int main(int argc, char* argv[])
{
int msg;
MPI_Comm parent, child;
int rank, size;
char hostname[512];
pid_t pid;
int i;
char *cmds[2];
char *argv0[] = { "foo", NULL };
char *argv1[] = { "bar", NULL };
char **spawn_argv[2];
int maxprocs[] = { 1, 1 };
MPI_Info info[] = { MPI_INFO_NULL, MPI_INFO_NULL };
cmds[1] = cmds[0] = argv[0];
spawn_argv[0] = argv0;
spawn_argv[1] = argv1;
MPI_Init(NULL, NULL);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &size);
MPI_Comm_get_parent(&parent);
/* If we get COMM_NULL back, then we're the parent */
if (MPI_COMM_NULL == parent) {
pid = getpid();
printf("Parent [pid %ld] about to spawn!\n", (long)pid);
MPI_Comm_spawn_multiple(2, cmds, spawn_argv, maxprocs,
info, 0, MPI_COMM_WORLD,
&child, MPI_ERRCODES_IGNORE);
printf("Parent done with spawn\n");
if (0 == rank) {
msg = 38;
printf("Parent sending message to children\n");
MPI_Send(&msg, 1, MPI_INT, 0, 1, child);
MPI_Send(&msg, 1, MPI_INT, 1, 1, child);
}
MPI_Comm_disconnect(&child);
printf("Parent disconnected\n");
}
/* Otherwise, we're the child */
else {
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &size);
gethostname(hostname, 512);
pid = getpid();
printf("Hello from the child %d of %d on host %s pid %ld: argv[1] = %s\n", rank, size, hostname, (long)pid, argv[1]);
if (0 == rank) {
MPI_Recv(&msg, 1, MPI_INT, 0, 1, parent, MPI_STATUS_IGNORE);
printf("Child %d received msg: %d\n", rank, msg);
}
MPI_Comm_disconnect(&parent);
printf("Child %d disconnected\n", rank);
}
MPI_Finalize();
return 0;
}
void mpi_comm_spawn_multiple_f(MPI_Fint *count, char *array_commands,
char *array_argv,
MPI_Fint *array_maxprocs,
MPI_Fint *array_info, MPI_Fint *root,
MPI_Fint *comm, MPI_Fint *intercomm,
MPI_Fint *array_errcds, MPI_Fint *ierr,
int cmd_len, int argv_len)
{
MPI_Comm c_comm, c_new_comm;
MPI_Info *c_info;
int size, array_size, i;
int *c_errs;
char **c_array_commands;
char ***c_array_argv;
OMPI_ARRAY_NAME_DECL(array_maxprocs);
OMPI_ARRAY_NAME_DECL(array_errcds);
c_comm = MPI_Comm_f2c(*comm);
MPI_Comm_size(c_comm, &size);
array_size = OMPI_FINT_2_INT(*count);
/* It's allowed to ignore the errcodes */
if (OMPI_IS_FORTRAN_ERRCODES_IGNORE(array_errcds)) {
c_errs = MPI_ERRCODES_IGNORE;
} else {
OMPI_ARRAY_FINT_2_INT_ALLOC(array_errcds, size);
c_errs = OMPI_ARRAY_NAME_CONVERT(array_errcds);
}
/* It's allowed to have no argv */
if (OMPI_IS_FORTRAN_ARGVS_NULL(array_argv)) {
c_array_argv = MPI_ARGVS_NULL;
} else {
ompi_fortran_multiple_argvs_f2c(OMPI_FINT_2_INT(*count), array_argv,
argv_len, &c_array_argv);
}
OMPI_ARRAY_FINT_2_INT(array_maxprocs, array_size);
ompi_fortran_argv_f2c(array_commands, cmd_len, &c_array_commands);
c_info = malloc (array_size * sizeof(MPI_Info));
for (i = 0; i < array_size; ++i) {
c_info[i] = MPI_Info_f2c(array_info[i]);
