/*************************************************************************
* This function implements a simple graph adaption strategy.
**************************************************************************/
void AdaptGraph2(graph_t *graph, idx_t afactor, MPI_Comm comm)
{
idx_t i, j, k, nvtxs, firstvtx, lastvtx;
idx_t npes, mype, mypwgt, max, min, sum;
idx_t *vwgt, *xadj, *adjncy, *adjwgt;
gkMPI_Comm_size(comm, &npes);
gkMPI_Comm_rank(comm, &mype);
srand(mype*afactor);
nvtxs = graph->nvtxs;
xadj = graph->xadj;
adjncy = graph->adjncy;
if (graph->adjwgt == NULL)
adjwgt = graph->adjwgt = ismalloc(graph->nedges, 1, "AdaptGraph: adjwgt");
else
adjwgt = graph->adjwgt;
vwgt = graph->vwgt;
firstvtx = graph->vtxdist[mype];
lastvtx = graph->vtxdist[mype+1];
/* if (RandomInRange(npes+1) < .05*npes) { */
if (RandomInRange(npes+1) < 2) {
printf("[%"PRIDX"] is adapting\n", mype);
for (i=0; i<nvtxs; i++)
vwgt[i] = afactor*vwgt[i];
}
for (i=0; i<nvtxs; i++) {
for (j=xadj[i]; j<xadj[i+1]; j++) {
k = adjncy[j];
if (k >= firstvtx && k < lastvtx) {
adjwgt[j] = (int)pow(1.0*(gk_min(vwgt[i],vwgt[k-firstvtx])), .6667);
if (adjwgt[j] == 0)
adjwgt[j] = 1;
}
}
}
mypwgt = isum(nvtxs, vwgt, 1);
gkMPI_Allreduce((void *)&mypwgt, (void *)&max, 1, IDX_T, MPI_MAX, comm);
gkMPI_Allreduce((void *)&mypwgt, (void *)&min, 1, IDX_T, MPI_MIN, comm);
gkMPI_Allreduce((void *)&mypwgt, (void *)&sum, 1, IDX_T, MPI_SUM, comm);
if (mype == 0)
printf("Initial Load Imbalance: %5.4"PRREAL", [%5"PRIDX" %5"PRIDX" %5"PRIDX"]\n", (1.0*max*npes)/(1.0*sum), min, max, sum);
}
int getTopSimUsers(gk_csr_t *adjMat, int user, gk_fkv_t *topUsers, int nsim) {
//to store the restart distributon of vertices
float *pr;
int i, j, count;
int iter = 0;
gk_fkv_t *pRanks;
pr = (float*) malloc(sizeof(float)*adjMat->nrows);
//initialize the restart distribution for user
pr[user] = 1.0;
iter = gk_rw_PageRank(adjMat, 0.5, 0.000001, 100, pr);
fprintf(stderr, "Iter: %d\n", iter);
count = 0;
//count the non-zero values in pr and sort them
for (i = 0; i < adjMat->nrows; i++) {
if (pr[i] > 0 && i != user) {
count++;
}
}
//printf("Found %d similar users.\n", count);
pRanks = gk_fkvmalloc(count, "store page ranks");
for (i = 0, j = 0; i < adjMat->nrows; i++) {
if (pr[i] > 0 && i != user) {
pRanks[j].key = pr[i];
pRanks[j].val = i;
j++;
}
}
nsim = gk_min(nsim, count);
gk_dfkvkselect(count, nsim, pRanks);
gk_fkvsortd(nsim, pRanks);
gk_fkvcopy(nsim, pRanks, topUsers);
free(pr);
free(pRanks);
return nsim;
}
/*************************************************************************
* This function is the entry point of the initial partition algorithm
* that does recursive bissection.
* This algorithm assembles the graph to all the processors and preceeds
* by parallelizing the recursive bisection step.
**************************************************************************/
void InitPartition(ctrl_t *ctrl, graph_t *graph)
{
idx_t i, j, ncon, mype, npes, gnvtxs, ngroups;
idx_t *xadj, *adjncy, *adjwgt, *vwgt;
idx_t *part, *gwhere0, *gwhere1;
idx_t *tmpwhere, *tmpvwgt, *tmpxadj, *tmpadjncy, *tmpadjwgt;
graph_t *agraph;
idx_t lnparts, fpart, fpe, lnpes;
idx_t twoparts=2, moptions[METIS_NOPTIONS], edgecut, max_cut;
real_t *tpwgts, *tpwgts2, *lbvec, lbsum, min_lbsum, wsum;
MPI_Comm ipcomm;
struct {
double sum;
int rank;
} lpesum, gpesum;
WCOREPUSH;
ncon = graph->ncon;
ngroups = gk_max(gk_min(RIP_SPLIT_FACTOR, ctrl->npes), 1);
IFSET(ctrl->dbglvl, DBG_TIME, gkMPI_Barrier(ctrl->comm));
IFSET(ctrl->dbglvl, DBG_TIME, starttimer(ctrl->InitPartTmr));
lbvec = rwspacemalloc(ctrl, ncon);
/* assemble the graph to all the processors */
agraph = AssembleAdaptiveGraph(ctrl, graph);
gnvtxs = agraph->nvtxs;
/* make a copy of the graph's structure for later */
xadj = icopy(gnvtxs+1, agraph->xadj, iwspacemalloc(ctrl, gnvtxs+1));
vwgt = icopy(gnvtxs*ncon, agraph->vwgt, iwspacemalloc(ctrl, gnvtxs*ncon));
adjncy = icopy(agraph->nedges, agraph->adjncy, iwspacemalloc(ctrl, agraph->nedges));
adjwgt = icopy(agraph->nedges, agraph->adjwgt, iwspacemalloc(ctrl, agraph->nedges));
part = iwspacemalloc(ctrl, gnvtxs);
/* create different processor groups */
gkMPI_Comm_split(ctrl->gcomm, ctrl->mype % ngroups, 0, &ipcomm);
gkMPI_Comm_rank(ipcomm, &mype);
gkMPI_Comm_size(ipcomm, &npes);
/* Go into the recursive bisection */
METIS_SetDefaultOptions(moptions);
moptions[METIS_OPTION_SEED] = ctrl->sync + (ctrl->mype % ngroups) + 1;
tpwgts = ctrl->tpwgts;
tpwgts2 = rwspacemalloc(ctrl, 2*ncon);
lnparts = ctrl->nparts;
fpart = fpe = 0;
lnpes = npes;
while (lnpes > 1 && lnparts > 1) {
/* determine the weights of the two partitions as a function of the
weight of the target partition weights */
for (j=(lnparts>>1), i=0; i<ncon; i++) {
tpwgts2[i] = rsum(j, tpwgts+fpart*ncon+i, ncon);
tpwgts2[ncon+i] = rsum(lnparts-j, tpwgts+(fpart+j)*ncon+i, ncon);
wsum = 1.0/(tpwgts2[i] + tpwgts2[ncon+i]);
tpwgts2[i] *= wsum;
tpwgts2[ncon+i] *= wsum;
}
METIS_PartGraphRecursive(&agraph->nvtxs, &ncon, agraph->xadj, agraph->adjncy,
agraph->vwgt, NULL, agraph->adjwgt, &twoparts, tpwgts2, NULL, moptions,
&edgecut, part);
/* pick one of the branches */
if (mype < fpe+lnpes/2) {
KeepPart(ctrl, agraph, part, 0);
lnpes = lnpes/2;
lnparts = lnparts/2;
}
else {
KeepPart(ctrl, agraph, part, 1);
fpart = fpart + lnparts/2;
fpe = fpe + lnpes/2;
lnpes = lnpes - lnpes/2;
lnparts = lnparts - lnparts/2;
}
}
gwhere0 = iset(gnvtxs, 0, iwspacemalloc(ctrl, gnvtxs));
gwhere1 = iwspacemalloc(ctrl, gnvtxs);
if (lnparts == 1) { /* Case npes is greater than or equal to nparts */
/* Only the first process will assign labels (for the reduction to work) */
if (mype == fpe) {
for (i=0; i<agraph->nvtxs; i++)
gwhere0[agraph->label[i]] = fpart;
}
}
else { /* Case in which npes is smaller than nparts */
/* create the normalized tpwgts for the lnparts from ctrl->tpwgts */
tpwgts = rwspacemalloc(ctrl, lnparts*ncon);
for (j=0; j<ncon; j++) {
for (wsum=0.0, i=0; i<lnparts; i++) {
tpwgts[i*ncon+j] = ctrl->tpwgts[(fpart+i)*ncon+j];
wsum += tpwgts[i*ncon+j];
}
for (wsum=1.0/wsum, i=0; i<lnparts; i++)
tpwgts[i*ncon+j] *= wsum;
}
