FORTRAN_API void FORT_CALL mpi_group_compare_ ( MPI_Fint *group1, MPI_Fint *group2, MPI_Fint *result, MPI_Fint *__ierr ) { int l_result; *__ierr = MPI_Group_compare( MPI_Group_f2c(*group1), MPI_Group_f2c(*group2), &l_result ); *result = l_result; }
JNIEXPORT jint JNICALL Java_mpi_Group_compare( JNIEnv *env, jclass jthis, jlong group1, jlong group2) { int result, rc; rc = MPI_Group_compare((MPI_Group)group1, (MPI_Group)group2, &result); ompi_java_exceptionCheck(env, rc); return result; }
int main(int argc, char *argv[]) { int rank; MPI_Status status; MPI_Init(&argc, &argv); MPI_Group gw, g1, g2, g3; MPI_Comm_group(MPI_COMM_WORLD, &gw); int new_ranks[] = {0, 2, 3}; MPI_Group_incl(gw, 3, new_ranks, &g1); int new_ranks2[] = {2, 3, 0}; MPI_Group_incl(gw, 3, new_ranks2, &g2); MPI_Group_incl(gw, 3, new_ranks2, &g3); int size1, size2, size3; MPI_Group_size(g1, &size1); MPI_Group_size(g1, &size2); MPI_Group_size(g1, &size3); if (size1 != size2 || size2 != size3 || size1 != 3) { return 1; } int r1, r2; MPI_Group_compare(g1, g2, &r1); MPI_Group_compare(g2, g3, &r2); if (r1 != MPI_SIMILAR || r2 != MPI_IDENT) { return 1; } MPI_Group_free(&g1); MPI_Group_free(&g3); MPI_Group_free(&g2); return 0; }
/* * Class: mpi_Group * Method: Compare * Signature: (Lmpi/Group;Lmpi/Group;)I */ JNIEXPORT jint JNICALL Java_mpi_Group_Compare(JNIEnv *env, jclass jthis, jobject group1, jobject group2) { int result; ompi_java_clearFreeList(env) ; MPI_Group_compare((MPI_Group)((*env)->GetLongField(env,group1,ompi_java.GrouphandleID)), (MPI_Group)((*env)->GetLongField(env,group2,ompi_java.GrouphandleID)), &result); return result; }
static VALUE group_equal(VALUE self, VALUE rgrp2) { int rv, flag; MPI_Group *grp1, *grp2; Data_Get_Struct(self, MPI_Group, grp1); Data_Get_Struct(grp2, MPI_Group, grp2); rv = MPI_Group_compare(*grp1, *grp2, &flag); mpi_exception(rv); return flag == MPI_IDENT ? Qtrue : Qfalse; }
void mpi_group_compare_f(MPI_Fint *group1, MPI_Fint *group2, MPI_Fint *result, MPI_Fint *ierr) { ompi_group_t *c_group1, *c_group2; OMPI_SINGLE_NAME_DECL(result); /* make the fortran to c representation conversion */ c_group1 = MPI_Group_f2c(*group1); c_group2 = MPI_Group_f2c(*group2); *ierr = OMPI_INT_2_FINT(MPI_Group_compare(c_group1, c_group2, OMPI_SINGLE_NAME_CONVERT(result) )); if (MPI_SUCCESS == OMPI_FINT_2_INT(*ierr)) { OMPI_SINGLE_INT_2_FINT(result); } }
int c2ffile_ ( int *file ) { MPI_File cFile = MPI_File_f2c( *file ); MPI_Group group, wgroup; int result; MPI_File_get_group( cFile, &group ); MPI_Comm_group( MPI_COMM_WORLD, &wgroup ); MPI_Group_compare( group, wgroup, &result ); if (result != MPI_IDENT) { fprintf( stderr, "File: did not get expected group\n" ); return 1; } MPI_Group_free( &group ); MPI_Group_free( &wgroup ); return 0; }
