_buf, count, MPI_DOUBLE, next, tag,
MPI_COMM_WORLD, (requests+1));
MPI_Recv_init(recv_buf, TEST_SIZE, MPI_DOUBLE, MPI_ANY_SOURCE,
MPI_ANY_TAG, MPI_COMM_WORLD, requests);
if (rank == 0) {
init_test_data(send_buf,TEST_SIZE,0);
MPI_Startall(2, requests);
reqcount = 0;
while (reqcount != 2) {
MPI_Testsome(2, requests, &outcount, indices, statuses);
for (i=0; i<outcount; i++) {
if (indices[i] == 0) {
msg_check( recv_buf, prev, tag, count, (statuses+i),
TEST_SIZE, "testsome");
}
reqcount++;
}
}
}
else {
MPI_Start(requests);
flag = 0;
while (!flag)
MPI_Testany(1, requests, &index, &flag, statuses);
msg_check( recv_buf, prev, tag, count, statuses, TEST_SIZE, "testany" );
init_test_data(send_buf,TEST_SIZE,1);
MPI_Start((requests+1));
MPI_Wait((requests+1), &status);
}
MPI_Request_free(requests);
MPI_Request_free((requests+1));
VT_END_REGION( pers_ssends );
/*
* Send/receive.
*/
VT_BEGIN_REGION( sendrecv );
if (rank == 0)
printf ("Sendrecv\n");
tag = 4123;
count = TEST_SIZE / 5;
clear_test_data(recv_buf,TEST_SIZE);
if (rank == 0) {
init_test_data(send_buf,TEST_SIZE,0);
MPI_Sendrecv(send_buf, count, MPI_DOUBLE, next, tag,
recv_buf, count, MPI_DOUBLE, prev, tag,
MPI_COMM_WORLD, &status );
msg_check( recv_buf, prev, tag, count, &status, TEST_SIZE,
"sendrecv");
}
else {
MPI_Recv(recv_buf, TEST_SIZE, MPI_DOUBLE, MPI_ANY_SOURCE,
MPI_ANY_TAG, MPI_COMM_WORLD, &status);
msg_check( recv_buf, prev, tag, count, &status, TEST_SIZE,
"recv/send"); init_test_data(recv_buf,TEST_SIZE,1);
MPI_Send(recv_buf, count, MPI_DOUBLE, next, tag, MPI_COMM_WORLD);
}
VT_END_REGION( sendrecv );
#ifdef V_T
VT_flush();
#endif
/*
* Send/receive replace.
*/
VT_BEGIN_REGION( sendrecv_repl );
if (rank == 0)
printf ("Sendrecv_replace\n");
tag = 4456;
count = TEST_SIZE / 3;
if (rank == 0) {
init_test_data(recv_buf, TEST_SIZE,0);
for (i=count; i< TEST_SIZE; i++)
recv_buf[i] = 0.0;
MPI_Sendrecv_replace(recv_buf, count, MPI_DOUBLE,
next, tag, prev, tag, MPI_COMM_WORLD, &status);
msg_check( recv_buf, prev, tag, count, &status, TEST_SIZE,
"sendrecvreplace");
}
else {
clear_test_data(recv_buf,TEST_SIZE);
MPI_Recv(recv_buf, TEST_SIZE, MPI_DOUBLE, MPI_ANY_SOURCE,
MPI_ANY_TAG, MPI_COMM_WORLD, &status);
msg_check( recv_buf, prev, tag, count, &status, TEST_SIZE,
"recv/send for replace"); init_test_data(recv_buf,TEST_SIZE,1);
MPI_Send(recv_buf, count, MPI_DOUBLE, next, tag, MPI_COMM_WORLD);
}
VT_END_REGION( sendrecv_repl );
/*
* Send/Receive via inter-communicator
*/
VT_BEGIN_REGION( intercomm );
MPI_Intercomm_create(MPI_COMM_SELF, 0, MPI_COMM_WORLD, next, 1, &intercom);
if (rank == 0)
printf ("Send via inter-communicator\n");
tag = 4018;
count = TEST_SIZE / 5;
clear_test_data(recv_buf,TEST_SIZE);
if (rank == 0) {
init_test_data(send_buf,TEST_SIZE,0);
LOCDEF();
MPI_Send(send_buf, count, MPI_DOUBLE, 0, tag, intercom);
MPI_Recv(recv_buf, TEST_SIZE, MPI_DOUBLE, MPI_ANY_SOURCE,
MPI_ANY_TAG, intercom, &status);
msg_check(recv_buf, 0, tag, count, &status, TEST_SIZE, "send and recv via inter-communicator");
}
else if (rank == 1) {
LOCDEF();
MPI_Recv(recv_buf, TEST_SIZE, MPI_DOUBLE,MPI_ANY_SOURCE, MPI_ANY_TAG,
intercom, &status);
