Isorropia: Partitioning, Load Balancing and more
graphedge_weights.cpp

This example shows how to use graph partitioning methods with edge weights.

//@HEADER
//************************************************************************
//
// Isorropia: Partitioning and Load Balancing Package
// Copyright (2006) Sandia Corporation
//
//Under terms of Contract DE-AC04-94AL85000, there is a non-exclusive
//license for use of this work by or on behalf of the U.S. Government.
//
// Redistribution and use in source and binary forms, with or without
// modification, are permitted provided that the following conditions are
// met:
//
// 1. Redistributions of source code must retain the above copyright
// notice, this list of conditions and the following disclaimer.
//
// 2. Redistributions in binary form must reproduce the above copyright
// notice, this list of conditions and the following disclaimer in the
// documentation and/or other materials provided with the distribution.
//
// 3. Neither the name of the Corporation nor the names of the
// contributors may be used to endorse or promote products derived from
// this software without specific prior written permission.
//
// THIS SOFTWARE IS PROVIDED BY SANDIA CORPORATION "AS IS" AND ANY
// EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
// IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
// PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL SANDIA CORPORATION OR THE
// CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
// EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
// PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
// PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
// LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
// NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
// SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
//
//************************************************************************
//@HEADER
//--------------------------------------------------------------------
//This file is a self-contained example of creating an Epetra_RowMatrix
//object, and using Isorropia to create a rebalanced copy of it.
//Furthermore graph-edge weights are used to influence the repartitioning.
//--------------------------------------------------------------------
//Include Isorropia_Exception.hpp only because the helper functions at
//the bottom of this file (which create the epetra objects) can
//potentially throw exceptions.
//The Isorropia symbols being demonstrated are declared
//in these headers:
#ifdef HAVE_MPI
#include <mpi.h>
#endif
#ifdef HAVE_EPETRA
#ifdef HAVE_MPI
#include <Epetra_MpiComm.h>
#else
#include <Epetra_SerialComm.h>
#endif
#include <Epetra_Map.h>
#include <Epetra_Vector.h>
#include <Epetra_CrsMatrix.h>
#include <Epetra_FECrsMatrix.h>
#endif
int main(int argc, char** argv) {
#if defined(HAVE_MPI) && defined(HAVE_EPETRA)
int numProcs = 1;
int localProc = 0;
//first, set up our MPI environment...
MPI_Init(&argc, &argv);
MPI_Comm_rank(MPI_COMM_WORLD, &localProc);
MPI_Comm_size(MPI_COMM_WORLD, &numProcs);
// This program can only run on 3 processors, to make things
// work out easy.
//
if (numProcs != 3) {
std::cout << "num-procs="<<numProcs<<". This program can only "
<< "run on 3 procs. Exiting."<<std::endl;
MPI_Finalize();
return(0);
}
//Consider the following mesh of 4 2-D quad elements:
//
// *-------*-------*
// 8| 7| 6|
// | E2 | E3 |
// *-------*-------*
// 3| 2| 5|
// | E0 | E1 |
// *-------*-------*
// 0 1 4
//
// Node-ids are to the lower-left of each node (*).
//
// Mimicing a finite-element application, we will say that
// each node has 1 scalar degree-of-freedom, and assemble
// a matrix which would have 9 global rows and columns.
//
// Each processor will have 3 rows. We'll set up a strange
// initial map, where nodes are distributed as follows:
//
// proc 0: nodes 0,3,8,
// proc 1: nodes 1,2,7
// proc 2: nodes 4,5,6.
//
// After we assemble our matrix, we'll create another matrix
// and populate it with graph edge weights such that the
// partitioner repartitions the problem so that nodes are
// laid out as follows:
//
// proc 0: nodes 0, 1, 4
// proc 1: nodes 3, 2, 5
// proc 2: nodes 8, 7, 6
//
int nodesPerElem = 4;
int global_n = 9;
//First, set up the initial map:
std::vector<int> mynodes(3);
if (localProc == 0) {
mynodes[0] = 0; mynodes[1] = 3; mynodes[2] = 8;
}
if (localProc == 1) {
mynodes[0] = 1; mynodes[1] = 2; mynodes[2] = 7;
}
if (localProc == 2) {
mynodes[0] = 4; mynodes[1] = 5; mynodes[2] = 6;
}
Epetra_MpiComm comm(MPI_COMM_WORLD);
Epetra_Map origmap(global_n, 3, &mynodes[0], 0, comm);
Teuchos::RCP<Epetra_FECrsMatrix> matrix =
Teuchos::rcp(new Epetra_FECrsMatrix(Copy, origmap, 0));
//We'll assemble elements E0 and E1 on proc 0,
// element E2 or proc 1,
// element E3 on proc 2.