}
*ierr =
OMPI_INT_2_FINT(MPI_Comm_spawn_multiple(OMPI_FINT_2_INT(*count),
c_array_commands,
c_array_argv,
OMPI_ARRAY_NAME_CONVERT(array_maxprocs),
c_info,
OMPI_FINT_2_INT(*root),
c_comm, &c_new_comm,
c_errs));
if (MPI_SUCCESS == OMPI_FINT_2_INT(*ierr)) {
*intercomm = MPI_Comm_c2f(c_new_comm);
}
if (!OMPI_IS_FORTRAN_ERRCODES_IGNORE(array_errcds)) {
OMPI_ARRAY_INT_2_FINT(array_errcds, size);
}
OMPI_ARRAY_FINT_2_INT_CLEANUP(array_maxprocs);
opal_argv_free(c_array_commands);
if (MPI_ARGVS_NULL != c_array_argv && NULL != c_array_argv) {
for (i = 0; i < OMPI_FINT_2_INT(*count); ++i) {
opal_argv_free(c_array_argv[i]);
}
}
free(c_array_argv);
}
void repairComm(MPI_Comm * broken, MPI_Comm * repaired, int iteration, int * listFails, int * numFails,
int * numNodeFails, int sumPrevNumNodeFails, int argc, char ** argv, int verbosity) {
MPI_Comm tempShrink, unorderIntracomm, tempIntercomm;
int i, ret, result, procsNeeded = 0, oldRank, newRank, oldGroupSize, rankKey = 0, flag;
int * tempRanks, * failedRanks, * errCodes, rank, hostfileLineIndex;
MPI_Group oldGroup, failedGroup, shrinkGroup;
int hostfileLastLineIndex, tempLineIndex, * failedNodeList = NULL, * nodeList = NULL, totNodeFailed = 0;
double startTime = 0.0, endTime;
int nprocs, j, * shrinkMergeList;
char hostName[128];
gethostname(hostName, sizeof(hostName));
char ** appToLaunch;
char *** argvToLaunch;
int * procsNeededToLaunch;
MPI_Info * hostInfoToLaunch;
char ** hostNameToLaunch;
MPI_Comm_rank(*broken, &rank);
if(rank == 0)
startTime = MPI_Wtime();
#ifndef GLOBAL_DETECTION
MPI_Comm_size(*broken, &oldGroupSize);
MPI_Comm_group(*broken, &oldGroup);
MPI_Comm_rank(*broken, &oldRank);
OMPI_Comm_failure_ack(*broken);
OMPI_Comm_failure_get_acked(*broken, &failedGroup);
MPI_Group_size(failedGroup, &procsNeeded);
errCodes = (int *) malloc(sizeof(int) * procsNeeded);
// Figure out ranks of the processes which had failed
tempRanks = (int *) malloc(sizeof(int) * oldGroupSize);
failedRanks = (int *) malloc(sizeof(int) * oldGroupSize);
#pragma omp parallel for default(shared)
for(i = 0; i < oldGroupSize; i++)
tempRanks[i] = i;
MPI_Group_translate_ranks(failedGroup, procsNeeded, tempRanks, oldGroup, failedRanks);
#endif
double shrinkTime = MPI_Wtime();
// Shrink the broken communicator to remove failed procs
if(MPI_SUCCESS != (ret = OMPI_Comm_shrink(*broken, &tempShrink)))
printf("Iteration %d: OMPI_Comm_shrink (parent): ERROR!\n", iteration);
else {
if(verbosity > 1 )
printf("Iteration %d: OMPI_Comm_shrink (parent): SUCCESS\n", iteration);
}
if (verbosity > 0 && rank == 0)
printf("OMPI_Comm_shrink takes %0.6f Sec\n", MPI_Wtime() - shrinkTime);
#ifdef GLOBAL_DETECTION
MPI_Comm_group(*broken, &oldGroup);
MPI_Comm_group(tempShrink, &shrinkGroup);
MPI_Comm_size(*broken, &oldGroupSize);