METIS_PartGraphKway(&agraph->nvtxs, &ncon, agraph->xadj, agraph->adjncy,
agraph->vwgt, NULL, agraph->adjwgt, &lnparts, tpwgts, NULL, moptions,
&edgecut, part);
for (i=0; i<agraph->nvtxs; i++)
gwhere0[agraph->label[i]] = fpart + part[i];
}
gkMPI_Allreduce((void *)gwhere0, (void *)gwhere1, gnvtxs, IDX_T, MPI_SUM, ipcomm);
if (ngroups > 1) {
tmpxadj = agraph->xadj;
tmpadjncy = agraph->adjncy;
tmpadjwgt = agraph->adjwgt;
tmpvwgt = agraph->vwgt;
tmpwhere = agraph->where;
agraph->xadj = xadj;
agraph->adjncy = adjncy;
agraph->adjwgt = adjwgt;
agraph->vwgt = vwgt;
agraph->where = gwhere1;
agraph->vwgt = vwgt;
agraph->nvtxs = gnvtxs;
edgecut = ComputeSerialEdgeCut(agraph);
ComputeSerialBalance(ctrl, agraph, gwhere1, lbvec);
lbsum = rsum(ncon, lbvec, 1);
gkMPI_Allreduce((void *)&edgecut, (void *)&max_cut, 1, IDX_T, MPI_MAX, ctrl->gcomm);
gkMPI_Allreduce((void *)&lbsum, (void *)&min_lbsum, 1, REAL_T, MPI_MIN, ctrl->gcomm);
lpesum.sum = lbsum;
if (min_lbsum < UNBALANCE_FRACTION*ncon) {
if (lbsum < UNBALANCE_FRACTION*ncon)
lpesum.sum = edgecut;
else
lpesum.sum = max_cut;
}
lpesum.rank = ctrl->mype;
gkMPI_Allreduce((void *)&lpesum, (void *)&gpesum, 1, MPI_DOUBLE_INT,
MPI_MINLOC, ctrl->gcomm);
gkMPI_Bcast((void *)gwhere1, gnvtxs, IDX_T, gpesum.rank, ctrl->gcomm);
agraph->xadj = tmpxadj;
agraph->adjncy = tmpadjncy;
agraph->adjwgt = tmpadjwgt;
agraph->vwgt = tmpvwgt;
agraph->where = tmpwhere;
}
icopy(graph->nvtxs, gwhere1+graph->vtxdist[ctrl->mype], graph->where);
FreeGraph(agraph);
gkMPI_Comm_free(&ipcomm);
IFSET(ctrl->dbglvl, DBG_TIME, gkMPI_Barrier(ctrl->comm));
IFSET(ctrl->dbglvl, DBG_TIME, stoptimer(ctrl->InitPartTmr));
WCOREPOP;
}
void FM_2WayCutRefine(ctrl_t *ctrl, graph_t *graph, real_t *ntpwgts, idx_t niter)
{
idx_t i, ii, j, k, kwgt, nvtxs, nbnd, nswaps, from, to, pass, me, limit, tmp;
idx_t *xadj, *vwgt, *adjncy, *adjwgt, *where, *id, *ed, *bndptr, *bndind, *pwgts;
idx_t *moved, *swaps, *perm;
rpq_t *queues[2];
idx_t higain, mincut, mindiff, origdiff, initcut, newcut, mincutorder, avgvwgt;
idx_t tpwgts[2];
WCOREPUSH;
nvtxs = graph->nvtxs;
xadj = graph->xadj;
vwgt = graph->vwgt;
adjncy = graph->adjncy;
adjwgt = graph->adjwgt;
where = graph->where;
id = graph->id;
ed = graph->ed;
pwgts = graph->pwgts;
bndptr = graph->bndptr;
bndind = graph->bndind;
moved = iwspacemalloc(ctrl, nvtxs);
swaps = iwspacemalloc(ctrl, nvtxs);
perm = iwspacemalloc(ctrl, nvtxs);
tpwgts[0] = graph->tvwgt[0]*ntpwgts[0];
tpwgts[1] = graph->tvwgt[0]-tpwgts[0];
limit = gk_min(gk_max(0.01*nvtxs, 15), 100);
avgvwgt = gk_min((pwgts[0]+pwgts[1])/20, 2*(pwgts[0]+pwgts[1])/nvtxs);
queues[0] = rpqCreate(nvtxs);
queues[1] = rpqCreate(nvtxs);
IFSET(ctrl->dbglvl, METIS_DBG_REFINE,
Print2WayRefineStats(ctrl, graph, ntpwgts, 0, -2));
origdiff = iabs(tpwgts[0]-pwgts[0]);
iset(nvtxs, -1, moved);
for (pass=0; pass<niter; pass++) { /* Do a number of passes */
rpqReset(queues[0]);
rpqReset(queues[1]);
mincutorder = -1;
newcut = mincut = initcut = graph->mincut;
mindiff = iabs(tpwgts[0]-pwgts[0]);
ASSERT(ComputeCut(graph, where) == graph->mincut);
ASSERT(CheckBnd(graph));
/* Insert boundary nodes in the priority queues */
nbnd = graph->nbnd;
irandArrayPermute(nbnd, perm, nbnd, 1);
for (ii=0; ii<nbnd; ii++) {
i = perm[ii];
ASSERT(ed[bndind[i]] > 0 || id[bndind[i]] == 0);
ASSERT(bndptr[bndind[i]] != -1);
rpqInsert(queues[where[bndind[i]]], bndind[i], ed[bndind[i]]-id[bndind[i]]);
}
for (nswaps=0; nswaps<nvtxs; nswaps++) {
from = (tpwgts[0]-pwgts[0] < tpwgts[1]-pwgts[1] ? 0 : 1);
to = (from+1)%2;
if ((higain = rpqGetTop(queues[from])) == -1)
break;
ASSERT(bndptr[higain] != -1);
newcut -= (ed[higain]-id[higain]);
INC_DEC(pwgts[to], pwgts[from], vwgt[higain]);
if ((newcut < mincut && iabs(tpwgts[0]-pwgts[0]) <= origdiff+avgvwgt) ||
(newcut == mincut && iabs(tpwgts[0]-pwgts[0]) < mindiff)) {
mincut = newcut;
mindiff = iabs(tpwgts[0]-pwgts[0]);
mincutorder = nswaps;
}
else if (nswaps-mincutorder > limit) { /* We hit the limit, undo last move */
newcut += (ed[higain]-id[higain]);
INC_DEC(pwgts[from], pwgts[to], vwgt[higain]);
break;
}
where[higain] = to;
moved[higain] = nswaps;
swaps[nswaps] = higain;
IFSET(ctrl->dbglvl, METIS_DBG_MOVEINFO,
printf("Moved %6"PRIDX" from %"PRIDX". [%3"PRIDX" %3"PRIDX"] %5"PRIDX" [%4"PRIDX" %4"PRIDX"]\n", higain, from, ed[higain]-id[higain], vwgt[higain], newcut, pwgts[0], pwgts[1]));
/**************************************************************
* Update the id[i]/ed[i] values of the affected nodes
***************************************************************/
SWAP(id[higain], ed[higain], tmp);
if (ed[higain] == 0 && xadj[higain] < xadj[higain+1])
BNDDelete(nbnd, bndind, bndptr, higain);
for (j=xadj[higain]; j<xadj[higain+1]; j++) {
k = adjncy[j];
kwgt = (to == where[k] ? adjwgt[j] : -adjwgt[j]);
INC_DEC(id[k], ed[k], kwgt);
/* Update its boundary information and queue position */
if (bndptr[k] != -1) { /* If k was a boundary vertex */
if (ed[k] == 0) { /* Not a boundary vertex any more */
BNDDelete(nbnd, bndind, bndptr, k);
if (moved[k] == -1) /* Remove it if in the queues */
rpqDelete(queues[where[k]], k);
}
else { /* If it has not been moved, update its position in the queue */
if (moved[k] == -1)
rpqUpdate(queues[where[k]], k, ed[k]-id[k]);
}
}
else {
if (ed[k] > 0) { /* It will now become a boundary vertex */
BNDInsert(nbnd, bndind, bndptr, k);
if (moved[k] == -1)
rpqInsert(queues[where[k]], k, ed[k]-id[k]);
}
}
}
}
/****************************************************************
* Roll back computations
*****************************************************************/
for (i=0; i<nswaps; i++)
moved[swaps[i]] = -1; /* reset moved array */
for (nswaps--; nswaps>mincutorder; nswaps--) {
higain = swaps[nswaps];
to = where[higain] = (where[higain]+1)%2;
SWAP(id[higain], ed[higain], tmp);
if (ed[higain] == 0 && bndptr[higain] != -1 && xadj[higain] < xadj[higain+1])
BNDDelete(nbnd, bndind, bndptr, higain);
else if (ed[higain] > 0 && bndptr[higain] == -1)
BNDInsert(nbnd, bndind, bndptr, higain);
INC_DEC(pwgts[to], pwgts[(to+1)%2], vwgt[higain]);
for (j=xadj[higain]; j<xadj[higain+1]; j++) {
k = adjncy[j];
