Fill estimation for blocked sparse matrices and tensors

Abstract

Many sparse matrices and tensors from a variety of applications, such as finite element methods and computational chemistry, have a natural aligned rectangular nonzero block structure. Researchers have designed high-performance blocked sparse operations which can take advantage of this sparse structure to reduce the complexity of storing the locations of nonzeros. The performance of a blocked sparse operation depends on how well a particular blocking scheme, or tiling of the sparse matrix into blocks, reflects the structure of nonzeros in the tensor. Since sparse tensor structure is generally unknown until runtime, blocking-scheme selection must be efficient. The fill is a quantity which, for some blocking scheme, relates the number of nonzero blocks to the number of nonzeros. Many performance models use the fill to help choose a blocking scheme. The fill is expensive to compute exactly, however. This thesis presents a sampling-based algorithm called PHIL that efficiently estimates the fill of sparse matrices and tensors in any format. Much of the thesis will appear in a paper coauthored with Peter Ahrens and Nicholas Schiefer. We provide theoretical guarantees for sparse matrices and tensors, and experimental results for matrices. The existing state-of-the-art fill-estimation algorithm, which we will call OSKI, runs in time linear in the number of elements in the tensor. In contrast, the number of samples PHIL needs to compute a fill estimate is unrelated to the number of nonzeros in the tensor. We compared PHIL and OSKI on a suite of hundreds of sparse matrices and found that on most inputs, PHIL estimates the fill at least 2 times faster and often more than 20 times faster than OSKI. PHIL consistently produced accurate estimates and was faster and/or more accurate than OSKI on all cases. Finally, we found that PHIL and OSKI produced comparable speedups in parallel blocked sparse matrix-vector multiplication.