MPI_Group_compare(oldGroup, shrinkGroup, &result);
if(result != MPI_IDENT)
MPI_Group_difference(oldGroup, shrinkGroup, &failedGroup);
MPI_Comm_rank(*broken, &oldRank);
MPI_Group_size(failedGroup, &procsNeeded);
errCodes = (int *) malloc(sizeof(int)*procsNeeded);
// Figure out ranks of the processes which had failed
tempRanks = (int*)malloc(sizeof(int)*oldGroupSize);
failedRanks = (int*)malloc(sizeof(int)*oldGroupSize);
#pragma omp parallel for default(shared)
for(i = 0; i < oldGroupSize; i++)
tempRanks[i] = i;
MPI_Group_translate_ranks(failedGroup, procsNeeded, tempRanks, oldGroup, failedRanks);
MPI_Group_free(&shrinkGroup);
#endif
// Assign number of failed processes
*numFails = procsNeeded;
hostNameToLaunch = (char **) malloc(procsNeeded * sizeof(char *));
if(verbosity > 0 && rank == 0)
printf("*** Iteration %d: Application: Number of process(es) failed in the corresponding "
"communicator is %d ***\n", iteration, procsNeeded);
if(rank == 0) {
endTime = MPI_Wtime();
printf("[%d]----- Creating failed process list takes %0.6f Sec (MPI_Wtime) -----\n", rank, endTime - startTime);
}
#ifdef RECOV_ON_SPARE_NODES
// Determining total number of node failed, and a list of them
hostfileLastLineIndex = getHostfileLastLineIndex(); //started from 0
nodeList = (int *) malloc((hostfileLastLineIndex+1) * sizeof(int));
memset(nodeList, 0, (hostfileLastLineIndex+1)*sizeof(int)); // initialize nodeList with 0's
for(int i = 0; i < procsNeeded; ++i) {
tempLineIndex = failedRanks[i]/SLOTS; //started from 0
nodeList[tempLineIndex] = 1;
}
for(int nodeCounter = 0; nodeCounter < (hostfileLastLineIndex+1); ++nodeCounter)
totNodeFailed += nodeList[nodeCounter];
*numNodeFails = totNodeFailed;
// Check if there is sufficient spare node available for recovery
if((hostfileLastLineIndex - totNodeFailed -sumPrevNumNodeFails) < (oldGroupSize-1)/SLOTS) {
if(rank == 0)
printf("[%d] There is no sufficient spare node available for recovery.\n", rank);
exit(0);
}
failedNodeList = (int *) malloc(totNodeFailed * sizeof(int));
memset(failedNodeList, 0, totNodeFailed * sizeof(int)); // initialize failedNodeList with 0's
int failedNodeCounter = 0;
for(int nodeCounter = 0; nodeCounter < (hostfileLastLineIndex+1); ++nodeCounter) {
if(nodeList[nodeCounter] == 1)
failedNodeList[failedNodeCounter++] = nodeCounter;
}
#endif
char * hostNameFailed = NULL;
#pragma omp parallel for default(shared)
for(i = 0; i < procsNeeded; ++i) {
// Assign list of processes failed
listFails[i] = failedRanks[i];
#ifdef RUN_ON_COMPUTE_NODES
tempLineIndex = failedRanks[i]/SLOTS; //started from 0
#ifdef RECOV_ON_SPARE_NODES
for(int j = 0; j < totNodeFailed; ++j) {
if(failedNodeList[j] == tempLineIndex)
hostfileLineIndex = hostfileLastLineIndex - j - sumPrevNumNodeFails;