kwgt = (to == where[k] ? adjwgt[j] : -adjwgt[j]);
INC_DEC(id[k], ed[k], kwgt);
if (bndptr[k] != -1 && ed[k] == 0)
BNDDelete(nbnd, bndind, bndptr, k);
if (bndptr[k] == -1 && ed[k] > 0)
BNDInsert(nbnd, bndind, bndptr, k);
}
}
graph->mincut = mincut;
graph->nbnd = nbnd;
IFSET(ctrl->dbglvl, METIS_DBG_REFINE,
Print2WayRefineStats(ctrl, graph, ntpwgts, 0, mincutorder));
if (mincutorder <= 0 || mincut == initcut)
break;
}
rpqDestroy(queues[0]);
rpqDestroy(queues[1]);
WCOREPOP;
}
void FM_Mc2WayCutRefine(ctrl_t *ctrl, graph_t *graph, real_t *ntpwgts, idx_t niter)
{
idx_t i, ii, j, k, l, kwgt, nvtxs, ncon, nbnd, nswaps, from, to, pass,
me, limit, tmp, cnum;
idx_t *xadj, *adjncy, *vwgt, *adjwgt, *pwgts, *where, *id, *ed,
*bndptr, *bndind;
idx_t *moved, *swaps, *perm, *qnum;
idx_t higain, mincut, initcut, newcut, mincutorder;
real_t *invtvwgt, *ubfactors, *minbalv, *newbalv;
real_t origbal, minbal, newbal, rgain, ffactor;
rpq_t **queues;
WCOREPUSH;
nvtxs = graph->nvtxs;
ncon = graph->ncon;
xadj = graph->xadj;
vwgt = graph->vwgt;
adjncy = graph->adjncy;
adjwgt = graph->adjwgt;
invtvwgt = graph->invtvwgt;
where = graph->where;
id = graph->id;
ed = graph->ed;
pwgts = graph->pwgts;
bndptr = graph->bndptr;
bndind = graph->bndind;
moved = iwspacemalloc(ctrl, nvtxs);
swaps = iwspacemalloc(ctrl, nvtxs);
perm = iwspacemalloc(ctrl, nvtxs);
qnum = iwspacemalloc(ctrl, nvtxs);
ubfactors = rwspacemalloc(ctrl, ncon);
newbalv = rwspacemalloc(ctrl, ncon);
minbalv = rwspacemalloc(ctrl, ncon);
limit = gk_min(gk_max(0.01*nvtxs, 25), 150);
/* Determine a fudge factor to allow the refinement routines to get out
of tight balancing constraints. */
ffactor = .5/gk_max(20, nvtxs);
/* Initialize the queues */
queues = (rpq_t **)wspacemalloc(ctrl, 2*ncon*sizeof(rpq_t *));
for (i=0; i<2*ncon; i++)
queues[i] = rpqCreate(nvtxs);
for (i=0; i<nvtxs; i++)
qnum[i] = iargmax_nrm(ncon, vwgt+i*ncon, invtvwgt);
/* Determine the unbalance tolerance for each constraint. The tolerance is
equal to the maximum of the original load imbalance and the user-supplied
allowed tolerance. The rationale behind this approach is to allow the
refinement routine to improve the cut, without having to worry about fixing
load imbalance problems. The load imbalance is addressed by the balancing
routines. */
origbal = ComputeLoadImbalanceDiffVec(graph, 2, ctrl->pijbm, ctrl->ubfactors, ubfactors);
for (i=0; i<ncon; i++)
ubfactors[i] = (ubfactors[i] > 0 ? ctrl->ubfactors[i]+ubfactors[i] : ctrl->ubfactors[i]);
IFSET(ctrl->dbglvl, METIS_DBG_REFINE,
Print2WayRefineStats(ctrl, graph, ntpwgts, origbal, -2));
iset(nvtxs, -1, moved);
for (pass=0; pass<niter; pass++) { /* Do a number of passes */
for (i=0; i<2*ncon; i++)
rpqReset(queues[i]);
mincutorder = -1;
newcut = mincut = initcut = graph->mincut;
minbal = ComputeLoadImbalanceDiffVec(graph, 2, ctrl->pijbm, ubfactors, minbalv);
ASSERT(ComputeCut(graph, where) == graph->mincut);
ASSERT(CheckBnd(graph));
/* Insert boundary nodes in the priority queues */
nbnd = graph->nbnd;
irandArrayPermute(nbnd, perm, nbnd/5, 1);
for (ii=0; ii<nbnd; ii++) {
i = bndind[perm[ii]];
ASSERT(ed[i] > 0 || id[i] == 0);
ASSERT(bndptr[i] != -1);
//rgain = 1.0*(ed[i]-id[i])/sqrt(vwgt[i*ncon+qnum[i]]+1);
//rgain = (ed[i]-id[i] > 0 ? 1.0*(ed[i]-id[i])/sqrt(vwgt[i*ncon+qnum[i]]+1) : ed[i]-id[i]);
rgain = ed[i]-id[i];
rpqInsert(queues[2*qnum[i]+where[i]], i, rgain);
}
for (nswaps=0; nswaps<nvtxs; nswaps++) {
SelectQueue(graph, ctrl->pijbm, ubfactors, queues, &from, &cnum);
to = (from+1)%2;
if (from == -1 || (higain = rpqGetTop(queues[2*cnum+from])) == -1)
break;
ASSERT(bndptr[higain] != -1);
newcut -= (ed[higain]-id[higain]);
iaxpy(ncon, 1, vwgt+higain*ncon, 1, pwgts+to*ncon, 1);
iaxpy(ncon, -1, vwgt+higain*ncon, 1, pwgts+from*ncon, 1);
newbal = ComputeLoadImbalanceDiffVec(graph, 2, ctrl->pijbm, ubfactors, newbalv);
if ((newcut < mincut && newbal <= ffactor) ||
(newcut == mincut && (newbal < minbal ||
(newbal == minbal && BetterBalance2Way(ncon, minbalv, newbalv))))) {
mincut = newcut;
minbal = newbal;
mincutorder = nswaps;
rcopy(ncon, newbalv, minbalv);
}
else if (nswaps-mincutorder > limit) { /* We hit the limit, undo last move */
newcut += (ed[higain]-id[higain]);
iaxpy(ncon, 1, vwgt+higain*ncon, 1, pwgts+from*ncon, 1);
iaxpy(ncon, -1, vwgt+higain*ncon, 1, pwgts+to*ncon, 1);
break;
}
where[higain] = to;
moved[higain] = nswaps;
swaps[nswaps] = higain;
if (ctrl->dbglvl&METIS_DBG_MOVEINFO) {
printf("Moved%6"PRIDX" from %"PRIDX"(%"PRIDX") Gain:%5"PRIDX", "
"Cut:%5"PRIDX", NPwgts:", higain, from, cnum, ed[higain]-id[higain], newcut);
for (l=0; l<ncon; l++)
printf("(%.3"PRREAL" %.3"PRREAL")", pwgts[l]*invtvwgt[l], pwgts[ncon+l]*invtvwgt[l]);
printf(" %+.3"PRREAL" LB: %.3"PRREAL"(%+.3"PRREAL")\n",
minbal, ComputeLoadImbalance(graph, 2, ctrl->pijbm), newbal);
}
/**************************************************************
* Update the id[i]/ed[i] values of the affected nodes
***************************************************************/
SWAP(id[higain], ed[higain], tmp);
if (ed[higain] == 0 && xadj[higain] < xadj[higain+1])
BNDDelete(nbnd, bndind, bndptr, higain);
for (j=xadj[higain]; j<xadj[higain+1]; j++) {
k = adjncy[j];
kwgt = (to == where[k] ? adjwgt[j] : -adjwgt[j]);
INC_DEC(id[k], ed[k], kwgt);
/* Update its boundary information and queue position */
if (bndptr[k] != -1) { /* If k was a boundary vertex */
if (ed[k] == 0) { /* Not a boundary vertex any more */
BNDDelete(nbnd, bndind, bndptr, k);
if (moved[k] == -1) /* Remove it if in the queues */
rpqDelete(queues[2*qnum[k]+where[k]], k);
}
else { /* If it has not been moved, update its position in the queue */
if (moved[k] == -1) {
//rgain = 1.0*(ed[k]-id[k])/sqrt(vwgt[k*ncon+qnum[k]]+1);
//rgain = (ed[k]-id[k] > 0 ?
// 1.0*(ed[k]-id[k])/sqrt(vwgt[k*ncon+qnum[k]]+1) : ed[k]-id[k]);
rgain = ed[k]-id[k];
rpqUpdate(queues[2*qnum[k]+where[k]], k, rgain);
}
}
}
else {
if (ed[k] > 0) { /* It will now become a boundary vertex */
BNDInsert(nbnd, bndind, bndptr, k);
if (moved[k] == -1) {
//rgain = 1.0*(ed[k]-id[k])/sqrt(vwgt[k*ncon+qnum[k]]+1);
//rgain = (ed[k]-id[k] > 0 ?