int main( int argc, char *argv[] ) { MPI_Group g1, g2, g4, g5, g45, selfgroup, g6; int ranks[16], size, rank, myrank, range[1][3]; int errs = 0; int i, rin[16], rout[16], result; MPI_Init(&argc,&argv); MPI_Comm_group( MPI_COMM_WORLD, &g1 ); MPI_Comm_rank( MPI_COMM_WORLD, &myrank ); MPI_Comm_size( MPI_COMM_WORLD, &size ); if (size < 8) { fprintf( stderr, "Test requires 8 processes (16 prefered) only %d provided\n", size ); errs++; } /* 16 members, this process is rank 0, return in group 1 */ ranks[0] = myrank; ranks[1] = 2; ranks[2] = 7; if (myrank == 2) ranks[1] = 3; if (myrank == 7) ranks[2] = 6; MPI_Group_incl( g1, 3, ranks, &g2 ); /* Check the resulting group */ MPI_Group_size( g2, &size ); MPI_Group_rank( g2, &rank ); if (size != 3) { fprintf( stderr, "Size should be %d, is %d\n", 3, size ); errs++; } if (rank != 0) { fprintf( stderr, "Rank should be %d, is %d\n", 0, rank ); errs++; } rin[0] = 0; rin[1] = 1; rin[2] = 2; MPI_Group_translate_ranks( g2, 3, rin, g1, rout ); for (i=0; i<3; i++) { if (rout[i] != ranks[i]) { fprintf( stderr, "translated rank[%d] %d should be %d\n", i, rout[i], ranks[i] ); errs++; } } /* Translate the process of the self group against another group */ MPI_Comm_group( MPI_COMM_SELF, &selfgroup ); rin[0] = 0; MPI_Group_translate_ranks( selfgroup, 1, rin, g1, rout ); if (rout[0] != myrank) { fprintf( stderr, "translated of self is %d should be %d\n", rout[0], myrank ); errs++; } for (i=0; i<size; i++) rin[i] = i; MPI_Group_translate_ranks( g1, size, rin, selfgroup, rout ); for (i=0; i<size; i++) { if (i == myrank && rout[i] != 0) { fprintf( stderr, "translated world to self of %d is %d\n", i, rout[i] ); errs++; } else if (i != myrank && rout[i] != MPI_UNDEFINED) { fprintf( stderr, "translated world to self of %d should be undefined, is %d\n", i, rout[i] ); errs++; } } MPI_Group_free( &selfgroup ); /* Exclude everyone in our group */ { int ii, *lranks, g1size; MPI_Group_size( g1, &g1size ); lranks = (int *)malloc( g1size * sizeof(int) ); for (ii=0; ii<g1size; ii++) lranks[ii] = ii; MPI_Group_excl( g1, g1size, lranks, &g6 ); if (g6 != MPI_GROUP_EMPTY) { fprintf( stderr, "Group formed by excluding all ranks not empty\n" ); errs++; MPI_Group_free( &g6 ); } free( lranks ); } /* Add tests for additional group operations */ /* g2 = incl 1,3,7 g3 = excl 1,3,7 intersect ( w, g2 ) => g2 intersect ( w, g3 ) => g3 intersect ( g2, g3 ) => empty g4 = rincl 1:n-1:2 g5 = rexcl 1:n-1:2 union( g4, g5 ) => world g6 = rincl n-1:1:-1 g7 = rexcl n-1:1:-1 union( g6, g7 ) => concat of entries, similar to world diff( w, g2 ) => g3 */ MPI_Group_free( &g2 ); range[0][0] = 1; range[0][1] = size-1; range[0][2] = 2; MPI_Group_range_excl( g1, 1, range, &g5 ); range[0][0] = 1; range[0][1] = size-1; range[0][2] = 2; MPI_Group_range_incl( g1, 1, range, &g4 ); MPI_Group_union( g4, g5, &g45 ); MPI_Group_compare( MPI_GROUP_EMPTY, g4, &result ); if (result != MPI_UNEQUAL) { errs++; fprintf( stderr, "Comparison with empty group gave %d, not 3\n", result ); } MPI_Group_free( &g4 ); MPI_Group_free( &g5 ); MPI_Group_free( &g45 ); /* Now, duplicate the test, but using negative strides */ range[0][0] = size-1; range[0][1] = 1; range[0][2] = -2; MPI_Group_range_excl( g1, 1, range, &g5 ); range[0][0] = size-1; range[0][1] = 1; range[0][2] = -2; MPI_Group_range_incl( g1, 