Sparse Data Structures for Weighted Bipartite Matching

Inspired by the success of blocking to improve the performance of algorithms for sparse matrix vector multiplication [5] and sparse direct factorization [1], we explore the benefits of blocking in related sparse graph algorithms. A natural question is whether the benefits of local blocking extend to other sparse graph algorithms. Here we examine algorithms for finding a maximum-weight complete matching in a sparse bipartite graph, otherwise known as the linear sum assignment problem, or lsap. Like sparse matrix-vector products, algorithms for lsap combine data from two sources. But rather than computing a dot product, the use a complicated max reduction or depth-first search (dfs). Maximum weight matchings find uses in sparse matrix algorithms which could otherwise benefit from blocked matrix representations. A matching provides statically chosen pivots for parallel sparse LU factorization [6]. The LU factorization is followed by matrix-vector products in iterative refinement, so we would like take a blocked matrix structure as input. Preliminary work shows that the blocked representations do not decrease performance significantly although they do increase the number of operations performed, and so we can use a matrix data structure that improves the iterative refinement performance.