// 1.0*(ed[k]-id[k])/sqrt(vwgt[k*ncon+qnum[k]]+1) : ed[k]-id[k]);
rgain = ed[k]-id[k];
rpqInsert(queues[2*qnum[k]+where[k]], k, rgain);
}
}
}
}
}
/****************************************************************
* Roll back computations
*****************************************************************/
for (i=0; i<nswaps; i++)
moved[swaps[i]] = -1; /* reset moved array */
for (nswaps--; nswaps>mincutorder; nswaps--) {
higain = swaps[nswaps];
to = where[higain] = (where[higain]+1)%2;
SWAP(id[higain], ed[higain], tmp);
if (ed[higain] == 0 && bndptr[higain] != -1 && xadj[higain] < xadj[higain+1])
BNDDelete(nbnd, bndind, bndptr, higain);
else if (ed[higain] > 0 && bndptr[higain] == -1)
BNDInsert(nbnd, bndind, bndptr, higain);
iaxpy(ncon, 1, vwgt+higain*ncon, 1, pwgts+to*ncon, 1);
iaxpy(ncon, -1, vwgt+higain*ncon, 1, pwgts+((to+1)%2)*ncon, 1);
for (j=xadj[higain]; j<xadj[higain+1]; j++) {
k = adjncy[j];
kwgt = (to == where[k] ? adjwgt[j] : -adjwgt[j]);
INC_DEC(id[k], ed[k], kwgt);
if (bndptr[k] != -1 && ed[k] == 0)
BNDDelete(nbnd, bndind, bndptr, k);
if (bndptr[k] == -1 && ed[k] > 0)
BNDInsert(nbnd, bndind, bndptr, k);
}
}
graph->mincut = mincut;
graph->nbnd = nbnd;
IFSET(ctrl->dbglvl, METIS_DBG_REFINE,
Print2WayRefineStats(ctrl, graph, ntpwgts, minbal, mincutorder));
if (mincutorder <= 0 || mincut == initcut)
break;
}
for (i=0; i<2*ncon; i++)
rpqDestroy(queues[i]);
WCOREPOP;
}
/*************************************************************************
* This function performs a k-way directed diffusion
**************************************************************************/
real_t WavefrontDiffusion(ctrl_t *ctrl, graph_t *graph, idx_t *home)
{
idx_t ii, i, j, k, l, nvtxs, nedges, nparts;
idx_t from, to, edge, done, nswaps, noswaps, totalv, wsize;
idx_t npasses, first, second, third, mind, maxd;
idx_t *xadj, *adjncy, *adjwgt, *where, *perm;
idx_t *rowptr, *colind, *ed, *psize;
real_t *transfer, *tmpvec;
real_t balance = -1.0, *load, *solution, *workspace;
real_t *nvwgt, *npwgts, flowFactor, cost, ubfactor;
matrix_t matrix;
ikv_t *cand;
idx_t ndirty, nclean, dptr, clean;
nvtxs = graph->nvtxs;
nedges = graph->nedges;
xadj = graph->xadj;
nvwgt = graph->nvwgt;
adjncy = graph->adjncy;
adjwgt = graph->adjwgt;
where = graph->where;
nparts = ctrl->nparts;
ubfactor = ctrl->ubvec[0];
matrix.nrows = nparts;
flowFactor = 0.35;
flowFactor = (ctrl->mype == 2) ? 0.50 : flowFactor;
flowFactor = (ctrl->mype == 3) ? 0.75 : flowFactor;
flowFactor = (ctrl->mype == 4) ? 1.00 : flowFactor;
/* allocate memory */
solution = rmalloc(4*nparts+2*nedges, "WavefrontDiffusion: solution");
tmpvec = solution + nparts;
npwgts = solution + 2*nparts;
load = solution + 3*nparts;
matrix.values = solution + 4*nparts;
transfer = matrix.transfer = solution + 4*nparts + nedges;
perm = imalloc(2*nvtxs+2*nparts+nedges+1, "WavefrontDiffusion: perm");
ed = perm + nvtxs;
psize = perm + 2*nvtxs;
rowptr = matrix.rowptr = perm + 2*nvtxs + nparts;
colind = matrix.colind = perm + 2*nvtxs + 2*nparts + 1;
/*GKTODO - Potential problem with this malloc */
wsize = gk_max(sizeof(real_t)*nparts*6, sizeof(idx_t)*(nvtxs+nparts*2+1));
workspace = (real_t *)gk_malloc(wsize, "WavefrontDiffusion: workspace");
cand = ikvmalloc(nvtxs, "WavefrontDiffusion: cand");
/*****************************/
/* Populate empty subdomains */
/*****************************/
iset(nparts, 0, psize);
for (i=0; i<nvtxs; i++)
psize[where[i]]++;
mind = iargmin(nparts, psize);
maxd = iargmax(nparts, psize);
if (psize[mind] == 0) {
for (i=0; i<nvtxs; i++) {
k = (RandomInRange(nvtxs)+i)%nvtxs;
if (where[k] == maxd) {
where[k] = mind;
psize[mind]++;
psize[maxd]--;
break;
}
}
}
iset(nvtxs, 0, ed);
rset(nparts, 0.0, npwgts);
for (i=0; i<nvtxs; i++) {
npwgts[where[i]] += nvwgt[i];
for (j=xadj[i]; j<xadj[i+1]; j++)
ed[i] += (where[i] != where[adjncy[j]] ? adjwgt[j] : 0);
}
ComputeLoad(graph, nparts, load, ctrl->tpwgts, 0);
done = 0;
/* zero out the tmpvec array */
rset(nparts, 0.0, tmpvec);
npasses = gk_min(nparts/2, NGD_PASSES);
for (l=0; l<npasses; l++) {
/* Set-up and solve the diffusion equation */
nswaps = 0;
/************************/
/* Solve flow equations */
/************************/
SetUpConnectGraph(graph, &matrix, (idx_t *)workspace);
/* check for disconnected subdomains */
for(i=0; i<matrix.nrows; i++) {
if (matrix.rowptr[i]+1 == matrix.rowptr[i+1]) {
cost = (real_t)(ctrl->mype);
goto CleanUpAndExit;
}
}
ConjGrad2(&matrix, load, solution, 0.001, workspace);
ComputeTransferVector(1, &matrix, solution, transfer, 0);
GetThreeMax(nparts, load, &first, &second, &third);
if (l%3 == 0) {
FastRandomPermute(nvtxs, perm, 1);
}
else {
/*****************************/
/* move dirty vertices first */
/*****************************/
ndirty = 0;
for (i=0; i<nvtxs; i++) {
if (where[i] != home[i])
ndirty++;
}
dptr = 0;
for (i=0; i<nvtxs; i++) {
if (where[i] != home[i])
perm[dptr++] = i;
else
perm[ndirty++] = i;
}
PASSERT(ctrl, ndirty == nvtxs);
ndirty = dptr;
nclean = nvtxs-dptr;
FastRandomPermute(ndirty, perm, 0);
FastRandomPermute(nclean, perm+ndirty, 0);
}
if (ctrl->mype == 0) {
for (j=nvtxs, k=0, ii=0; ii<nvtxs; ii++) {
i = perm[ii];
if (ed[i] != 0) {
cand[k].key = -ed[i];
cand[k++].val = i;
}
else {
cand[--j].key = 0;
cand[j].val = i;
}
}
ikvsorti(k, cand);
}
for (ii=0; ii<nvtxs/3; ii++) {
i = (ctrl->mype == 0) ? cand[ii].val : perm[ii];
from = where[i];
/* don't move out the last vertex in a subdomain */
if (psize[from] == 1)
continue;
clean = (from == home[i]) ? 1 : 0;
/* only move from top three or dirty vertices */
if (from != first && from != second && from != third && clean)
continue;
/* Scatter the sparse transfer row into the dense tmpvec row */
for (j=rowptr[from]+1; j<rowptr[from+1]; j++)
tmpvec[colind[j]] = transfer[j];
for (j=xadj[i]; j<xadj[i+1]; j++) {
to = where[adjncy[j]];
if (from != to) {
if (tmpvec[to] > (flowFactor * nvwgt[i])) {
tmpvec[to] -= nvwgt[i];
INC_DEC(psize[to], psize[from], 1);
INC_DEC(npwgts[to], npwgts[from], nvwgt[i]);
INC_DEC(load[to], load[from], nvwgt[i]);
where[i] = to;
nswaps++;
/* Update external degrees */
ed[i] = 0;
for (k=xadj[i]; k<xadj[i+1]; k++) {
edge = adjncy[k];
ed[i] += (to != where[edge] ? adjwgt[k] : 0);
if (where[edge] == from)
ed[edge] += adjwgt[k];
if (where[edge] == to)
ed[edge] -= adjwgt[k];
}
break;
}
}
}
/* Gather the dense tmpvec row into the sparse transfer row */
for (j=rowptr[from]+1; j<rowptr[from+1]; j++) {
transfer[j] = tmpvec[colind[j]];
tmpvec[colind[j]] = 0.0;
}
ASSERT(fabs(rsum(nparts, tmpvec, 1)) < .0001)
}
if (l % 2 == 1) {
balance = rmax(nparts, npwgts)*nparts;
if (balance < ubfactor + 0.035)
done = 1;
if (GlobalSESum(ctrl, done) > 0)
break;
noswaps = (nswaps > 0) ? 0 : 1;
if (GlobalSESum(ctrl, noswaps) > ctrl->npes/2)
break;
}
}
graph->mincut = ComputeSerialEdgeCut(graph);
totalv = Mc_ComputeSerialTotalV(graph, home);
cost = ctrl->ipc_factor * (real_t)graph->mincut + ctrl->redist_factor * (real_t)totalv;
CleanUpAndExit:
gk_free((void **)&solution, (void **)&perm, (void **)&workspace, (void **)&cand, LTERM);
return cost;
}
/***********************************************************************************
* This function is the entry point of the parallel multilevel local diffusion
* algorithm. It uses parallel undirected diffusion followed by adaptive k-way
* refinement. This function utilizes local coarsening.