msg_check( recv_buf, 0, tag, count, &status, TEST_SIZE,"send and recv via inter-communicator");
init_test_data(recv_buf,TEST_SIZE,0);
MPI_Send(recv_buf, count, MPI_DOUBLE, 0, tag, intercom);
}
VT_END_REGION( normal_sends );
MPI_Comm_free(&intercom);
MPI_Comm_free(&dupcom);
}
void
img (const char *FileNameImg)
{
FILE *FileImg;
COLOR *TabColor, *Color, *TileColor;
STRING Name;
INDEX i, j, rank;
BYTE Byte;
int N = 18988, err, provided;
int next_proc;
MPI_Request rs;
MPI_Status status;
MPI_Init_thread(NULL, NULL, MPI_THREAD_MULTIPLE, &provided);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &P);
next_proc = (rank + 1) % P;
if (next_proc == 0) next_proc++;
P--;
MPI_Type_vector(1, 3, 0, MPI_FLOAT, &MPI_COLOR);
MPI_Type_commit(&MPI_COLOR);
if (rank == 0) {
strcpy (Name, FileNameImg);
strcat (Name, ".ppm");
INIT_FILE (FileImg, Name, "w");
fprintf (FileImg, "P6\n%d %d\n255\n", Img.Pixel.i, Img.Pixel.j);
}
// number of tiles
Ci = Img.Pixel.i / TILE_SIZE + (Img.Pixel.i % TILE_SIZE?1:0); // number of tiles in dimension i
Cj = Img.Pixel.j / TILE_SIZE + (Img.Pixel.i % TILE_SIZE?1:0); // number of tiles in dimension j
int C = Ci * Cj;
int q = (C+P-1)/P;
int size = Img.Pixel.i * Img.Pixel.j ;
N = C/2+1;
// buffer for each tile
INIT_MEM (TileColor, TILE_SIZE * TILE_SIZE, COLOR);
if (rank != 0) {
struct TileQueue tiles = {NULL, NULL};
// Init tasks
init(&tiles,rank,q,N,C);
// Init mutex, semaphores & threads
pthread_mutex_init(&mutex,NULL);
pthread_mutex_init(&mutex_time,NULL);
sem_init(&wait_work, 0, 0);
sem_init(&ask_work, 0, 0);
pthread_t tid[NB_THREADS];
for (i = 0; i < NB_THREADS; i++){
err = pthread_create(&(tid[i]), NULL, (void*)tile_fill, (void*)&tiles);
if (err != 0)
printf("\ncan't create thread :[%s]", strerror(err));
}
// vol de travail ?
if (vol){
// Main thread: Communicator
while (!terminated)
{
int flag = 0, msg;
MPI_Iprobe(MPI_ANY_SOURCE, MPI_ANY_TAG, MPI_COMM_WORLD, &flag, &status);
if (flag) {
// We DID receive a communication so we CAN do a blocking receive
MPI_Recv(&msg, 1, MPI_INT, status.MPI_SOURCE, status.MPI_TAG, MPI_COMM_WORLD, &status);
// TAG contains the information about what the message is
switch (status.MPI_TAG){
case TERMINATE: // No more jobs; threads can finish their jobs and return
MPI_Isend(&msg, 1, MPI_INT, next_proc, TERMINATE, MPI_COMM_WORLD, &rs);
terminated = 1;
break;
case WORK_ASK: // msg-th process is seeking for job
pthread_mutex_lock(&mutex);
if(!isEmpty(&tiles)){
int tile = firstElement(&tiles);
pop(&tiles);
pthread_mutex_unlock(&mutex);
MPI_Isend(&tile, 1, MPI_INT, msg, WORK_SEND, MPI_COMM_WORLD, &rs);
}
else {
pthread_mutex_unlock(&mutex);
if (msg == rank){
MPI_Isend(&msg, 1, MPI_INT, next_proc, TERMINATE, MPI_COMM_WORLD, &rs);
terminated = 1;
} else {
MPI_Isend(&msg, 1, MPI_INT, next_proc, WORK_ASK, MPI_COMM_WORLD, &rs);
}
}
break;
case WORK_SEND: // Received a job
pthread_mutex_lock(&mutex);
addTile(&tiles, msg);
pthread_mutex_unlock(&mutex);
sem_post(&wait_work);
break;
default:
fprintf(stderr, "Err: Unknown message: %d, with tag %d\n", msg,status.MPI_TAG);
break;
}
}
if (sem_trywait(&ask_work) == 0){
MPI_Isend(&rank, 1, MPI_INT, next_proc, WORK_ASK, MPI_COMM_WORLD, &rs);
