void Zoltan_RIB_Free_Structure(ZZ *zz) { /* Deallocate the persistent RIB data structures in zz->Structure. */ RIB_STRUCT *rib; /* Data structure for RIB. */ rib = (RIB_STRUCT *) (zz->LB.Data_Structure); if (rib != NULL) { ZOLTAN_FREE(&(rib->Tree_Ptr)); ZOLTAN_FREE(&(rib->Global_IDs)); ZOLTAN_FREE(&(rib->Local_IDs)); Zoltan_Free_And_Reset_Dot_Structure(&rib->Dots); ZOLTAN_FREE(&(zz->LB.Data_Structure)); } }
static int rib_fn( ZZ *zz, /* The Zoltan structure with info for the RIB balancer. */ int *num_import, /* Number of non-local objects assigned to this processor in the new decomposition. When LB.Return_Lists==CANDIDATE_LISTS, num_import returns the number of input objects as given by ZOLTAN_NUM_OBJ_FN. */ ZOLTAN_ID_PTR *import_global_ids, /* Returned value: array of global IDs for non-local objects in this processor's new decomposition. When LB.Return_Lists==CANDIDATE_LISTS, this array contains GIDs for all input objs as given by ZOLTAN_OBJ_LIST_FN.*/ ZOLTAN_ID_PTR *import_local_ids, /* Returned value: array of local IDs for non-local objects in this processor's new decomposition. When LB.Return_Lists==CANDIDATE_LISTS, this array contains LIDs for all input objs as given by ZOLTAN_OBJ_LIST_FN.*/ int **import_procs, /* Returned value: array of processor IDs for processors owning the non-local objects in this processor's new decomposition. When LB.Return_Lists==CANDIDATE_LISTS, the returned array is NULL. */ int **import_to_part, /* Returned value: array of parts to which objects are imported. When LB.Return_Lists==CANDIDATE_LISTS, the returned array is NULL. */ int *num_export, /* Returned value only when LB.Return_Lists==CANDIDATE_LISTS; number of input objs as given by ZOLTAN_NUM_OBJ_FN */ ZOLTAN_ID_PTR *export_global_ids, /* Returned value only when LB.Return_Lists==CANDIDATE_LISTS; for each input obj (from ZOLTAN_OBJ_LIST_FN), return a candidate obj from the part to which the obj is assigned; used in PHG matching */ double overalloc, /* amount to overallocate by when realloc of dot array must be done. 1.0 = no extra; 1.5 = 50% extra; etc. */ int wgtflag, /* No. of weights per dot supplied by user. */ int check_geom, /* Check input & output for consistency? */ int stats, /* Print timing & count summary? */ int gen_tree, /* (0) do not (1) do generate full treept */ int average_cuts, /* (0) don't (1) compute the cut to be the average of the closest dots. */ float *part_sizes /* Input: Array of size zz->Num_Global_Parts * max(zz->Obj_Weight_Dim, 1) containing the percentage of work to be assigned to each part. */ ) { char yo[] = "rib_fn"; int proc,nprocs; /* my proc id, total # of procs */ struct Dot_Struct *dotpt; /* temporary pointer to local dot arrays */ int pdotnum; /* # of dots - decomposition changes it */ int *dotmark = NULL; /* which side of median for each dot */ int dotnum; /* number of dots */ int dotmax = 0; /* max # of dots arrays can hold */ int dottop; /* dots >= this index are new */ int proclower; /* 1st proc in lower set */ int procmid; /* 1st proc in upper set */ int partlower; /* 1st part in lower set */ int partmid; /* 1st part in upper set */ int set; /* which set processor is in = 0/1 */ int old_set; /* set processor was in last cut = 0/1 */ int root; /* part that stores last cut */ int num_procs; /* number of procs in current set */ int num_parts; /* number of parts in current set */ int ierr = ZOLTAN_OK; /* error