1, range, &g4 ); MPI_Group_union( g4, g5, &g45 ); MPI_Group_compare( MPI_GROUP_EMPTY, g4, &result ); if (result != MPI_UNEQUAL) { errs++; fprintf( stderr, "Comparison with empty group (formed with negative strides) gave %d, not 3\n", result ); } MPI_Group_free( &g4 ); MPI_Group_free( &g5 ); MPI_Group_free( &g45 ); MPI_Group_free( &g1 ); if (myrank == 0) { if (errs == 0) { printf( " No Errors\n" ); } else { printf( "Found %d errors\n", errs ); } } MPI_Finalize(); return 0; }
int main( int argc, char **argv ) { int errs=0, toterr; MPI_Group basegroup; MPI_Group g1, g2, g3, g4, g5, g6, g7, g8, g9, g10; MPI_Group g3a, g3b; MPI_Comm comm, newcomm, splitcomm, dupcomm; int i, grp_rank, rank, grp_size, size, result; int nranks, *ranks, *ranks_out; int range[1][3]; int worldrank; MPI_Init( &argc, &argv ); MPI_Comm_rank( MPI_COMM_WORLD, &worldrank ); comm = MPI_COMM_WORLD; MPI_Comm_group( comm, &basegroup ); MPI_Comm_rank( comm, &rank ); MPI_Comm_size( comm, &size ); /* Get the basic information on this group */ MPI_Group_rank( basegroup, &grp_rank ); if (grp_rank != rank) { errs++; fprintf( stdout, "group rank %d != comm rank %d\n", grp_rank, rank ); } MPI_Group_size( basegroup, &grp_size ); if (grp_size != size) { errs++; fprintf( stdout, "group size %d != comm size %d\n", grp_size, size ); } /* Form a new communicator with inverted ranking */ MPI_Comm_split( comm, 0, size - rank, &newcomm ); MPI_Comm_group( newcomm, &g1 ); ranks = (int *)malloc( size * sizeof(int) ); ranks_out = (int *)malloc( size * sizeof(int) ); for (i=0; i<size; i++) ranks[i] = i; nranks = size; MPI_Group_translate_ranks( g1, nranks, ranks, basegroup, ranks_out ); for (i=0; i<size; i++) { if (ranks_out[i] != (size - 1) - i) { errs++; fprintf( stdout, "Translate ranks got %d expected %d\n", ranks_out[i], (size - 1) - i ); } } /* Check Compare */ MPI_Group_compare( basegroup, g1, &result ); if (result != MPI_SIMILAR) { errs++; fprintf( stdout, "Group compare should have been similar, was %d\n", result ); } MPI_Comm_dup( comm, &dupcomm ); MPI_Comm_group( dupcomm, &g2 ); MPI_Group_compare( basegroup, g2, &result ); if (result != MPI_IDENT) { errs++; fprintf( stdout, "Group compare should have been ident, was %d\n", result ); } MPI_Comm_split( comm, rank < size/2, rank, &splitcomm ); MPI_Comm_group( splitcomm, &g3 ); MPI_Group_compare( basegroup, g3, &result ); if (result != MPI_UNEQUAL) { errs++; fprintf( stdout, "Group compare should have been unequal, was %d\n", result ); } /* Build two groups that have this process and one other, but do not have the same processes */ ranks[0] = rank; ranks[1] = (rank + 1) % size; MPI_Group_incl( basegroup, 2, ranks, &g3a ); ranks[1] = (rank + size - 1) % size; MPI_Group_incl( basegroup, 2, ranks, &g3b ); MPI_Group_compare( g3a, g3b, &result ); if (result != MPI_UNEQUAL) { errs++; fprintf( stdout, "Group compare of equal sized but different groups should have been unequal, was %d\n", result ); } /* Build two new groups by excluding members; use Union to put them