************************************************************************************/
int ParMETIS_V3_RefineKway(idx_t *vtxdist, idx_t *xadj, idx_t *adjncy, idx_t *vwgt,
idx_t *adjwgt, idx_t *wgtflag, idx_t *numflag, idx_t *ncon, idx_t *nparts,
real_t *tpwgts, real_t *ubvec, idx_t *options, idx_t *edgecut, idx_t *part,
MPI_Comm *comm)
{
idx_t npes, mype, status;
ctrl_t *ctrl=NULL;
graph_t *graph=NULL;
size_t curmem;
/* Check the input parameters and return if an error */
status = CheckInputsPartKway(vtxdist, xadj, adjncy, vwgt, adjwgt, wgtflag,
numflag, ncon, nparts, tpwgts, ubvec, options, edgecut, part, comm);
if (GlobalSEMinComm(*comm, status) == 0)
return METIS_ERROR;
status = METIS_OK;
gk_malloc_init();
curmem = gk_GetCurMemoryUsed();
/* Setup ctrl */
ctrl = SetupCtrl(PARMETIS_OP_RMETIS, options, *ncon, *nparts, tpwgts, ubvec, *comm);
npes = ctrl->npes;
mype = ctrl->mype;
/* Take care the nparts == 1 case */
if (*nparts == 1) {
iset(vtxdist[mype+1]-vtxdist[mype], (*numflag == 0 ? 0 : 1), part);
*edgecut = 0;
goto DONE;
}
/* setup the graph */
if (*numflag > 0)
ChangeNumbering(vtxdist, xadj, adjncy, part, npes, mype, 1);
graph = SetupGraph(ctrl, *ncon, vtxdist, xadj, vwgt, NULL, adjncy, adjwgt, *wgtflag);
if (ctrl->ps_relation == PARMETIS_PSR_COUPLED)
iset(graph->nvtxs, mype, graph->home);
else
icopy(graph->nvtxs, part, graph->home);
/* Allocate workspace */
AllocateWSpace(ctrl, 10*graph->nvtxs);
/* Partition and Remap */
STARTTIMER(ctrl, ctrl->TotalTmr);
ctrl->CoarsenTo = gk_min(vtxdist[npes]+1, 50*(*ncon)*gk_max(npes, *nparts));
Adaptive_Partition(ctrl, graph);
ParallelReMapGraph(ctrl, graph);
icopy(graph->nvtxs, graph->where, part);
*edgecut = graph->mincut;
STOPTIMER(ctrl, ctrl->TotalTmr);
/* Take care of output */
IFSET(ctrl->dbglvl, DBG_TIME, PrintTimingInfo(ctrl));
IFSET(ctrl->dbglvl, DBG_TIME, gkMPI_Barrier(ctrl->gcomm));
IFSET(ctrl->dbglvl, DBG_INFO, PrintPostPartInfo(ctrl, graph, 1));
FreeInitialGraphAndRemap(graph);
if (*numflag > 0)
ChangeNumbering(vtxdist, xadj, adjncy, part, npes, mype, 0);
DONE:
FreeCtrl(&ctrl);
if (gk_GetCurMemoryUsed() - curmem > 0) {
printf("ParMETIS appears to have a memory leak of %zdbytes. Report this.\n",
(ssize_t)(gk_GetCurMemoryUsed() - curmem));
}
gk_malloc_cleanup(0);
return (int)status;
}
void FM_2WayNodeRefine1Sided(ctrl_t *ctrl, graph_t *graph, idx_t niter)
{
idx_t i, ii, j, k, jj, kk, nvtxs, nbnd, nswaps, nmind, iend;
idx_t *xadj, *vwgt, *adjncy, *where, *pwgts, *edegrees, *bndind, *bndptr;
idx_t *mptr, *mind, *swaps;
rpq_t *queue;
nrinfo_t *rinfo;
idx_t higain, mincut, initcut, mincutorder;
idx_t pass, to, other, limit;
idx_t badmaxpwgt, mindiff, newdiff;
real_t mult;
WCOREPUSH;
nvtxs = graph->nvtxs;
xadj = graph->xadj;
adjncy = graph->adjncy;
vwgt = graph->vwgt;
bndind = graph->bndind;
bndptr = graph->bndptr;
where = graph->where;
pwgts = graph->pwgts;
rinfo = graph->nrinfo;
queue = rpqCreate(nvtxs);
swaps = iwspacemalloc(ctrl, nvtxs);
mptr = iwspacemalloc(ctrl, nvtxs+1);
mind = iwspacemalloc(ctrl, 2*nvtxs);
mult = 0.5*ctrl->ubfactors[0];
badmaxpwgt = (idx_t)(mult*(pwgts[0]+pwgts[1]+pwgts[2]));
IFSET(ctrl->dbglvl, METIS_DBG_REFINE,
printf("Partitions-N1: [%6"PRIDX" %6"PRIDX"] Nv-Nb[%6"PRIDX" %6"PRIDX"]. ISep: %6"PRIDX"\n", pwgts[0], pwgts[1], graph->nvtxs, graph->nbnd, graph->mincut));
to = (pwgts[0] < pwgts[1] ? 1 : 0);
for (pass=0; pass<2*niter; pass++) { /* the 2*niter is for the two sides */
other = to;
to = (to+1)%2;
rpqReset(queue);
mincutorder = -1;
initcut = mincut = graph->mincut;
nbnd = graph->nbnd;
/* use the swaps array in place of the traditional perm array to save memory */
my_irandArrayPermute_r(nbnd, swaps, nbnd, 1, &ctrl->curseed);
for (ii=0; ii<nbnd; ii++) {
i = bndind[swaps[ii]];
ASSERT(where[i] == 2);
rpqInsert(queue, i, vwgt[i]-rinfo[i].edegrees[other]);
}
ASSERT(CheckNodeBnd(graph, nbnd));
ASSERT(CheckNodePartitionParams(graph));
limit = (ctrl->compress ? gk_min(5*nbnd, 500) : gk_min(3*nbnd, 300));
/******************************************************
* Get into the FM loop
*******************************************************/
IFSET(ctrl->dbglvl, METIS_DBG_TIME, gk_startwctimer(ctrl->Aux3Tmr));
mptr[0] = nmind = 0;
mindiff = iabs(pwgts[0]-pwgts[1]);
for (nswaps=0; nswaps<nvtxs; nswaps++) {
if ((higain = rpqGetTop(queue)) == -1)
break;
ASSERT(bndptr[higain] != -1);
/* The following check is to ensure we break out if there is a posibility
of over-running the mind array. */
if (nmind + xadj[higain+1]-xadj[higain] >= 2*nvtxs-1)
break;
if (pwgts[to]+vwgt[higain] > badmaxpwgt)
break; /* No point going any further. Balance will be bad */
pwgts[2] -= (vwgt[higain]-rinfo[higain].edegrees[other]);
newdiff = iabs(pwgts[to]+vwgt[higain] - (pwgts[other]-rinfo[higain].edegrees[other]));
if (pwgts[2] < mincut || (pwgts[2] == mincut && newdiff < mindiff)) {
mincut = pwgts[2];
mincutorder = nswaps;
mindiff = newdiff;
}
else {
if (nswaps - mincutorder > 3*limit ||
(nswaps - mincutorder > limit && pwgts[2] > 1.10*mincut)) {
pwgts[2] += (vwgt[higain]-rinfo[higain].edegrees[other]);
break; /* No further improvement, break out */
}
}
BNDDelete(nbnd, bndind, bndptr, higain);
pwgts[to] += vwgt[higain];
where[higain] = to;
swaps[nswaps] = higain;
/**********************************************************
* Update the degrees of the affected nodes
***********************************************************/
IFSET(ctrl->dbglvl, METIS_DBG_TIME, gk_startwctimer(ctrl->Aux1Tmr));
for (j=xadj[higain]; j<xadj[higain+1]; j++) {
k = adjncy[j];
if (where[k] == 2) { /* For the in-separator vertices modify their edegree[to] */
rinfo[k].edegrees[to] += vwgt[higain];
}
else if (where[k] == other) { /* This vertex is pulled into the separator */
ASSERTP(bndptr[k] == -1, ("%"PRIDX" %"PRIDX" %"PRIDX"\n", k, bndptr[k], where[k]));
BNDInsert(nbnd, bndind, bndptr, k);
mind[nmind++] = k; /* Keep track for rollback */
where[k] = 2;
pwgts[other] -= vwgt[k];
edegrees = rinfo[k].edegrees;
edegrees[0] = edegrees[1] = 0;
for (jj=xadj[k], iend=xadj[k+1]; jj<iend; jj++) {
kk = adjncy[jj];
if (where[kk] != 2)
edegrees[where[kk]] += vwgt[kk];
else {
rinfo[kk].edegrees[other] -= vwgt[k];
/* Since the moves are one-sided this vertex has not been moved yet */
rpqUpdate(queue, kk, vwgt[kk]-rinfo[kk].edegrees[other]);
}
}
/* Insert the new vertex into the priority queue. Safe due to one-sided moves */
rpqInsert(queue, k, vwgt[k]-edegrees[other]);
}
}
mptr[nswaps+1] = nmind;
IFSET(ctrl->dbglvl, METIS_DBG_TIME, gk_stopwctimer(ctrl->Aux1Tmr));
IFSET(ctrl->dbglvl, METIS_DBG_MOVEINFO,
printf("Moved %6"PRIDX" to %3"PRIDX", Gain: %5"PRIDX" [%5"PRIDX"] \t[%5"PRIDX" %5"PRIDX" %5"PRIDX"] [%3"PRIDX" %2"PRIDX"]\n",
higain, to, (vwgt[higain]-rinfo[higain].edegrees[other]), vwgt[higain],
pwgts[0], pwgts[1], pwgts[2], nswaps, limit));
}
IFSET(ctrl->dbglvl, METIS_DBG_TIME, gk_stopwctimer(ctrl->Aux3Tmr));
/****************************************************************
* Roll back computation
*****************************************************************/
IFSET(ctrl->dbglvl, METIS_DBG_TIME, gk_startwctimer(ctrl->Aux2Tmr));
for (nswaps--; nswaps>mincutorder; nswaps--) {
higain = swaps[nswaps];
ASSERT(CheckNodePartitionParams(graph));
ASSERT(where[higain] == to);
INC_DEC(pwgts[2], pwgts[to], vwgt[higain]);
where[higain] = 2;
BNDInsert(nbnd, bndind, bndptr, higain);
edegrees = rinfo[higain].edegrees;
edegrees[0] = edegrees[1] = 0;
for (j=xadj[higain]; j<xadj[higain+1]; j++) {
k = adjncy[j];
if (where[k] == 2)
rinfo[k].edegrees[to] -= vwgt[higain];