}
}
for (i = 0; i < NB_THREADS; i++){
sem_post(&wait_work);
}
}
for (i = 0; i < NB_THREADS; i++){
pthread_join(tid[i],NULL);
}
pthread_mutex_destroy(&mutex);
pthread_mutex_destroy(&mutex_time);
sem_destroy(&wait_work);
sem_destroy(&ask_work);
fprintf(stderr, "%d %ld\n", rank, local_time);
}
// process 0 gathers all the tiles
if (rank == 0){
// If fake tasks: we don't receive anything and don't write the image
FILE* fd = fopen("config","r");
if (fd != NULL){
int fake;
fscanf(fd,"%d\n",&fake);
if (fake){
EXIT_FILE(FileImg);
EXIT_MEM(TileColor);
MPI_Finalize();
return;
}
}
// final image buffer that will receive the tiles
INIT_MEM (TabColor, size, COLOR);
// Receive tiles from other procs
for (i = 0; i < C ; i++){
MPI_Recv(TileColor, TILE_SIZE * TILE_SIZE, MPI_COLOR, MPI_ANY_SOURCE, MPI_ANY_TAG, MPI_COMM_WORLD, &status);
int current_tile = status.MPI_TAG - TILE_TAG_INDEX;
int j_begin = rank_j(current_tile,Cj);
int index_begin = rank_i(current_tile,Ci) + j_begin * Img.Pixel.i;
for (j = 0; j < TILE_SIZE && j_begin + j < Img.Pixel.j; j++) {
memcpy(&TabColor[index_begin + j * Img.Pixel.i],&TileColor[j * TILE_SIZE],MIN(Img.Pixel.i - rank_i(current_tile,Ci),TILE_SIZE) * sizeof(COLOR));
}
}
// writing in file
for (j = 0, Color = TabColor; j < size; j++, Color++) {
Byte = Color->r < 1.0 ? 255.0*Color->r : 255.0;
putc (Byte, FileImg);
Byte = Color->g < 1.0 ? 255.0*Color->g : 255.0;
putc (Byte, FileImg);
Byte = Color->b < 1.0 ? 255.0*Color->b : 255.0;
putc (Byte, FileImg);
}
EXIT_FILE (FileImg);
printf("Copied in file\n");
EXIT_MEM (TabColor);
}
EXIT_MEM (TileColor);
MPI_Finalize();
}
/**
* Kontrola a zpracovani zprav pomoci MPI.
*/
void mpi_handle() {
MPI_Status mpi_status;
int mpi_flag;
unsigned int l;
unsigned int v = 1;
char *b = NULL;
MPI_Iprobe(MPI_ANY_SOURCE, MPI_ANY_TAG, MPI_COMM_WORLD, &mpi_flag,
&mpi_status);
// odeslu nejnizsi nalezenou penalizaci (podmineno viz. telo funkce)
mpi_send_penalty();
if(mpi_flag) {
srpdebug("msg", node, "zprava <tag=%d>", mpi_status.MPI_TAG);
switch(mpi_status.MPI_TAG) {
case MSG_REQUEST:
srpdebug("msg", node, "MSG_REQUEST");
MPI_Recv(&v, 1, MPI_UNSIGNED, mpi_status.MPI_SOURCE, MSG_REQUEST,
MPI_COMM_WORLD, &mpi_status);
srpdebug("mpi", node, "prijata zadost o praci <src=%d>",
mpi_status.MPI_SOURCE);
// TODO rozdelit praci a odpovedet MSG_STACK nebo MSG_NOSTACK
if(s->s_real > 1) {
mpi_send_stack(2, mpi_status.MPI_SOURCE);
} else {
mpi_send_nostack(mpi_status.MPI_SOURCE);
}
break;
case MSG_STACK:
srpdebug("msg", node, "MSG_STACK");
mpi_recv_stack(mpi_status.MPI_SOURCE);
break;
case MSG_NOSTACK:
srpdebug("msg", node, "MSG_NOSTACK");
mpi_recv_nostack(mpi_status.MPI_SOURCE);
break;
case MSG_PENALTY:
srpdebug("msg", node, "MSG_PENALTY");
mpi_recv_penalty(mpi_status.MPI_SOURCE);
break;
case MSG_TOKEN:
srpdebug("msg", node, "MSG_TOKEN");
mpi_recv_token(mpi_status.MPI_SOURCE);
break;
case MSG_FINALIZE:
srpdebug("msg", node, "MSG_FINALIZE");
mpi_recv_finalize(mpi_status.MPI_SOURCE);
finalize();
break;
default:
// neznama zprava
srpprintf(stderr, "msg", node, "neznama zprava");
exit(EXIT_FAILURE);
}
} /* else {
srpdebug("mpi", node, "zadne zpravy");
} */
}