void AlgPTScotch<Adapter>::partition(
const RCP<PartitioningSolution<Adapter> > &solution
)
{
HELLO;
size_t numGlobalParts = solution->getTargetGlobalNumberOfParts();
SCOTCH_Num partnbr=0;
TPL_Traits<SCOTCH_Num, size_t>::ASSIGN_TPL_T(partnbr, numGlobalParts, env);
#ifdef HAVE_ZOLTAN2_MPI
int ierr = 0;
int me = problemComm->getRank();
const SCOTCH_Num baseval = 0; // Base value for array indexing.
// GraphModel returns GNOs from base 0.
SCOTCH_Strat stratstr; // Strategy string
// TODO: Set from parameters
SCOTCH_stratInit(&stratstr);
// Allocate and initialize PTScotch Graph data structure.
SCOTCH_Dgraph *gr = SCOTCH_dgraphAlloc(); // Scotch distributed graph
ierr = SCOTCH_dgraphInit(gr, mpicomm);
env->globalInputAssertion(__FILE__, __LINE__, "SCOTCH_dgraphInit",
!ierr, BASIC_ASSERTION, problemComm);
// Get vertex info
ArrayView<const gno_t> vtxID;
ArrayView<StridedData<lno_t, scalar_t> > xyz;
ArrayView<StridedData<lno_t, scalar_t> > vwgts;
size_t nVtx = model->getVertexList(vtxID, xyz, vwgts);
SCOTCH_Num vertlocnbr=0;
TPL_Traits<SCOTCH_Num, size_t>::ASSIGN_TPL_T(vertlocnbr, nVtx, env);
SCOTCH_Num vertlocmax = vertlocnbr; // Assumes no holes in global nums.
// Get edge info
ArrayView<const gno_t> edgeIds;
ArrayView<const int> procIds;
ArrayView<const lno_t> offsets;
ArrayView<StridedData<lno_t, scalar_t> > ewgts;
size_t nEdge = model->getEdgeList(edgeIds, procIds, offsets, ewgts);
SCOTCH_Num edgelocnbr=0;
TPL_Traits<SCOTCH_Num, size_t>::ASSIGN_TPL_T(edgelocnbr, nEdge, env);
const SCOTCH_Num edgelocsize = edgelocnbr; // Assumes adj array is compact.
SCOTCH_Num *vertloctab; // starting adj/vtx
TPL_Traits<SCOTCH_Num, lno_t>::ASSIGN_TPL_T_ARRAY(&vertloctab, offsets, env);
SCOTCH_Num *edgeloctab; // adjacencies
TPL_Traits<SCOTCH_Num, gno_t>::ASSIGN_TPL_T_ARRAY(&edgeloctab, edgeIds, env);
// We don't use these arrays, but we need them as arguments to Scotch.
SCOTCH_Num *vendloctab = NULL; // Assume consecutive storage for adj
SCOTCH_Num *vlblloctab = NULL; // Vertex label array
SCOTCH_Num *edgegsttab = NULL; // Array for ghost vertices
// Get weight info.
SCOTCH_Num *velotab = NULL; // Vertex weights
SCOTCH_Num *edlotab = NULL; // Edge weights
int nVwgts = model->getNumWeightsPerVertex();
int nEwgts = model->getNumWeightsPerEdge();
if (nVwgts > 1 && me == 0) {
std::cerr << "Warning: NumWeightsPerVertex is " << nVwgts
<< " but Scotch allows only one weight. "
<< " Zoltan2 will use only the first weight per vertex."
<< std::endl;
}
if (nEwgts > 1 && me == 0) {
std::cerr << "Warning: NumWeightsPerEdge is " << nEwgts
<< " but Scotch allows only one weight. "
<< " Zoltan2 will use only the first weight per edge."