std::vector<int> indices(nodesPerElem);
std::vector<double> coefs(nodesPerElem*nodesPerElem,2.0);
if (localProc == 0) {
//element E0:
indices[0] = 0; indices[1] = 1; indices[2] = 2; indices[3] = 3;
matrix->InsertGlobalValues(nodesPerElem, &indices[0], &coefs[0]);
//element E1:
indices[0] = 1; indices[1] = 4; indices[2] = 5; indices[3] = 2;
matrix->InsertGlobalValues(nodesPerElem, &indices[0], &coefs[0]);
}
else if (localProc == 1) {
//element E2:
indices[0] = 3; indices[1] = 2; indices[2] = 7; indices[3] = 8;
matrix->InsertGlobalValues(nodesPerElem, &indices[0], &coefs[0]);
}
else { //localProc==2
//element E3:
indices[0] = 2; indices[1] = 5; indices[2] = 6; indices[3] = 7;
matrix->InsertGlobalValues(nodesPerElem, &indices[0], &coefs[0]);
}
int err = matrix->GlobalAssemble();
if (err != 0) {
std::cout << "err="<<err<<" returned from matrix->GlobalAssemble()"
<< std::endl;
}
// std::cout << "matrix: " << std::endl;
// std::cout << *matrix << std::endl;
//We'll need a Teuchos::ParameterList object to pass to the
//Isorropia::Epetra::Partitioner class.
Teuchos::ParameterList paramlist;
#ifdef HAVE_ISORROPIA_ZOLTAN
// If Zoltan is available, we'll specify that the Zoltan package be
// used for the partitioning operation, by creating a parameter
// sublist named "Zoltan".
// In the sublist, we'll set parameters that we want sent to Zoltan.
paramlist.set("PARTITIONING METHOD", "GRAPH");
paramlist.set("PRINT ZOLTAN METRICS", "2");
Teuchos::ParameterList& sublist = paramlist.sublist("Zoltan");
sublist.set("GRAPH_PACKAGE", "PHG"); // Use ParMetis or Scotch if you have it
sublist.set("EDGE_WEIGHT_DIM", "1"); // One weight per edge
//sublist.set("DEBUG_LEVEL", "1"); // Zoltan will print out parameters
//sublist.set("DEBUG_LEVEL", "5"); // proc 0 will trace Zoltan calls
//sublist.set("DEBUG_MEMORY", "2"); // Zoltan will trace alloc & free
#else
// If Zoltan is not available, a simple linear partitioner will be
// used to partition such that the number of nonzeros is equal (or
// close to equal) on each processor. No parameter is necessary to
// specify this.
#endif
Teuchos::RCP<Isorropia::Epetra::CostDescriber> costs =
//Next create a matrix which is a copy of the matrix we just
//assembled, but we'll replace the values with graph edge weights.
Teuchos::RCP<Epetra_FECrsMatrix> ge_weights =
Teuchos::rcp(new Epetra_FECrsMatrix(*matrix));
Teuchos::RCP<Epetra_CrsMatrix> crs_ge_weights;
crs_ge_weights = ge_weights;
//Fill the matrix with a "default" weight of 1.0.
crs_ge_weights->PutScalar(1.0);
//Now we'll put a "large" weight on edges that connect nodes
//0 and 1, 1 and 4,
//3 and 2, 2 and 5,
//8 and 7, 7 and 6.