}
#else // Recovery on the same node (no node failure, only process failure)
hostfileLineIndex = tempLineIndex;
#endif
hostNameToLaunch[i] = getHostToLaunch(hostfileLineIndex);
hostNameFailed = getHostToLaunch(tempLineIndex);
#else // Run on head node or personal machine
hostNameToLaunch[i] = (char *)hostName;
hostNameFailed = (char *)hostName;
#endif
if(verbosity > 0 && rank == 0)
printf("--- Iteration %d: Application: Process %d on node %s is failed! ---\n", iteration, failedRanks[i], hostNameFailed);
}
// Release memory of hostNameFailed
free(hostNameFailed);
appToLaunch = (char **) malloc(procsNeeded * sizeof(char *));
argvToLaunch = (char ***) malloc(procsNeeded * sizeof(char **));
procsNeededToLaunch = (int *) malloc(procsNeeded * sizeof(int));
hostInfoToLaunch = (MPI_Info *) malloc(procsNeeded * sizeof(MPI_Info));
argv[argc] = NULL;
#pragma omp parallel for default(shared)
for(i = 0; i < procsNeeded; i++) {
appToLaunch[i] = (char *)argv[0];
argvToLaunch[i] = (char **)argv;
procsNeededToLaunch[i] = 1;
// Host information where to spawn the processes
MPI_Info_create(&hostInfoToLaunch[i]);
MPI_Info_set(hostInfoToLaunch[i], (char *)"host", hostNameToLaunch[i]);
//MPI_Info_set(hostInfoToLaunch[i], "hostfile", "hostfile");
}
double spawnTime = MPI_Wtime();
#ifdef HANG_ON_REMOVE
OMPI_Comm_agree(tempShrink, &flag);
#endif
// Spawn the new process(es)
if(MPI_SUCCESS != (ret = MPI_Comm_spawn_multiple(procsNeeded, appToLaunch, argvToLaunch, procsNeededToLaunch,
hostInfoToLaunch, 0, tempShrink, &tempIntercomm, MPI_ERRCODES_IGNORE))) {
free(tempRanks);
free(failedRanks);
free(errCodes);
if(MPI_ERR_COMM == ret)
printf("Iteration %d: MPI_Comm_spawn_multiple: Invalid communicator (parent)\n", iteration);
if(MPI_ERR_ARG == ret)
printf("Iteration %d: MPI_Comm_spawn_multiple: Invalid argument (parent)\n", iteration);
if(MPI_ERR_INFO == ret)
printf("Iteration %d: MPI_Comm_spawn_multiple: Invalid info (parent)\n", iteration);
if((MPI_ERR_PROC_FAILED == ret) || (MPI_ERR_REVOKED == ret)) {
OMPI_Comm_revoke(tempShrink);
return repairComm(broken, repaired, iteration, listFails, numFails, numNodeFails,
sumPrevNumNodeFails, argc, argv, verbosity);
}
else {
fprintf(stderr, "Iteration %d: Unknown error with MPI_Comm_spawn_multiple (parent): %d\n", iteration, ret);
exit(1);
}
}
else {
if(verbosity > 0 && rank == 0) {
for(i = 0; i < procsNeeded; i++)
printf("Iteration %d: MPI_Comm_spawn_multiple (parent) [spawning failed process %d on "
"node %s]: SUCCESS\n", iteration, failedRanks[i], hostNameToLaunch[i]);
}
// Memory release. Moving the last two to the end of the function causes segmentation faults for 4 processes failure
}
if (verbosity > 0 && rank == 0)
printf("MPI_Comm_spawn_multiple takes %0.6f Sec\n", MPI_Wtime() - spawnTime);