together again */ /* Exclude 0 */ for (i=0; i<size; i++) ranks[i] = i; MPI_Group_excl( basegroup, 1, ranks, &g4 ); /* Exclude 1-(size-1) */ MPI_Group_excl( basegroup, size-1, ranks+1, &g5 ); MPI_Group_union( g5, g4, &g6 ); MPI_Group_compare( basegroup, g6, &result ); if (result != MPI_IDENT) { int usize; errs++; /* See ordering requirements on union */ fprintf( stdout, "Group excl and union did not give ident groups\n" ); fprintf( stdout, "[%d] result of compare was %d\n", rank, result ); MPI_Group_size( g6, &usize ); fprintf( stdout, "Size of union is %d, should be %d\n", usize, size ); } MPI_Group_union( basegroup, g4, &g7 ); MPI_Group_compare( basegroup, g7, &result ); if (result != MPI_IDENT) { int usize; errs++; fprintf( stdout, "Group union of overlapping groups failed\n" ); fprintf( stdout, "[%d] result of compare was %d\n", rank, result ); MPI_Group_size( g7, &usize ); fprintf( stdout, "Size of union is %d, should be %d\n", usize, size ); } /* Use range_excl instead of ranks */ /* printf ("range excl\n" ); fflush( stdout ); */ range[0][0] = 1; range[0][1] = size-1; range[0][2] = 1; MPI_Group_range_excl( basegroup, 1, range, &g8 ); /* printf( "out of range excl\n" ); fflush( stdout ); */ MPI_Group_compare( g5, g8, &result ); /* printf( "out of compare\n" ); fflush( stdout ); */ if (result != MPI_IDENT) { errs++; fprintf( stdout, "Group range excl did not give ident groups\n" ); } /* printf( "intersection\n" ); fflush( stdout ); */ MPI_Group_intersection( basegroup, g4, &g9 ); MPI_Group_compare( g9, g4, &result ); if (result != MPI_IDENT) { errs++; fprintf( stdout, "Group intersection did not give ident groups\n" ); } /* Exclude EVERYTHING and check against MPI_GROUP_EMPTY */ /* printf( "range excl all\n" ); fflush( stdout ); */ range[0][0] = 0; range[0][1] = size-1; range[0][2] = 1; MPI_Group_range_excl( basegroup, 1, range, &g10 ); /* printf( "done range excl all\n" ); fflush(stdout); */ MPI_Group_compare( g10, MPI_GROUP_EMPTY, &result ); /* printf( "done compare to MPI_GROUP_EMPTY\n" ); fflush(stdout); */ if (result != MPI_IDENT) { errs++; fprintf( stdout, "MPI_GROUP_EMPTY didn't compare against empty group\n"); } /* printf( "freeing groups\n" ); fflush( stdout ); */ MPI_Group_free( &basegroup ); MPI_Group_free( &g1 ); MPI_Group_free( &g2 ); MPI_Group_free( &g3 ); MPI_Group_free( &g3a ); MPI_Group_free( &g3b ); MPI_Group_free( &g4 ); MPI_Group_free( &g5 ); MPI_Group_free( &g6 ); MPI_Group_free( &g7 ); MPI_Group_free( &g8 ); MPI_Group_free( &g9 ); MPI_Group_free( &g10 ); MPI_Comm_free( &dupcomm ); MPI_Comm_free( &splitcomm ); MPI_Comm_free( &newcomm ); MPI_Allreduce( &errs, &toterr, 1, MPI_INT, MPI_SUM, MPI_COMM_WORLD ); if (worldrank == 0) { if (toterr == 0) printf( " No Errors\n" ); else printf( "Found %d errors in MPI Group routines\n", toterr ); } MPI_Finalize(); return toterr; }