else
edegrees[where[k]] += vwgt[k];
}
/* Push nodes out of the separator */
for (j=mptr[nswaps]; j<mptr[nswaps+1]; j++) {
k = mind[j];
ASSERT(where[k] == 2);
where[k] = other;
INC_DEC(pwgts[other], pwgts[2], vwgt[k]);
BNDDelete(nbnd, bndind, bndptr, k);
for (jj=xadj[k], iend=xadj[k+1]; jj<iend; jj++) {
kk = adjncy[jj];
if (where[kk] == 2)
rinfo[kk].edegrees[other] += vwgt[k];
}
}
}
IFSET(ctrl->dbglvl, METIS_DBG_TIME, gk_stopwctimer(ctrl->Aux2Tmr));
ASSERT(mincut == pwgts[2]);
IFSET(ctrl->dbglvl, METIS_DBG_REFINE,
printf("\tMinimum sep: %6"PRIDX" at %5"PRIDX", PWGTS: [%6"PRIDX" %6"PRIDX"], NBND: %6"PRIDX"\n", mincut, mincutorder, pwgts[0], pwgts[1], nbnd));
graph->mincut = mincut;
graph->nbnd = nbnd;
if (pass%2 == 1 && (mincutorder == -1 || mincut >= initcut))
break;
}
rpqDestroy(queue);
WCOREPOP;
}
void FM_2WayNodeRefine2Sided(ctrl_t *ctrl, graph_t *graph, idx_t niter)
{
idx_t i, ii, j, k, jj, kk, nvtxs, nbnd, nswaps, nmind;
idx_t *xadj, *vwgt, *adjncy, *where, *pwgts, *edegrees, *bndind, *bndptr;
idx_t *mptr, *mind, *moved, *swaps;
rpq_t *queues[2];
nrinfo_t *rinfo;
idx_t higain, oldgain, mincut, initcut, mincutorder;
idx_t pass, to, other, limit;
idx_t badmaxpwgt, mindiff, newdiff;
idx_t u[2], g[2];
real_t mult;
WCOREPUSH;
nvtxs = graph->nvtxs;
xadj = graph->xadj;
adjncy = graph->adjncy;
vwgt = graph->vwgt;
bndind = graph->bndind;
bndptr = graph->bndptr;
where = graph->where;
pwgts = graph->pwgts;
rinfo = graph->nrinfo;
queues[0] = rpqCreate(nvtxs);
queues[1] = rpqCreate(nvtxs);
moved = iwspacemalloc(ctrl, nvtxs);
swaps = iwspacemalloc(ctrl, nvtxs);
mptr = iwspacemalloc(ctrl, nvtxs+1);
mind = iwspacemalloc(ctrl, 2*nvtxs);
mult = 0.5*ctrl->ubfactors[0];
badmaxpwgt = (idx_t)(mult*(pwgts[0]+pwgts[1]+pwgts[2]));
IFSET(ctrl->dbglvl, METIS_DBG_REFINE,
printf("Partitions-N2: [%6"PRIDX" %6"PRIDX"] Nv-Nb[%6"PRIDX" %6"PRIDX"]. ISep: %6"PRIDX"\n", pwgts[0], pwgts[1], graph->nvtxs, graph->nbnd, graph->mincut));
for (pass=0; pass<niter; pass++) {
iset(nvtxs, -1, moved);
rpqReset(queues[0]);
rpqReset(queues[1]);
mincutorder = -1;
initcut = mincut = graph->mincut;
nbnd = graph->nbnd;
/* use the swaps array in place of the traditional perm array to save memory */
my_irandArrayPermute_r(nbnd, swaps, nbnd, 1, &ctrl->curseed);
for (ii=0; ii<nbnd; ii++) {
i = bndind[swaps[ii]];
ASSERT(where[i] == 2);
rpqInsert(queues[0], i, vwgt[i]-rinfo[i].edegrees[1]);
rpqInsert(queues[1], i, vwgt[i]-rinfo[i].edegrees[0]);
}
ASSERT(CheckNodeBnd(graph, nbnd));
ASSERT(CheckNodePartitionParams(graph));
limit = (ctrl->compress ? gk_min(5*nbnd, 400) : gk_min(2*nbnd, 300));
/******************************************************
* Get into the FM loop
*******************************************************/
mptr[0] = nmind = 0;
mindiff = iabs(pwgts[0]-pwgts[1]);
to = (pwgts[0] < pwgts[1] ? 0 : 1);
for (nswaps=0; nswaps<nvtxs; nswaps++) {
u[0] = rpqSeeTopVal(queues[0]);
u[1] = rpqSeeTopVal(queues[1]);
if (u[0] != -1 && u[1] != -1) {
g[0] = vwgt[u[0]]-rinfo[u[0]].edegrees[1];
g[1] = vwgt[u[1]]-rinfo[u[1]].edegrees[0];
to = (g[0] > g[1] ? 0 : (g[0] < g[1] ? 1 : pass%2));
if (pwgts[to]+vwgt[u[to]] > badmaxpwgt)
to = (to+1)%2;
}
else if (u[0] == -1 && u[1] == -1) {
break;
}
else if (u[0] != -1 && pwgts[0]+vwgt[u[0]] <= badmaxpwgt) {
to = 0;
}
else if (u[1] != -1 && pwgts[1]+vwgt[u[1]] <= badmaxpwgt) {
to = 1;
}
else
break;
other = (to+1)%2;
higain = rpqGetTop(queues[to]);
if (moved[higain] == -1) /* Delete if it was in the separator originally */
rpqDelete(queues[other], higain);
ASSERT(bndptr[higain] != -1);
/* The following check is to ensure we break out if there is a posibility
of over-running the mind array. */
if (nmind + xadj[higain+1]-xadj[higain] >= 2*nvtxs-1)
break;
pwgts[2] -= (vwgt[higain]-rinfo[higain].edegrees[other]);
newdiff = iabs(pwgts[to]+vwgt[higain] - (pwgts[other]-rinfo[higain].edegrees[other]));
if (pwgts[2] < mincut || (pwgts[2] == mincut && newdiff < mindiff)) {
mincut = pwgts[2];
mincutorder = nswaps;
mindiff = newdiff;
}
else {
if (nswaps - mincutorder > 2*limit ||
(nswaps - mincutorder > limit && pwgts[2] > 1.10*mincut)) {
pwgts[2] += (vwgt[higain]-rinfo[higain].edegrees[other]);
break; /* No further improvement, break out */
}
}
BNDDelete(nbnd, bndind, bndptr, higain);
pwgts[to] += vwgt[higain];
where[higain] = to;
moved[higain] = nswaps;
swaps[nswaps] = higain;
/**********************************************************
* Update the degrees of the affected nodes
***********************************************************/
for (j=xadj[higain]; j<xadj[higain+1]; j++) {
k = adjncy[j];
if (where[k] == 2) { /* For the in-separator vertices modify their edegree[to] */
oldgain = vwgt[k]-rinfo[k].edegrees[to];
rinfo[k].edegrees[to] += vwgt[higain];
if (moved[k] == -1 || moved[k] == -(2+other))
rpqUpdate(queues[other], k, oldgain-vwgt[higain]);
}
else if (where[k] == other) { /* This vertex is pulled into the separator */
ASSERTP(bndptr[k] == -1, ("%"PRIDX" %"PRIDX" %"PRIDX"\n", k, bndptr[k], where[k]));
BNDInsert(nbnd, bndind, bndptr, k);
mind[nmind++] = k; /* Keep track for rollback */
where[k] = 2;
pwgts[other] -= vwgt[k];
edegrees = rinfo[k].edegrees;
edegrees[0] = edegrees[1] = 0;
for (jj=xadj[k]; jj<xadj[k+1]; jj++) {
kk = adjncy[jj];
if (where[kk] != 2)
edegrees[where[kk]] += vwgt[kk];
else {
oldgain = vwgt[kk]-rinfo[kk].edegrees[other];
rinfo[kk].edegrees[other] -= vwgt[k];
if (moved[kk] == -1 || moved[kk] == -(2+to))
rpqUpdate(queues[to], kk, oldgain+vwgt[k]);
}
}
/* Insert the new vertex into the priority queue. Only one side! */
if (moved[k] == -1) {
rpqInsert(queues[to], k, vwgt[k]-edegrees[other]);
moved[k] = -(2+to);
}
}
}
mptr[nswaps+1] = nmind;
IFSET(ctrl->dbglvl, METIS_DBG_MOVEINFO,
printf("Moved %6"PRIDX" to %3"PRIDX", Gain: %5"PRIDX" [%5"PRIDX"] [%4"PRIDX" %4"PRIDX"] \t[%5"PRIDX" %5"PRIDX" %5"PRIDX"]\n", higain, to, g[to], g[other], vwgt[u[to]], vwgt[u[other]], pwgts[0], pwgts[1], pwgts[2]));
}
/****************************************************************
* Roll back computation
*****************************************************************/
for (nswaps--; nswaps>mincutorder; nswaps--) {
higain = swaps[nswaps];
ASSERT(CheckNodePartitionParams(graph));
to = where[higain];
other = (to+1)%2;
INC_DEC(pwgts[2], pwgts[to], vwgt[higain]);
where[higain] = 2;
BNDInsert(nbnd, bndind, bndptr, higain);
edegrees = rinfo[higain].edegrees;
edegrees[0] = edegrees[1] = 0;
for (j=xadj[higain]; j<xadj[higain+1]; j++) {
k = adjncy[j];
if (where[k] == 2)
rinfo[k].edegrees[to] -= vwgt[higain];
else
edegrees[where[k]] += vwgt[k];
}
/* Push nodes out of the separator */
for (j=mptr[nswaps]; j<mptr[nswaps+1]; j++) {
k = mind[j];
ASSERT(where[k] == 2);
where[k] = other;
INC_DEC(pwgts[other], pwgts[2], vwgt[k]);
BNDDelete(nbnd, bndind, bndptr, k);
for (jj=xadj[k]; jj<xadj[k+1]; jj++) {
kk = adjncy[jj];
if (where[kk] == 2)
rinfo[kk].edegrees[other] += vwgt[k];
}
}
}
ASSERT(mincut == pwgts[2]);
IFSET(ctrl->dbglvl, METIS_DBG_REFINE,
printf("\tMinimum sep: %6"PRIDX" at %5"PRIDX", PWGTS: [%6"PRIDX" %6"PRIDX"], NBND: %6"PRIDX"\n", mincut, mincutorder, pwgts[0], pwgts[1], nbnd));
graph->mincut = mincut;
graph->nbnd = nbnd;
if (mincutorder == -1 || mincut >= initcut)
break;
}
rpqDestroy(queues[0]);
rpqDestroy(queues[1]);
WCOREPOP;
}
/***********************************************************************************
* This function is the entry point of the parallel kmetis algorithm that uses
* coordinates to compute an initial graph distribution.