<< std::endl;
}
if (nVwgts) {
velotab = new SCOTCH_Num[nVtx+1]; // +1 since Scotch wants all procs
// to have non-NULL arrays
scale_weights(nVtx, vwgts[0], velotab);
}
if (nEwgts) {
edlotab = new SCOTCH_Num[nEdge+1]; // +1 since Scotch wants all procs
// to have non-NULL arrays
scale_weights(nEdge, ewgts[0], edlotab);
}
// Build PTScotch distributed data structure
ierr = SCOTCH_dgraphBuild(gr, baseval, vertlocnbr, vertlocmax,
vertloctab, vendloctab, velotab, vlblloctab,
edgelocnbr, edgelocsize,
edgeloctab, edgegsttab, edlotab);
env->globalInputAssertion(__FILE__, __LINE__, "SCOTCH_dgraphBuild",
!ierr, BASIC_ASSERTION, problemComm);
// Create array for Scotch to return results in.
ArrayRCP<part_t> partList(new part_t[nVtx], 0, nVtx,true);
SCOTCH_Num *partloctab = NULL;
if (nVtx && (sizeof(SCOTCH_Num) == sizeof(part_t))) {
// Can write directly into the solution's memory
partloctab = (SCOTCH_Num *) partList.getRawPtr();
}
else {
// Can't use solution memory directly; will have to copy later.
// Note: Scotch does not like NULL arrays, so add 1 to always have non-null.
// ParMETIS has this same "feature." See Zoltan bug 4299.
partloctab = new SCOTCH_Num[nVtx+1];
}
// Get target part sizes
float *partsizes = new float[numGlobalParts];
if (!solution->criteriaHasUniformPartSizes(0))
for (size_t i=0; i<numGlobalParts; i++)
partsizes[i] = solution->getCriteriaPartSize(0, i);
else
for (size_t i=0; i<numGlobalParts; i++)
partsizes[i] = 1.0 / float(numGlobalParts);
// Allocate and initialize PTScotch target architecture data structure
SCOTCH_Arch archdat;
SCOTCH_archInit(&archdat);
SCOTCH_Num velosum = 0;
SCOTCH_dgraphSize (gr, &velosum, NULL, NULL, NULL);
SCOTCH_Num *goalsizes = new SCOTCH_Num[partnbr];
// TODO: The goalsizes are set as in Zoltan; not sure it is correct there
// or here.
// It appears velosum is global NUMBER of vertices, not global total
// vertex weight. I think we should use the latter.
// Fix this when we add vertex weights.
for (SCOTCH_Num i = 0; i < partnbr; i++)
goalsizes[i] = SCOTCH_Num(ceil(velosum * partsizes[i]));
delete [] partsizes;
SCOTCH_archCmpltw(&archdat, partnbr, goalsizes);
// Call partitioning; result returned in partloctab.
ierr = SCOTCH_dgraphMap(gr, &archdat, &stratstr, partloctab);
env->globalInputAssertion(__FILE__, __LINE__, "SCOTCH_dgraphMap",
!ierr, BASIC_ASSERTION, problemComm);
SCOTCH_archExit(&archdat);
delete [] goalsizes;
// TODO - metrics
#ifdef SHOW_ZOLTAN2_SCOTCH_MEMORY
int me = env->comm_->getRank();
#endif
#ifdef HAVE_SCOTCH_ZOLTAN2_GETMEMORYMAX
if (me == 0){
size_t scotchBytes = SCOTCH_getMemoryMax();
std::cout << "Rank " << me << ": Maximum bytes used by Scotch: ";
std::cout << scotchBytes << std::endl;
}
#endif
// Clean up PTScotch
SCOTCH_dgraphExit(gr);
free(gr);
SCOTCH_stratExit(&stratstr);
// Load answer into the solution.
if ((sizeof(SCOTCH_Num) != sizeof(part_t)) || (nVtx == 0)) {
for (size_t i = 0; i < nVtx; i++) partList[i] = partloctab[i];
delete [] partloctab;
}
solution->setParts(partList);
env->memory("Zoltan2-Scotch: After creating solution");
// Clean up copies made due to differing data sizes.
TPL_Traits<SCOTCH_Num, lno_t>::DELETE_TPL_T_ARRAY(&vertloctab);
TPL_Traits<SCOTCH_Num, gno_t>::DELETE_TPL_T_ARRAY(&edgeloctab);
if (nVwgts) delete [] velotab;
if (nEwgts) delete [] edlotab;
#else // DO NOT HAVE_MPI
// TODO: Handle serial case with calls to Scotch.
// TODO: For now, assign everything to rank 0 and assume only one part.