double mergeTime = MPI_Wtime();
// Merge the new processes into a new communicator
if(MPI_SUCCESS != (ret = MPI_Intercomm_merge(tempIntercomm, false, &unorderIntracomm))) {
free(tempRanks);
free(failedRanks);
if((MPI_ERR_PROC_FAILED == ret) || (MPI_ERR_REVOKED == ret)) {
// Start the recovery over again if there is a failure
OMPI_Comm_revoke(tempIntercomm);
return repairComm(broken, repaired, iteration, listFails, numFails,
numNodeFails, sumPrevNumNodeFails, argc, argv, verbosity);
}
else if(MPI_ERR_COMM == ret) {
fprintf(stderr, "Iteration %d: Invalid communicator in MPI_Intercomm_merge (parent) %d\n", iteration, ret);
exit(1);
}
else if(MPI_ERR_INTERN == ret) {
fprintf(stderr, "Iteration %d: Acquaring memory error in MPI_Intercomm_merge ()%d\n", iteration, ret);
exit(1);
}
else {
fprintf(stderr, "Iteration %d: Unknown error with MPI_Intercomm_merge: %d\n", iteration, ret);
exit(1);
}
}
else {
if(verbosity > 1 )
printf("Iteration %d: MPI_Intercomm_merge (parent): SUCCESS\n", iteration);
}
if (verbosity > 0 && rank == 0)
printf("MPI_Intercomm_merge takes %0.6f Sec\n", MPI_Wtime() - mergeTime);
double agreeTime = MPI_Wtime();
// Synchronize. sometimes hangs in without this
// position of code and intercommunicator (not intra) is important
#ifdef HANG_ON_REMOVE
//MPI_Barrier(tempIntercomm);
OMPI_Comm_agree(tempIntercomm, &flag);// since some of the times MPI_Barrier hangs
#endif
if (verbosity > 0 && rank == 0)
printf("OMPI_Comm_agree takes %0.6f Sec\n", MPI_Wtime() - agreeTime);
// Sending failed ranks and number of processes failed to the the newly created ranks.
// oldGroupSize is the size of communicator before failure.
// procsNeeded is the number of processes that are failed
int * child = (int *) malloc(procsNeeded*sizeof(int));
#pragma omp parallel for default(shared)
for(i = 0; i < procsNeeded; i++)
child[i] = oldGroupSize - procsNeeded + i;
MPI_Comm_rank(unorderIntracomm, &newRank);
if(newRank == 0) {
int send_val[2];
for(i = 0; i < procsNeeded; i++) {
send_val[0] = failedRanks[i];
send_val[1] = procsNeeded;
if(MPI_SUCCESS != (ret = MPI_Send(&send_val, 2, MPI_INT, child[i], MERGE_TAG, unorderIntracomm))) {
free(tempRanks);
free(failedRanks);
if((MPI_ERR_PROC_FAILED == ret) || (MPI_ERR_REVOKED == ret)) {
// Start the recovery over again if there is a failure
OMPI_Comm_revoke(unorderIntracomm);
return repairComm(broken, repaired, iteration, listFails, numFails,
numNodeFails, sumPrevNumNodeFails, argc, argv, verbosity);
}
else {
fprintf(stderr, "Iteration %d: Unknown error with MPI_Send1 (parent): %d\n", iteration, ret);
exit(1);
}
}
else {
if(verbosity > 1 )
printf("Iteration %d: MPI_Send1 (parent): SUCCESS\n", iteration);
}
}
}
// Split the current world (splitted from original) to order the ranks.