************************************************************************************/
int ParMETIS_V3_PartGeomKway(idx_t *vtxdist, idx_t *xadj, idx_t *adjncy,
idx_t *vwgt, idx_t *adjwgt, idx_t *wgtflag, idx_t *numflag, idx_t *ndims,
real_t *xyz, idx_t *ncon, idx_t *nparts, real_t *tpwgts, real_t *ubvec,
idx_t *options, idx_t *edgecut, idx_t *part, MPI_Comm *comm)
{
idx_t h, i, j, npes, mype, status, nvtxs, seed, dbglvl;
idx_t cut, gcut, maxnvtxs;
idx_t moptions[METIS_NOPTIONS];
ctrl_t *ctrl;
graph_t *graph, *mgraph;
real_t balance;
size_t curmem;
/* Check the input parameters and return if an error */
status = CheckInputsPartGeomKway(vtxdist, xadj, adjncy, vwgt, adjwgt, wgtflag,
numflag, ndims, xyz, ncon, nparts, tpwgts, ubvec, options,
edgecut, part, comm);
if (GlobalSEMinComm(*comm, status) == 0)
return METIS_ERROR;
status = METIS_OK;
gk_malloc_init();
curmem = gk_GetCurMemoryUsed();
/* Setup the ctrl */
ctrl = SetupCtrl(PARMETIS_OP_GKMETIS, options, *ncon, *nparts, tpwgts, ubvec, *comm);
npes = ctrl->npes;
mype = ctrl->mype;
/* Take care the nparts == 1 case */
if (*nparts == 1) {
iset(vtxdist[mype+1]-vtxdist[mype], (*numflag == 0 ? 0 : 1), part);
*edgecut = 0;
goto DONE;
}
/* Take care of npes == 1 case */
if (npes == 1) {
nvtxs = vtxdist[1] - vtxdist[0]; /* subtraction is required when numflag==1 */
METIS_SetDefaultOptions(moptions);
moptions[METIS_OPTION_NUMBERING] = *numflag;
status = METIS_PartGraphKway(&nvtxs, ncon, xadj, adjncy, vwgt, NULL, adjwgt,
nparts, tpwgts, ubvec, moptions, edgecut, part);
goto DONE;
}
/* Setup the graph */
if (*numflag > 0)
ChangeNumbering(vtxdist, xadj, adjncy, part, npes, mype, 1);
graph = SetupGraph(ctrl, *ncon, vtxdist, xadj, vwgt, NULL, adjncy, adjwgt, *wgtflag);
gk_free((void **)&graph->nvwgt, LTERM);
/* Allocate the workspace */
AllocateWSpace(ctrl, 10*graph->nvtxs);
/* Compute the initial npes-way partitioning geometric partitioning */
STARTTIMER(ctrl, ctrl->TotalTmr);
Coordinate_Partition(ctrl, graph, *ndims, xyz, 1);
STOPTIMER(ctrl, ctrl->TotalTmr);
/* Move the graph according to the partitioning */
STARTTIMER(ctrl, ctrl->MoveTmr);
ctrl->nparts = npes;
mgraph = MoveGraph(ctrl, graph);
ctrl->nparts = *nparts;
SetupGraph_nvwgts(ctrl, mgraph); /* compute nvwgts for the moved graph */
if (ctrl->dbglvl&DBG_INFO) {
CommInterfaceData(ctrl, graph, graph->where, graph->where+graph->nvtxs);
for (cut=0, i=0; i<graph->nvtxs; i++) {
for (j=graph->xadj[i]; j<graph->xadj[i+1]; j++) {
if (graph->where[i] != graph->where[graph->adjncy[j]])
cut += graph->adjwgt[j];
}
}
gcut = GlobalSESum(ctrl, cut)/2;
maxnvtxs = GlobalSEMax(ctrl, mgraph->nvtxs);
balance = (real_t)(maxnvtxs)/((real_t)(graph->gnvtxs)/(real_t)(npes));
rprintf(ctrl, "XYZ Cut: %6"PRIDX" \tBalance: %6.3"PRREAL" [%"PRIDX" %"PRIDX" %"PRIDX"]\n",
gcut, balance, maxnvtxs, graph->gnvtxs, npes);
}
STOPTIMER(ctrl, ctrl->MoveTmr);
/* Compute the partition of the moved graph */
STARTTIMER(ctrl, ctrl->TotalTmr);
ctrl->CoarsenTo = gk_min(vtxdist[npes]+1, 25*(*ncon)*gk_max(npes, *nparts));
if (vtxdist[npes] < SMALLGRAPH
|| vtxdist[npes] < npes*20
|| GlobalSESum(ctrl, mgraph->nedges) == 0) { /* serially */
IFSET(ctrl->dbglvl, DBG_INFO,
rprintf(ctrl, "Partitioning a graph of size %"PRIDX" serially\n", vtxdist[npes]));
PartitionSmallGraph(ctrl, mgraph);
}
else { /* in parallel */
Global_Partition(ctrl, mgraph);
}
ParallelReMapGraph(ctrl, mgraph);
/* Invert the ordering back to the original graph */
ctrl->nparts = npes;
ProjectInfoBack(ctrl, graph, part, mgraph->where);
ctrl->nparts = *nparts;
*edgecut = mgraph->mincut;
STOPTIMER(ctrl, ctrl->TotalTmr);
/* Print some stats */
IFSET(ctrl->dbglvl, DBG_TIME, PrintTimingInfo(ctrl));
IFSET(ctrl->dbglvl, DBG_TIME, gkMPI_Barrier(ctrl->gcomm));
IFSET(ctrl->dbglvl, DBG_INFO, PrintPostPartInfo(ctrl, mgraph, 0));
FreeGraph(mgraph);
FreeInitialGraphAndRemap(graph);
if (*numflag > 0)
ChangeNumbering(vtxdist, xadj, adjncy, part, npes, mype, 0);
DONE:
FreeCtrl(&ctrl);
if (gk_GetCurMemoryUsed() - curmem > 0) {
printf("ParMETIS appears to have a memory leak of %zdbytes. Report this.\n",
(ssize_t)(gk_GetCurMemoryUsed() - curmem));
}
gk_malloc_cleanup(0);
return (int)status;
}
/*************************************************************************
* This function is the entry point of the initial balancing algorithm.
* This algorithm assembles the graph to all the processors and preceeds
* with the balancing step.