// TODO: Can probably use the code above for loading solution,
// TODO: instead of duplicating it here.
// TODO
// TODO: Actual logic should call Scotch when number of processes == 1.
ArrayView<const gno_t> vtxID;
ArrayView<StridedData<lno_t, scalar_t> > xyz;
ArrayView<StridedData<lno_t, scalar_t> > vwgts;
size_t nVtx = model->getVertexList(vtxID, xyz, vwgts);
ArrayRCP<part_t> partList(new part_t[nVtx], 0, nVtx, true);
for (size_t i = 0; i < nVtx; i++) partList[i] = 0;
solution->setParts(partList);
#endif // DO NOT HAVE_MPI
}
void AlgPTScotch(
const RCP<const Environment> &env, // parameters & app comm
const RCP<const Comm<int> > &problemComm, // problem comm
#ifdef HAVE_ZOLTAN2_MPI
MPI_Comm mpicomm,
#endif
const RCP<GraphModel<typename Adapter::base_adapter_t> > &model, // the graph
RCP<PartitioningSolution<Adapter> > &solution
)
{
HELLO;
typedef typename Adapter::lno_t lno_t;
typedef typename Adapter::gno_t gno_t;
typedef typename Adapter::scalar_t scalar_t;
int ierr = 0;
size_t numGlobalParts = solution->getTargetGlobalNumberOfParts();
SCOTCH_Num partnbr;
SCOTCH_Num_Traits<size_t>::ASSIGN_TO_SCOTCH_NUM(partnbr, numGlobalParts, env);
#ifdef HAVE_ZOLTAN2_MPI
const SCOTCH_Num baseval = 0; // Base value for array indexing.
// GraphModel returns GNOs from base 0.
SCOTCH_Strat stratstr; // Strategy string
// TODO: Set from parameters
SCOTCH_stratInit(&stratstr);
// Allocate & initialize PTScotch data structure.
SCOTCH_Dgraph *gr = SCOTCH_dgraphAlloc(); // Scotch distributed graph
ierr = SCOTCH_dgraphInit(gr, mpicomm);
env->globalInputAssertion(__FILE__, __LINE__, "SCOTCH_dgraphInit",
!ierr, BASIC_ASSERTION, problemComm);
// Get vertex info
ArrayView<const gno_t> vtxID;
ArrayView<StridedData<lno_t, scalar_t> > xyz;
ArrayView<StridedData<lno_t, scalar_t> > vtxWt;
size_t nVtx = model->getVertexList(vtxID, xyz, vtxWt);
SCOTCH_Num vertlocnbr;
SCOTCH_Num_Traits<size_t>::ASSIGN_TO_SCOTCH_NUM(vertlocnbr, nVtx, env);
SCOTCH_Num vertlocmax = vertlocnbr; // Assumes no holes in global nums.
// Get edge info
ArrayView<const gno_t> edgeIds;
ArrayView<const int> procIds;
ArrayView<const lno_t> offsets;
ArrayView<StridedData<lno_t, scalar_t> > ewgts;
size_t nEdges = model->getEdgeList(edgeIds, procIds, offsets, ewgts);
SCOTCH_Num edgelocnbr;
SCOTCH_Num_Traits<size_t>::ASSIGN_TO_SCOTCH_NUM(edgelocnbr, nEdges, env);
const SCOTCH_Num edgelocsize = edgelocnbr; // Assumes adj array is compact.
SCOTCH_Num *vertloctab; // starting adj/vtx
SCOTCH_Num_Traits<lno_t>::ASSIGN_SCOTCH_NUM_ARRAY(&vertloctab, offsets, env);
SCOTCH_Num *edgeloctab; // adjacencies
SCOTCH_Num_Traits<gno_t>::ASSIGN_SCOTCH_NUM_ARRAY(&edgeloctab, edgeIds, env);
// We don't use these arrays, but we need them as arguments to Scotch.
SCOTCH_Num *vendloctab = NULL; // Assume consecutive storage for adj
SCOTCH_Num *vlblloctab = NULL; // Vertex label array
SCOTCH_Num *edgegsttab = NULL; // Array for ghost vertices
// Get weight info.
// TODO: Actually get the weights; for now, not using weights.