int main(int argc, char **argv) { int rankLeft[4] = {0, 1, 2, 3}, rankRight[4] = {4, 5, 6, 7}; int i, result; char outStr[600]; int nProcs, myRank; MPI_Group grpWorld, grpNew; MPI_Comm commNew; MPI_Init(&argc, &argv); MPI_Comm_size(MPI_COMM_WORLD, &nProcs); MPI_Comm_rank(MPI_COMM_WORLD, &myRank); MPI_Comm_group(MPI_COMM_WORLD, &grpWorld); if (myRank < nProcs / 2) { MPI_Group_incl(grpWorld, nProcs / 2, rankLeft, &grpNew); } else { MPI_Group_incl(grpWorld, nProcs / 2, rankRight, &grpNew); } MPI_Comm_create(MPI_COMM_WORLD, grpNew, &commNew); int myRankCommNew, nProcsCommNew; int myRankGrpNew, nProcsGrpNew; MPI_Comm_rank(commNew, &myRankCommNew); MPI_Comm_size(commNew, &nProcsCommNew); MPI_Group_rank(grpNew, &myRankGrpNew); MPI_Group_size(grpNew, &nProcsGrpNew); fprintf(stdout, "WorldRank: %d in %d, NewCommRank: %d in %d, NewGrpRank: %d in %d\n", myRank, nProcs, myRankCommNew, nProcsCommNew, myRankGrpNew, nProcsGrpNew); MPI_Barrier(MPI_COMM_WORLD); int sendBuf = myRank, recvBuf; MPI_Allreduce(&sendBuf, &recvBuf, 1, MPI_INT, MPI_SUM, commNew); fprintf(stdout, "WorldRank = %d, sendBuf = %d, recvBuf = %d\n", myRank, sendBuf, recvBuf); fflush(stdout); MPI_Barrier(MPI_COMM_WORLD); int ranks1[8] = {0, 1, 2, 3, 4, 5, 6, 7}, ranks2[8]; MPI_Group_compare(grpWorld, grpNew, &result); MPI_Group_translate_ranks(grpWorld, nProcs, ranks1, grpNew, ranks2); if (myRank == 0) { fprintf(stdout, "result = %d\n", result); } sprintf_s(outStr, "rank %d: ", myRank); for (i = 0; i < nProcs; i++) { sprintf_s(outStr, "%s%d = %d ", outStr, ranks1[i], ranks2[i]); } fprintf(stdout, "%s\n", outStr); MPI_Comm_free(&commNew); MPI_Group_free(&grpNew); MPI_Group_free(&grpWorld); MPI_Finalize(); return 0; }
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()
/* * This test makes sure that after a failure, the correct group of failed * processes is returned from MPIX_Comm_failure_ack/get_acked. */ int main(int argc, char **argv) { int rank, size, err, result, i; char buf[10] = " No errors"; char error[MPI_MAX_ERROR_STRING]; MPI_Group failed_grp, one_grp, world_grp; int one[] = { 1 }; int world_ranks[] = { 0, 1, 2 }; int failed_ranks[3]; MPI_Init(&argc, &argv); MPI_Comm_rank(MPI_COMM_WORLD, &rank); MPI_Comm_size(MPI_COMM_WORLD, &size); if (size < 3) { fprintf(stderr, "Must run with at least 3 processes\n"); MPI_Abort(MPI_COMM_WORLD, 1); } MPI_Comm_set_errhandler(MPI_COMM_WORLD, MPI_ERRORS_RETURN); if (rank == 1) { exit(EXIT_FAILURE); } if (rank == 0) { err = MPI_Recv(buf, 10, MPI_CHAR, 1, 0, MPI_COMM_WORLD, MPI_STATUS_IGNORE); if (MPI_SUCCESS == err) { fprintf(stderr, "Expected a failure for receive from rank 1\n"); MPI_Abort(MPI_COMM_WORLD, 1); } err = MPIX_Comm_failure_ack(MPI_COMM_WORLD); if (MPI_SUCCESS != err) { int ec; MPI_Error_class(err, &ec); MPI_Error_string(err, error, &size); fprintf(stderr, "MPIX_Comm_failure_ack returned an error: %d\n%s", ec, error); MPI_Abort(MPI_COMM_WORLD, 1); } err = MPIX_Comm_failure_get_acked(MPI_COMM_WORLD, &failed_grp); if (MPI_SUCCESS != err) { int ec; MPI_Error_class(err, &ec); MPI_Error_string(err, error, &size); fprintf(stderr, "MPIX_Comm_failure_get_acked returned an error: %d\n%s", ec, error); MPI_Abort(MPI_COMM_WORLD, 