**************************************************************************/
void Balance_Partition(ctrl_t *ctrl, graph_t *graph)
{
idx_t i, j, nvtxs, nedges, ncon;
idx_t mype, npes, srnpes, srmype;
idx_t *vtxdist, *xadj, *adjncy, *adjwgt, *vwgt, *vsize;
idx_t *part, *lwhere, *home;
idx_t lnparts, fpart, fpe, lnpes, ngroups;
idx_t *rcounts, *rdispls;
idx_t twoparts=2, moptions[METIS_NOPTIONS], edgecut, max_cut;
idx_t sr_pe, gd_pe, sr, gd, who_wins;
real_t my_cut, my_totalv, my_cost = -1.0, my_balance = -1.0, wsum;
real_t rating, max_rating, your_cost = -1.0, your_balance = -1.0;
real_t lbsum, min_lbsum, *lbvec, *tpwgts, *tpwgts2, buffer[2];
graph_t *agraph, cgraph;
ctrl_t *myctrl;
MPI_Status status;
MPI_Comm ipcomm, srcomm;
struct {
double cost;
int rank;
} lpecost, gpecost;
IFSET(ctrl->dbglvl, DBG_TIME, starttimer(ctrl->InitPartTmr));
WCOREPUSH;
vtxdist = graph->vtxdist;
agraph = AssembleAdaptiveGraph(ctrl, graph);
nvtxs = cgraph.nvtxs = agraph->nvtxs;
nedges = cgraph.nedges = agraph->nedges;
ncon = cgraph.ncon = agraph->ncon;
xadj = cgraph.xadj = icopy(nvtxs+1, agraph->xadj, iwspacemalloc(ctrl, nvtxs+1));
vwgt = cgraph.vwgt = icopy(nvtxs*ncon, agraph->vwgt, iwspacemalloc(ctrl, nvtxs*ncon));
vsize = cgraph.vsize = icopy(nvtxs, agraph->vsize, iwspacemalloc(ctrl, nvtxs));
adjncy = cgraph.adjncy = icopy(nedges, agraph->adjncy, iwspacemalloc(ctrl, nedges));
adjwgt = cgraph.adjwgt = icopy(nedges, agraph->adjwgt, iwspacemalloc(ctrl, nedges));
part = cgraph.where = agraph->where = iwspacemalloc(ctrl, nvtxs);
lwhere = iwspacemalloc(ctrl, nvtxs);
home = iwspacemalloc(ctrl, nvtxs);
lbvec = rwspacemalloc(ctrl, graph->ncon);
/****************************************/
/****************************************/
if (ctrl->ps_relation == PARMETIS_PSR_UNCOUPLED) {
WCOREPUSH;
rcounts = iwspacemalloc(ctrl, ctrl->npes);
rdispls = iwspacemalloc(ctrl, ctrl->npes+1);
for (i=0; i<ctrl->npes; i++)
rdispls[i] = rcounts[i] = vtxdist[i+1]-vtxdist[i];
MAKECSR(i, ctrl->npes, rdispls);
gkMPI_Allgatherv((void *)graph->home, graph->nvtxs, IDX_T,
(void *)part, rcounts, rdispls, IDX_T, ctrl->comm);
for (i=0; i<agraph->nvtxs; i++)
home[i] = part[i];
WCOREPOP; /* local frees */
}
else {
for (i=0; i<ctrl->npes; i++) {
for (j=vtxdist[i]; j<vtxdist[i+1]; j++)
part[j] = home[j] = i;
}
}
/* Ensure that the initial partitioning is legal */
for (i=0; i<agraph->nvtxs; i++) {
if (part[i] >= ctrl->nparts)
part[i] = home[i] = part[i] % ctrl->nparts;
if (part[i] < 0)
part[i] = home[i] = (-1*part[i]) % ctrl->nparts;
}
/****************************************/
/****************************************/
IFSET(ctrl->dbglvl, DBG_REFINEINFO,
ComputeSerialBalance(ctrl, agraph, agraph->where, lbvec));
IFSET(ctrl->dbglvl, DBG_REFINEINFO,
rprintf(ctrl, "input cut: %"PRIDX", balance: ", ComputeSerialEdgeCut(agraph)));
for (i=0; i<agraph->ncon; i++)
IFSET(ctrl->dbglvl, DBG_REFINEINFO, rprintf(ctrl, "%.3"PRREAL" ", lbvec[i]));
IFSET(ctrl->dbglvl, DBG_REFINEINFO, rprintf(ctrl, "\n"));
/****************************************/
/* Split the processors into two groups */
/****************************************/
sr = (ctrl->mype % 2 == 0) ? 1 : 0;
gd = (ctrl->mype % 2 == 1) ? 1 : 0;
if (graph->ncon > MAX_NCON_FOR_DIFFUSION || ctrl->npes == 1) {
sr = 1;
gd = 0;
}
sr_pe = 0;
gd_pe = 1;
gkMPI_Comm_split(ctrl->gcomm, sr, 0, &ipcomm);
gkMPI_Comm_rank(ipcomm, &mype);
gkMPI_Comm_size(ipcomm, &npes);
if (sr == 1) { /* Half of the processors do scratch-remap */
ngroups = gk_max(gk_min(RIP_SPLIT_FACTOR, npes), 1);
gkMPI_Comm_split(ipcomm, mype % ngroups, 0, &srcomm);
gkMPI_Comm_rank(srcomm, &srmype);
gkMPI_Comm_size(srcomm, &srnpes);
METIS_SetDefaultOptions(moptions);
moptions[METIS_OPTION_SEED] = ctrl->sync + (mype % ngroups) + 1;
tpwgts = ctrl->tpwgts;
tpwgts2 = rwspacemalloc(ctrl, 2*ncon);
iset(nvtxs, 0, lwhere);
lnparts = ctrl->nparts;
fpart = fpe = 0;
lnpes = srnpes;
while (lnpes > 1 && lnparts > 1) {
PASSERT(ctrl, agraph->nvtxs > 1);
/* determine the weights of the two partitions as a function of the
weight of the target partition weights */
for (j=(lnparts>>1), i=0; i<ncon; i++) {
tpwgts2[i] = rsum(j, tpwgts+fpart*ncon+i, ncon);
tpwgts2[ncon+i] = rsum(lnparts-j, tpwgts+(fpart+j)*ncon+i, ncon);
wsum = 1.0/(tpwgts2[i] + tpwgts2[ncon+i]);
tpwgts2[i] *= wsum;
tpwgts2[ncon+i] *= wsum;
}
METIS_PartGraphRecursive(&agraph->nvtxs, &ncon, agraph->xadj,
agraph->adjncy, agraph->vwgt, NULL, agraph->adjwgt,
&twoparts, tpwgts2, NULL, moptions, &edgecut, part);
/* pick one of the branches */
if (srmype < fpe+lnpes/2) {
KeepPart(ctrl, agraph, part, 0);
lnpes = lnpes/2;
lnparts = lnparts/2;
}
else {
KeepPart(ctrl, agraph, part, 1);
fpart = fpart + lnparts/2;
fpe = fpe + lnpes/2;
lnpes = lnpes - lnpes/2;
lnparts = lnparts - lnparts/2;
}
}
if (lnparts == 1) { /* Case in which srnpes is greater or equal to nparts */
/* Only the first process will assign labels (for the reduction to work) */
if (srmype == fpe) {
for (i=0; i<agraph->nvtxs; i++)
lwhere[agraph->label[i]] = fpart;
}
}
else { /* Case in which srnpes is smaller than nparts */
/* create the normalized tpwgts for the lnparts from ctrl->tpwgts */
tpwgts = rwspacemalloc(ctrl, lnparts*ncon);
for (j=0; j<ncon; j++) {
for (wsum=0.0, i=0; i<lnparts; i++) {
tpwgts[i*ncon+j] = ctrl->tpwgts[(fpart+i)*ncon+j];
wsum += tpwgts[i*ncon+j];
}
for (wsum=1.0/wsum, i=0; i<lnparts; i++)
tpwgts[i*ncon+j] *= wsum;
}
METIS_PartGraphKway(&agraph->nvtxs, &ncon, agraph->xadj, agraph->adjncy,
agraph->vwgt, NULL, agraph->adjwgt, &lnparts, tpwgts, NULL, moptions,
&edgecut, part);
for (i=0; i<agraph->nvtxs; i++)
lwhere[agraph->label[i]] = fpart + part[i];
}
gkMPI_Allreduce((void *)lwhere, (void *)part, nvtxs, IDX_T, MPI_SUM, srcomm);
edgecut = ComputeSerialEdgeCut(&cgraph);
ComputeSerialBalance(ctrl, &cgraph, part, lbvec);
lbsum = rsum(ncon, lbvec, 1);
gkMPI_Allreduce((void *)&edgecut, (void *)&max_cut, 1, IDX_T, MPI_MAX, ipcomm);
gkMPI_Allreduce((void *)&lbsum, (void *)&min_lbsum, 1, REAL_T, MPI_MIN, ipcomm);
lpecost.rank = ctrl->mype;
lpecost.cost = lbsum;
if (min_lbsum < UNBALANCE_FRACTION * (real_t)(ncon)) {
if (lbsum < UNBALANCE_FRACTION * (real_t)(ncon))
lpecost.cost = (double)edgecut;
else
lpecost.cost = (double)max_cut + lbsum;
}
gkMPI_Allreduce((void *)&lpecost, (void *)&gpecost, 1, MPI_DOUBLE_INT,
MPI_MINLOC, ipcomm);
if (ctrl->mype == gpecost.rank && ctrl->mype != sr_pe)
gkMPI_Send((void *)part, nvtxs, IDX_T, sr_pe, 1, ctrl->comm);
if (ctrl->mype != gpecost.rank && ctrl->mype == sr_pe)
gkMPI_Recv((void *)part, nvtxs, IDX_T, gpecost.rank, 1, ctrl->comm, &status);
if (ctrl->mype == sr_pe) {
icopy(nvtxs, part, lwhere);
SerialRemap(ctrl, &cgraph, ctrl->nparts, home, lwhere, part, ctrl->tpwgts);
}
gkMPI_Comm_free(&srcomm);
}