SCOTCH_Num *veloloctab = NULL; // Vertex weights
SCOTCH_Num *edloloctab = NULL; // Edge weights
//TODO int vwtdim = model->getVertexWeightDim();
//TODO int ewtdim = model->getEdgeWeightDim();
//TODO if (vwtdim) veloloctab = new SCOTCH_Num[nVtx];
//TODO if (ewtdim) edloloctab = new SCOTCH_Num[nEdges];
//TODO scale weights to SCOTCH_Nums.
// Build PTScotch distributed data structure
ierr = SCOTCH_dgraphBuild(gr, baseval, vertlocnbr, vertlocmax,
vertloctab, vendloctab, veloloctab, vlblloctab,
edgelocnbr, edgelocsize,
edgeloctab, edgegsttab, edloloctab);
env->globalInputAssertion(__FILE__, __LINE__, "SCOTCH_dgraphBuild",
!ierr, BASIC_ASSERTION, problemComm);
// Create array for Scotch to return results in.
ArrayRCP<partId_t> partList(new partId_t [nVtx], 0, nVtx,true);
SCOTCH_Num *partloctab = NULL;
if (nVtx && (sizeof(SCOTCH_Num) == sizeof(partId_t))) {
// Can write directly into the solution's memory
partloctab = (SCOTCH_Num *) partList.getRawPtr();
}
else {
// Can't use solution memory directly; will have to copy later.
// Note: Scotch does not like NULL arrays, so add 1 to always have non-null.
// ParMETIS has this same "feature." See Zoltan bug 4299.
partloctab = new SCOTCH_Num[nVtx+1];
}
// Call partitioning; result returned in partloctab.
// TODO: Use SCOTCH_dgraphMap so can include a machine model in partitioning
ierr = SCOTCH_dgraphPart(gr, partnbr, &stratstr, partloctab);
env->globalInputAssertion(__FILE__, __LINE__, "SCOTCH_dgraphPart",
!ierr, BASIC_ASSERTION, problemComm);
// TODO - metrics
#ifdef SHOW_ZOLTAN2_SCOTCH_MEMORY
int me = env->comm_->getRank();
#endif
#ifdef HAVE_SCOTCH_ZOLTAN2_GETMEMORYMAX
if (me == 0){
size_t scotchBytes = SCOTCH_getMemoryMax();
std::cout << "Rank " << me << ": Maximum bytes used by Scotch: ";
std::cout << scotchBytes << std::endl;
}
#endif
// Clean up PTScotch
SCOTCH_dgraphExit(gr);
delete gr;
SCOTCH_stratExit(&stratstr);
// Load answer into the solution.
if ((sizeof(SCOTCH_Num) != sizeof(partId_t)) || (nVtx == 0)) {
for (size_t i = 0; i < nVtx; i++) partList[i] = partloctab[i];
delete [] partloctab;
}
ArrayRCP<const gno_t> gnos = arcpFromArrayView(vtxID);
solution->setParts(gnos, partList, true);
env->memory("Zoltan2-Scotch: After creating solution");
//if (me == 0) cout << " done." << endl;
// Clean up Zoltan2
//TODO if (vwtdim) delete [] velotab;
//TODO if (ewtdim) delete [] edlotab;
// Clean up copies made due to differing data sizes.
SCOTCH_Num_Traits<lno_t>::DELETE_SCOTCH_NUM_ARRAY(&vertloctab);
SCOTCH_Num_Traits<gno_t>::DELETE_SCOTCH_NUM_ARRAY(&edgeloctab);
#else // DO NOT HAVE_MPI
// TODO: Handle serial case with calls to Scotch.
// TODO: For now, assign everything to rank 0 and assume only one part.
// TODO: Can probably use the code above for loading solution,
// TODO: instead of duplicating it here.
// TODO
// TODO: Actual logic should call Scotch when number of processes == 1.
ArrayView<const gno_t> vtxID;
ArrayView<StridedData<lno_t, scalar_t> > xyz;
ArrayView<StridedData<lno_t, scalar_t> > vtxWt;
size_t nVtx = model->getVertexList(vtxID, xyz, vtxWt);
for (size_t i = 0; i < nVtx; i++) partList[i] = 0;
#endif // DO NOT HAVE_MPI
}