1); } MPI_Comm_group(MPI_COMM_WORLD, &world_grp); MPI_Group_incl(world_grp, 1, one, &one_grp); MPI_Group_compare(one_grp, failed_grp, &result); if (MPI_IDENT != result) { fprintf(stderr, "First failed group contains incorrect processes\n"); MPI_Group_size(failed_grp, &size); MPI_Group_translate_ranks(failed_grp, size, world_ranks, world_grp, failed_ranks); for (i = 0; i < size; i++) fprintf(stderr, "DEAD: %d\n", failed_ranks[i]); MPI_Abort(MPI_COMM_WORLD, 1); } MPI_Group_free(&failed_grp); err = MPI_Recv(buf, 10, MPI_CHAR, 2, 0, MPI_COMM_WORLD, MPI_STATUS_IGNORE); if (MPI_SUCCESS != err) { fprintf(stderr, "First receive failed\n"); MPI_Abort(MPI_COMM_WORLD, 1); } err = MPI_Recv(buf, 10, MPI_CHAR, 2, 0, MPI_COMM_WORLD, MPI_STATUS_IGNORE); if (MPI_SUCCESS == err) { fprintf(stderr, "Expected a failure for receive from rank 2\n"); MPI_Abort(MPI_COMM_WORLD, 1); } err = MPIX_Comm_failure_get_acked(MPI_COMM_WORLD, &failed_grp); if (MPI_SUCCESS != err) { int ec; MPI_Error_class(err, &ec); MPI_Error_string(err, error, &size); fprintf(stderr, "MPIX_Comm_failure_get_acked returned an error: %d\n%s", ec, error); MPI_Abort(MPI_COMM_WORLD, 1); } MPI_Group_compare(one_grp, failed_grp, &result); if (MPI_IDENT != result) { fprintf(stderr, "Second failed group contains incorrect processes\n"); MPI_Group_size(failed_grp, &size); MPI_Group_translate_ranks(failed_grp, size, world_ranks, world_grp, failed_ranks); for (i = 0; i < size; i++) fprintf(stderr, "DEAD: %d\n", failed_ranks[i]); MPI_Abort(MPI_COMM_WORLD, 1); } fprintf(stdout, " No errors\n"); } else if (rank == 2) { MPI_Ssend(buf, 10, MPI_CHAR, 0, 0, MPI_COMM_WORLD); exit(EXIT_FAILURE); } MPI_Group_free(&failed_grp); MPI_Group_free(&one_grp); MPI_Group_free(&world_grp); MPI_Finalize(); }
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 errs = 0; int rc, result; int ranks[1]; MPI_Group group, outgroup; MPI_Comm comm; MTest_Init(&argc, &argv); /* To improve reporting of problems about operations, we * change the error handler to errors return */ MPI_Comm_set_errhandler(MPI_COMM_WORLD, MPI_ERRORS_RETURN); while (MTestGetComm(&comm, 1)) { if (comm == MPI_COMM_NULL) continue; MPI_Comm_group(comm, &group); rc = MPI_Group_incl(group, 0, 0, &outgroup); if (rc) { errs++; MTestPrintError(rc); printf("Error in creating an empty group with (0,0)\n"); /* Some MPI implementations may reject a null "ranks" pointer */ rc = MPI_Group_incl(group, 0, ranks, &outgroup); if (rc) { errs++; MTestPrintError(rc); printf("Error in creating an empty group with (0,ranks)\n"); } } if (outgroup != MPI_GROUP_EMPTY) { /* Is the group equivalent to group empty? */ rc = MPI_Group_compare(outgroup, MPI_GROUP_EMPTY, &result); if (result != MPI_IDENT) { errs++; MTestPrintError(rc); printf("Did not create a group equivalent to an empty group\n"); } } rc = MPI_Group_free(&group); if (rc) { errs++; MTestPrintError(rc); } if (outgroup != MPI_GROUP_NULL) { rc = MPI_Group_free(&outgroup); if (rc) { errs++; MTestPrintError(rc); } } MTestFreeComm(&comm); } MTest_Finalize(errs); return MTestReturnValue(errs); }