double weight = 500.0;
if (localProc == 0) {
//row 0, edge 1
indices[0] = 1;
coefs[0] = weight;
crs_ge_weights->ReplaceGlobalValues(0, 1, &coefs[0], &indices[0]);
//row 3, edge 2
indices[0] = 2;
coefs[0] = weight;
crs_ge_weights->ReplaceGlobalValues(3, 1, &coefs[0], &indices[0]);
//row 8, edge 7
indices[0] = 7;
coefs[0] = weight;
crs_ge_weights->ReplaceGlobalValues(8, 1, &coefs[0], &indices[0]);
}
if (localProc == 1) {
//row 1, edges 0 and 4
indices[0] = 0; indices[1] = 4;
coefs[0] = weight; coefs[1] = weight;
crs_ge_weights->ReplaceGlobalValues(1, 2, &coefs[0], &indices[0]);
//row 2, edges 3 and 5
indices[0] = 3; indices[1] = 5;
coefs[0] = weight; coefs[1] = weight;
crs_ge_weights->ReplaceGlobalValues(2, 2, &coefs[0], &indices[0]);
//row 7, edges 6 and 8
indices[0] = 6; indices[1] = 8;
coefs[0] = weight;
crs_ge_weights->ReplaceGlobalValues(7, 2, &coefs[0], &indices[0]);
}
if (localProc == 2) {
//row 4, edge 1
indices[0] = 1;
coefs[0] = weight;
crs_ge_weights->ReplaceGlobalValues(4, 1, &coefs[0], &indices[0]);
//row 5, edge 2
indices[0] = 2;
coefs[0] = weight;
crs_ge_weights->ReplaceGlobalValues(5, 1, &coefs[0], &indices[0]);
//row 6, edge 7
indices[0] = 7;
coefs[0] = weight;
crs_ge_weights->ReplaceGlobalValues(6, 1, &coefs[0], &indices[0]);
}
// std::cout << "crs_ge_weights: " << std::endl
// << *crs_ge_weights << std::endl;
//Now give the graph edge weights to the CostDescriber:
costs->setGraphEdgeWeights(crs_ge_weights);
Teuchos::RCP<const Epetra_RowMatrix> rowmatrix;
rowmatrix = matrix;
//Now create the partitioner object using an Isorropia factory-like
//function...
Teuchos::RCP<Isorropia::Epetra::Partitioner> partitioner =
Teuchos::rcp(new Isorropia::Epetra::Partitioner(rowmatrix, costs, paramlist));
//Next create a Redistributor object and use it to create a
//repartitioned copy of the matrix
Teuchos::RCP<Epetra_CrsMatrix> bal_matrix;
//Use a try-catch block because Isorropia will throw an exception
//if it encounters an error.
if (localProc == 0) {
std::cout << " calling Isorropia::Epetra::Redistributor::redistribute..."
<< std::endl;
}
try {
bal_matrix = rd.redistribute(*rowmatrix);
}
catch(std::exception& exc) {
std::cout << "linsys example: Isorropia::Epetra::Redistributor threw "
<< "exception '" << exc.what() << "' on proc "
<< localProc << std::endl;
MPI_Finalize();
return(-1);
}
// Results
double bal0, bal1, cutn0, cutn1, cutl0, cutl1, cutWgt0, cutWgt1;
int numCuts0, numCuts1;
#if 1
// Balance and cut quality before partitioning
double goalWeight = 1.0 / (double)numProcs;
ispatest::compute_graph_metrics(*rowmatrix, *costs, goalWeight,
bal0, numCuts0, cutWgt0, cutn0, cutl0);
// Balance and cut quality after partitioning
Teuchos::RCP<Epetra_CrsMatrix> new_weights = rd.redistribute(*crs_ge_weights);
new_costs.setGraphEdgeWeights(new_weights);
ispatest::compute_graph_metrics(*bal_matrix, new_costs, goalWeight,
bal1, numCuts1, cutWgt1, cutn1, cutl1);
#else
std::vector<double> bal(2), cutwgt(2), cutn(2), cutl(2);
std::vector<int >ncuts(2);
Epetra_Import &importer = rd.get_importer();
costs->compareBeforeAndAfterGraph(*rowmatrix, *bal_matrix, importer,
bal, ncuts, cutwgt, cutn, cutl);
bal0 = bal[0]; cutn0 = cutn[0]; cutl0 = cutl[0]; cutWgt0 = cutwgt[0]; numCuts0 = ncuts[0];
bal1 = bal[1]; cutn1 = cutn[1]; cutl1 = cutl[1]; cutWgt1 = cutwgt[1]; numCuts1 = ncuts[1];
#endif
bal_matrix.release();
if (localProc == 0){
std::cout << "Before partitioning: Number of cuts " << numCuts0 << " Cut weight " << cutWgt0 << std::endl;
std::cout << " Balance " << bal0 << " cutN " << cutn0 << " cutL " << cutl0;
std::cout << std::endl;
std::cout << "After partitioning: Number of cuts " << numCuts1 << " Cut weight " << cutWgt1 << std::endl;
std::cout << " Balance " << bal1 << " cutN " << cutn1 << " cutL " << cutl1;
std::cout << std::endl;
}
MPI_Finalize();
#else
std::cout << "part_redist: must have both MPI and EPETRA. Make sure Trilinos "
<< "is configured with --enable-mpi and --enable-epetra." << std::endl;
#endif
return(0);
}