MPI_Comm_rank(unorderIntracomm, &newRank);
MPI_Comm_size(unorderIntracomm, &nprocs);
// For one or more process failure (ordering)
shrinkMergeList = (int *) malloc(nprocs*sizeof(int));
j = 0;
for(i = 0; i < nprocs; i++) {
if(rankIsNotOnFailedList(i, failedRanks, procsNeeded))
shrinkMergeList[j++] = i;
}
for(i = j; i < nprocs; i++)
shrinkMergeList[i] = failedRanks[i-j];
for(i = 0; i < (nprocs - procsNeeded); i++) {
if(newRank == i)
rankKey = shrinkMergeList[i];
}
if(MPI_SUCCESS != (MPI_Comm_split(unorderIntracomm, 0, rankKey, repaired))) {
if((MPI_ERR_PROC_FAILED == ret) || (MPI_ERR_REVOKED == ret)) {
// Start the recovery over again if there is a failure
OMPI_Comm_revoke(unorderIntracomm);
return repairComm(broken, repaired, iteration, listFails, numFails,
numNodeFails, sumPrevNumNodeFails, argc, argv, verbosity);
}
else {
fprintf(stderr, "Iteration %d: Unknown error with MPI_Comm_split (parent): %d\n", iteration, ret);
exit(1);
}
}
else {
if(verbosity > 1 )
printf("Iteration %d: MPI_Comm_split (parent): SUCCESS\n", iteration);
}
// Release memory
free(appToLaunch);
free(argvToLaunch);
free(procsNeededToLaunch);
free(hostInfoToLaunch);
free(hostNameToLaunch);
free(shrinkMergeList);
free(errCodes);
MPI_Comm_free(&tempShrink);
free(tempRanks);
free(failedRanks);
free(child);
MPI_Group_free(&failedGroup);
MPI_Group_free(&oldGroup);
MPI_Comm_free(&tempIntercomm);
MPI_Comm_free(&unorderIntracomm);
#ifdef RECOV_ON_SPARE_NODES
if(failedNodeList != NULL)
free(failedNodeList);
if(nodeList != NULL)
free(nodeList);
#endif
}//repairComm()
static void armci_mpi2_spawn()
{
int i;
char server_program[100];
char **command_arr=NULL, **hostname_arr=NULL, **nid_arr=NULL;
int *size_arr=NULL;
MPI_Info *info_arr;
/* we need to start 1 data server process on each node. So a total of
"armci_nclus" data servers */
armci_nserver = armci_nclus;
select_server_program(server_program, armci_nserver);
armci_mpi2_debug(0, "armci_mpi2_init(): Spawning %d data server processes "
"running %s\n", armci_nserver, server_program);
/* allocate necessary data structures */
{
command_arr = (char**) malloc(armci_nserver * sizeof(char*));
size_arr = (int*) malloc(armci_nserver * sizeof(int));
info_arr = (MPI_Info*) malloc(armci_nserver * sizeof(MPI_Info));
hostname_arr = (char**) malloc(armci_nserver * sizeof(char*));
#ifdef SPAWN_CRAY_XT
nid_arr = (char**) malloc(armci_nserver * sizeof(char*));;
#endif
for(i=0; i<armci_nserver; i++)
{
hostname_arr[i] = (char*)malloc(MPI_MAX_PROCESSOR_NAME*sizeof(char));
}
if(command_arr==NULL || size_arr==NULL || info_arr==NULL ||
hostname_arr==NULL)
{
armci_die("armci_mpi2_spawn: malloc failed.", 0);
}
}
/**
* 1. root process collects hostnames (i.e. machine names) of where to
* spawn dataservers. ARMCI masters of respective node will return their
* hostnames.
*/
armci_gather_hostnames(hostname_arr);
/** 2. initialize MPI_Comm_spawn_multiple() arguments */
{
for(i=0; i<armci_nserver; i++)
{
command_arr[i] = (*_armci_argv)[0]; /*CHECK: path needs fix */
size_arr[i] = 1; /* 1 data server in each node */
MPI_Info_create(&info_arr[i]);
#ifdef SPAWN_CRAY_XT
asprintf(&nid_arr[i], "%d", atoi((hostname_arr[i] + 3)));
MPI_Info_set(info_arr[i], "host", nid_arr[i]); /*portability? */
#else
MPI_Info_set(info_arr[i], "host", hostname_arr[i]); /*portability? */
#endif
}
}
/**
* 3. MPI_Comm_spawn_multiple(): This is a collective call.
* Intercommunicator "ds_intercomm" contains only new dataserver processes.
*/
MPI_Check(
MPI_Comm_spawn_multiple(armci_nserver, command_arr, MPI_ARGVS_NULL,
size_arr, info_arr, ARMCI_ROOT, MPI_COMM_WORLD,
&MPI_COMM_CLIENT2SERVER, MPI_ERRCODES_IGNORE)
);
{
for(i=0; i<armci_nserver; i++) free(hostname_arr[i]);
free(command_arr);
free(size_arr);
free(info_arr);
free(hostname_arr);
#ifdef SPAWN_CRAY_XT
free(nid_arr);
#endif
}
}