flag. */ double *value = NULL; /* temp array for median_find */ double *wgts = NULL; /* temp array for serial_rib */ double valuehalf; /* median cut position */ double cm[3]; /* Center of mass of objects */ double evec[3]; /* Eigenvector defining direction */ int first_guess = 0; /* flag if first guess for median search */ int allocflag; /* have to re-allocate space */ double time1=0,time2=0; /* timers */ double time3=0,time4=0; /* timers */ double timestart=0,timestop=0; /* timers */ double timers[4]={0.,0.,0.,0.}; /* diagnostic timers 0 = start-up time before recursion 1 = time before median iterations 2 = time in median iterations 3 = communication time */ ZOLTAN_GNO_TYPE counters[7]; /* diagnostic counts 0 = unused 1 = # of dots sent 2 = # of dots received 3 = most dots this proc ever owns 4 = most dot memory this proc ever allocs 5 = # of times a previous cut is re-used 6 = # of reallocs of dot array */ int i, j; /* local variables */ int use_ids; /* When true, global and local IDs will be stored along with dots in the RCB_STRUCT. When false, storage, manipulation, and communication of IDs is avoided. Set by call to Zoltan_RB_Use_IDs(). */ RIB_STRUCT *rib = NULL; /* Pointer to data structures for RIB */ struct rib_tree *treept = NULL; /* tree of cuts - single cut on exit*/ double start_time, end_time; double lb_time[2]={0,0}; int tfs[2], tmp_tfs[2]; /* added for Tflops_Special; max number of procs and parts over all processors in each iteration (while loop) of parallel partitioning. */ int old_nprocs; /* added for Tflops_Special */ int old_nparts; /* added for Tflops_Special */ double valuelo; /* smallest value of value[i] */ double valuehi; /* largest value of value[i] */ double weight[RB_MAX_WGTS]; /* weight for current set */ double weightlo[RB_MAX_WGTS]; /* weight of lower side of cut */ double weighthi[RB_MAX_WGTS]; /* weight of upper side of cut */ double fractionlo[RB_MAX_WGTS]; /* desired wt in lower half */ int *dotlist = NULL; /* list of dots for find_median. allocated above find_median for better efficiency (don't necessarily have to realloc for each find_median).*/ int rectilinear_blocks = 0; /* parameter for find_median (not used by rib) */ int fp=0; /* first part assigned to this proc. */ int np=0; /* number of parts assigned to this proc. */ int wgtdim; /* max(wgtflag,1) */ int *dindx = NULL, *tmpdindx = NULL; /* MPI data types and user functions */ MPI_Comm local_comm, tmp_comm; int free_comm = FALSE; /* Flag indicating whether MPI_Comm_free should be called on local_comm at end. */ ZOLTAN_TRACE_ENTER(zz, yo); if (stats || (zz->Debug_Level >= ZOLTAN_DEBUG_ATIME)) { MPI_Barrier(zz->Communicator); timestart = time1 = Zoltan_Time(zz->Timer); } /* setup for parallel */ proc = zz->Proc; nprocs = zz->Num_Proc; num_parts = zz->LB.Num_Global_Parts; /* * Determine whether to store, manipulate, and communicate global and * local IDs. */ use_ids = Zoltan_RB_Use_IDs(zz); /* * Build the RIB Data structure and * set pointers to information in it. */ start_time = Zoltan_Time(zz->Timer); ierr = Zoltan_RIB_Build_Structure(zz, &pdotnum, &dotmax, wgtflag, overalloc, use_ids); if (ierr < 0) { ZOLTAN_PRINT_ERROR(proc, yo, "Error returned from Zoltan_RIB_Build_Structure."); goto End; } rib = (RIB_STRUCT *) (zz->LB.Data_Structure); treept = rib->Tree_Ptr; end_time = Zoltan_Time(zz->Timer); lb_time[0] = end_time - start_time; start_time = end_time; /* local copies of calling parameters */ dottop = dotnum = pdotnum; /* initialize timers and counters */ counters[0] = 0; counters[1] = 0; counters[2] = 0; counters[3] = dotnum; counters[4] = dotmax; counters[5] = 0; counters[6] = 0; /* Ensure there are dots */ MPI_Allreduce(&dotnum, &i, 1, MPI_INT, MPI_MAX, zz->Communicator); if (i == 0){ if (proc == 0){ ZOLTAN_PRINT_WARN(proc, yo, "RIB partitioning called with no objects"); } timestart = timestop = 0; goto EndReporting; } /* If using RIB for matching, need to generate candidate lists. * Candidate lists include input GIDs, LIDs as provided by the application. * We need to capture that input here before we move any dots! * We return it in the import lists. * Candidates will be computed after partitioning and returned in the * export lists. */ if (zz->LB.Return_Lists == ZOLTAN_LB_CANDIDATE_LISTS) { ierr = Zoltan_RB_Candidates_Copy_Input(zz, dotnum, rib->Global_IDs, rib->Local_IDs, &rib->Dots, num_import, import_global_ids, import_local_ids, import_procs, import_to_part); if (ierr < 0) { ZOLTAN_PRINT_ERROR(proc,yo, "Error returned from Zoltan_RB_Return_Arguments."); goto End; } } /* create mark and list arrays for dots */ allocflag = 0; if (dotmax > 0) { if (!(dotmark = (int *) ZOLTAN_MALLOC(dotmax*sizeof(int))) || !(value = (double *) ZOLTAN_MALLOC(dotmax*sizeof(double))) || !(dotlist = (int *) ZOLTAN_MALLOC(dotmax*sizeof(int)))) { ierr = ZOLTAN_MEMERR; goto End; } } else { dotmark = NULL; value = NULL; dotlist = NULL; } /* set dot weights = 1.0 if user didn't and determine total weight */ dotpt = &rib->Dots; if (dotpt->nWeights == 0) { weightlo[0] = (double) dotnum; dotpt->uniformWeight = 1.0; wgtdim = 1; } else { double *wgt; for (j=0; j<dotpt->nWeights; j++){ weightlo[j] = 0.0; wgt = dotpt->Weight + j; for (i=0; i < dotnum; i++){ weightlo[j] += *wgt; wgt += dotpt->nWeights; } } wgtdim = dotpt->nWeights; } MPI_Allreduce(weightlo, weight, wgtdim, MPI_DOUBLE, MPI_SUM, zz->Communicator); if (check_geom) { ierr = Zoltan_RB_check_geom_input(zz, dotpt, dotnum); if (ierr < 0) { ZOLTAN_PRINT_ERROR(proc, yo, "Error returned from Zoltan_RB_check_geom_input"); goto End; } } /* create local communicator for use in recursion */ if (zz->Tflops_Special) local_comm = zz->Communicator; else { MPI_Comm_dup(zz->Communicator,&local_comm); free_comm = TRUE; } if (stats || (zz->Debug_Level >= ZOLTAN_DEBUG_ATIME)) { time2 = Zoltan_Time(zz->Timer); timers[0] = time2 - time1; } /* recursively halve until just one part or proc in set */ old_nprocs = num_procs = nprocs; old_nparts = num_parts; partlower = 0; root = 0; old_set = 1; ierr = Zoltan_LB_Proc_To_Part(zz, proc, &np, &fp); for (i = fp; i < (fp + np); i++) { treept[i].parent = 0; treept[i].left_leaf = 0; } if (zz->Tflops_Special) { proclower = 0; tfs[0] = nprocs; tfs[1] = num_parts; } while ((num_parts > 1 && num_procs > 1) || (zz->Tflops_Special && tfs[0] > 1 && tfs[1] > 1)) { ierr = Zoltan_Divide_Machine(zz, zz->Obj_Weight_Dim, part_sizes, proc, local_comm, &set, &proclower, &procmid, &num_procs, &partlower, &partmid, &num_parts, fractionlo); if (ierr < 0) { ZOLTAN_PRINT_ERROR(proc, yo, "Error in Zoltan_Divide_Machine."); goto End; } /* tfs[0] is max number of processors in all sets over all processors - * tfs[1] is max number of parts in all sets over all processors - * force all processors to go through all levels of parallel rib */ if (zz->Tflops_Special) { tmp_tfs[0] = num_procs; tmp_tfs[1] = num_parts; MPI_Allreduce(tmp_tfs, tfs, 2, MPI_INT, MPI_MAX, local_comm); } /* create mark array and active list for dots */ if (allocflag) { allocflag = 0; ZOLTAN_FREE(&dotmark); ZOLTAN_FREE(&value); ZOLTAN_FREE(&dotlist); if (!(dotmark = (int *) ZOLTAN_MALLOC(dotmax*sizeof(int))) || !(value = (double *) ZOLTAN_MALLOC(dotmax*sizeof(double))) || !(dotlist = (int *) ZOLTAN_MALLOC(dotmax*sizeof(int)))) { ierr = ZOLTAN_MEMERR; goto End; } } dotpt = &rib->Dots; if (old_nparts > 1 && old_nprocs > 1) { /* test added for Tflops_Special; compute values only if looping to decompose, not if looping to keep Tflops_Special happy. */ ierr = compute_rib_direction(zz, zz->Tflops_Special, rib->Num_Geom, &valuelo, &valuehi, dotpt, NULL, dotnum, wgtflag, cm, evec, value, local_comm, proc, old_nprocs, proclower); if (ierr < 0) { ZOLTAN_PRINT_ERROR(proc, yo, "Error returned from compute_rib_direction"); goto End; } } else { /* For Tflops_Special: initialize value when looping only for Tflops_Special */ for (i = 0; i < dotmax; i++) value[i] = 0.0; valuelo = valuehi = 0.0; } if (stats || (zz->Debug_Level >= ZOLTAN_DEBUG_ATIME)) time2 = Zoltan_Time(zz->Timer); if (!Zoltan_RB_find_median( zz->Tflops_Special, value, dotpt->Weight, dotpt->uniformWeight, dotmark, dotnum, proc, fractionlo, local_comm, &valuehalf, first_guess, nprocs, old_nprocs, proclower, old_nparts, wgtflag, valuelo, valuehi, weight[0], weightlo, weighthi, dotlist, rectilinear_blocks, average_cuts)) { ZOLTAN_PRINT_ERROR(proc, yo, "Error returned from Zoltan_RB_find_median."); ierr = ZOLTAN_FATAL; goto End; } if (set) /* set weight for current part */ for (j=0; j<wgtdim; j++) weight[j] = weighthi[j]; else for (j=0; j<wgtdim; j++) weight[j] = weightlo[j]; if (stats || (zz->Debug_Level >= ZOLTAN_DEBUG_ATIME)) time3 = Zoltan_Time(zz->Timer); /* store cut info in tree only if proc "owns" partmid */ /* test of partmid > 0 prevents treept[0] being set when this cut is only removing low-numbered processors (proclower to procmid-1) that have no parts in them from the processors remaining to be partitioned. */ if (partmid > 0 && partmid == fp) { treept[partmid].cm[0] = cm[0]; treept[partmid].cm[1] = cm[1]; treept[partmid].cm[2] = cm[2]; treept[partmid].ev[0] = evec[0]; treept[partmid].ev[1] = evec[1]; treept[partmid].ev[2] = evec[2]; treept[partmid].cut = valuehalf; treept[partmid].parent = old_set ? -(root+1) : root+1; /* The following two will get overwritten when the information is assembled if this is not a terminal cut */ treept[partmid].left_leaf = -partlower; treept[partmid].right_leaf = -partmid; } if (old_nprocs > 1 && partmid > 0 && partmid != partlower + old_nparts) { /* old_nprocs > 1 test: Don't reset these values if proc is in loop only * because of other procs for Tflops_Special. * partmid > 0 test: Don't reset these values if low-numbered processors * (proclower to procmid-1) have zero parts and this cut is removing * them from the processors remaining to be partitioned. * partmid != partlower + old_nparts test: Don't reset these values if * cut is removing high-numbered processors with zero parts from * the processors remaining to be partitioned. */ old_set = set; root = partmid; } ierr = Zoltan_RB_Send_Outgoing(zz, &(rib->Global_IDs), &(rib->Local_IDs), &(rib->Dots), &dotmark, &dottop, &dotnum, &dotmax, set, &allocflag, overalloc, stats, counters, use_ids, local_comm, proclower, old_nprocs, partlower, partmid); if (ierr < 0) { ZOLTAN_PRINT_ERROR(proc, yo, "Error returned from Zoltan_RB_Send_Outgoing."); goto End; } /* create new communicators */ if (zz->Tflops_Special) { if (set) { proclower = procmid; partlower = partmid; } old_nprocs = num_procs; old_nparts = num_parts; } else { if (set) partlower = partmid; MPI_Comm_split(local_comm,set,proc,&tmp_comm); MPI_Comm_free(&local_comm); local_comm = tmp_comm; old_nprocs = num_procs; old_nparts = num_parts; } if (stats || (zz->Debug_Level >= ZOLTAN_DEBUG_ATIME)) { time4 = Zoltan_Time(zz->Timer); timers[1] += time2 - time1; timers[2] += time3 - time2; timers[3] += time4 - time3; } } /* have recursed all the way to a single processor sub-domain */ /* Send dots to correct processors for their parts. This is needed most notably when a processor has zero parts on it, but still has some dots after the parallel partitioning. */ ierr = Zoltan_RB_Send_To_Part(zz, &(rib->Global_IDs), &(rib->Local_IDs), &(rib->Dots), &dotmark, &dottop, &dotnum, &dotmax, &allocflag, overalloc, stats, counters, use_ids); if (ierr < 0) { ZOLTAN_PRINT_ERROR(zz->Proc, yo, "Error returned from Zoltan_RB_Send_To_Part"); goto End; } /* All dots are now on the processors they will end up on; now generate * more parts if needed. */ if (num_parts > 1) { if (dotpt->nWeights) wgts = (double *) ZOLTAN_MALLOC(dotpt->nWeights * dotnum * sizeof(double)); dindx = (int *) ZOLTAN_MALLOC(dotnum * 2 * sizeof(int)); tmpdindx = dindx + dotnum; if (allocflag) { ZOLTAN_FREE(&dotmark); ZOLTAN_FREE(&value); ZOLTAN_FREE(&dotlist); if (!(dotmark = (int *) ZOLTAN_MALLOC(dotmax*sizeof(int))) || !(value = (double *) ZOLTAN_MALLOC(dotmax*sizeof(double))) || !(dotlist = (int *) ZOLTAN_MALLOC(dotmax*sizeof(int)))) { ZOLTAN_PRINT_ERROR(proc, yo, "Memory error."); ierr = ZOLTAN_MEMERR; goto End; } } for (i = 0; i < dotnum; i++) dindx[i] = i; ierr = serial_rib(zz, &rib->Dots, dotmark, dotlist, old_set, root, rib->Num_Geom, weight[0], dotnum, num_parts, &(dindx[0]), &(tmpdindx[0]), partlower, proc, wgtflag, stats, gen_tree, rectilinear_blocks, average_cuts, treept, value, wgts, part_sizes); if (ierr < 0) { ZOLTAN_PRINT_ERROR(proc, yo, "Error returned from serial_rib"); goto End; } ZOLTAN_FREE(&wgts); } end_time = Zoltan_Time(zz->Timer); lb_time[1] = end_time - start_time; if (stats || (zz->Debug_Level >= ZOLTAN_DEBUG_ATIME)) { MPI_Barrier(zz->Communicator); timestop = time1 = Zoltan_Time(zz->Timer); } /* error checking and statistics */ if (check_geom) { ierr = Zoltan_RB_check_geom_output(zz, &rib->Dots, part_sizes, np, fp, dotnum, pdotnum, NULL); if (ierr < 0) { ZOLTAN_PRINT_ERROR(proc, yo, "Error returned from Zoltan_RB_check_geom_output"); goto End; } } EndReporting: /* update calling routine parameters */ start_time = Zoltan_Time(zz->Timer); pdotnum = dotnum; /* Perform remapping (if requested) */ if (zz->LB.Remap_Flag) { ierr = Zoltan_RB_Remap(zz, &(rib->Global_IDs), &(rib->Local_IDs), &(rib->Dots), &dotnum, &dotmax, &allocflag, overalloc, stats, counters, use_ids); /* Note: dottop is no longer valid after remapping. Remapping might destroy the nice local-followed-by-non-local ordering of the dots array. Do not use dottop after remapping. */ if (ierr < 0) { ZOLTAN_PRINT_ERROR(zz->Proc, yo, "Error returned from Zoltan_RB_Remap."); goto End; } } /* build return arguments */ if (zz->LB.Return_Lists != ZOLTAN_LB_NO_LISTS && zz->LB.Return_Lists != ZOLTAN_LB_CANDIDATE_LISTS) { /* zz->LB.Return_Lists is true ==> use_ids is true */ ierr = Zoltan_RB_Return_Arguments(zz, rib->Global_IDs, rib->Local_IDs, &rib->Dots, num_import, import_global_ids, import_local_ids, import_procs, import_to_part, dotnum); if (ierr < 0) { ZOLTAN_PRINT_ERROR(proc, yo, "Error returned from Zoltan_RB_Return_Arguments."); goto End; } } else if (zz->LB.Return_Lists == ZOLTAN_LB_CANDIDATE_LISTS) { /* Select a candidate for each part and return it in the export_GIDs. */ ierr = Zoltan_RB_Candidates_Output(zz, dotnum, dindx, rib->Global_IDs, rib->Local_IDs, &rib->Dots, *num_import, *import_global_ids, num_export, export_global_ids); if (ierr < 0) { ZOLTAN_PRINT_ERROR(proc,yo, "Error returned from Zoltan_RB_Return_Candidates."); goto End; } } ZOLTAN_FREE(&dindx); if (gen_tree) { int *displ, *recvcount; int sendcount; struct rib_tree *treetmp = NULL; /* temporary tree of cuts; used to keep valgrind from reporting overlapped memory in MPI_Allgatherv */ treetmp = (struct rib_tree *) ZOLTAN_MALLOC(zz->LB.Num_Global_Parts* sizeof(struct rib_tree)); displ = (int *) ZOLTAN_MALLOC(2 * zz->Num_Proc * sizeof(int)); if (!displ || !treetmp) { ZOLTAN_PRINT_ERROR(zz->Proc, yo, "Memory error."); ierr = ZOLTAN_MEMERR; goto End; } recvcount = displ + zz->Num_Proc; ierr = Zoltan_RB_Tree_Gatherv(zz, sizeof(struct rib_tree), &sendcount, recvcount, displ); /* * Create copy of treept so that MPI_Allgatherv doesn't use same * memory for sending and receiving; removes valgrind warning. */ for (i = 0; i < zz->LB.Num_Global_Parts; i++) treetmp[i] = treept[i]; MPI_Allgatherv(&treetmp[fp], sendcount, MPI_BYTE, treept, recvcount, displ, MPI_BYTE, zz->Communicator); for (i = 1; i < zz->LB.Num_Global_Parts; i++){ if (treept[i].parent > 0) treept[treept[i].parent - 1].left_leaf = i; else if (treept[i].parent < 0) treept[-treept[i].parent - 1].right_leaf = i; } ZOLTAN_FREE(&displ); ZOLTAN_FREE(&treetmp); } else { treept[0].right_leaf = -1; } if (zz->Debug_Level >= ZOLTAN_DEBUG_ALL) print_rib_tree(zz, np, fp, &(treept[fp])); end_time = Zoltan_Time(zz->Timer); lb_time[0] += (end_time - start_time); if (stats || (zz->Debug_Level >= ZOLTAN_DEBUG_ATIME)) Zoltan_RB_stats(zz, timestop-timestart, &rib->Dots, dotnum, part_sizes, timers, counters, stats, NULL, NULL, FALSE); if (zz->Debug_Level >= ZOLTAN_DEBUG_ATIME) { if (zz->Proc == zz->Debug_Proc) printf("ZOLTAN RIB Times: \n"); Zoltan_Print_Stats(zz->Communicator, zz->Debug_Proc, lb_time[0], "ZOLTAN Build: "); Zoltan_Print_Stats(zz->Communicator, zz->Debug_Proc, lb_time[1], "ZOLTAN RIB: "); } if (zz->Debug_Level >= ZOLTAN_DEBUG_ALL) { /* zz->Debug_Level >= ZOLTAN_DEBUG_ALL ==> use_ids is true */ Zoltan_RB_Print_All(zz, rib->Global_IDs, &rib->Dots, dotnum, *num_import, *import_global_ids, *import_procs); } End: /* Free memory allocated by the algorithm. */ if (free_comm) MPI_Comm_free(&local_comm); ZOLTAN_FREE(&dotmark); ZOLTAN_FREE(&value); ZOLTAN_FREE(&dotlist); if (!gen_tree && /* don't need parts */ rib && (rib->Tran.Target_Dim < 0)) { /* don't need transformation */ /* Free all memory used. */ Zoltan_RIB_Free_Structure(zz); } else if (rib != NULL) { /* Free only Dots and IDs; keep other structures. */ ZOLTAN_FREE(&(rib->Global_IDs)); ZOLTAN_FREE(&(rib->Local_IDs)); Zoltan_Free_And_Reset_Dot_Structure(&(rib->Dots)); } ZOLTAN_TRACE_EXIT(zz, yo); return(ierr); }