SC21 Proceedings

The International Conference for High Performance Computing, Networking, Storage, and Analysis

Hashed-Coordinate Storage of Sparse Tensors

Authors: Robert Lowe (Southeast Missouri State University); MeiLi Charles (University of Tennessee, Knoxville); and Amritpreet Singh (Southeast Missouri State University)

Abstract: Tensors, or n-way arrays, are becoming increasingly important in many fields. In recent applications, tensors are extremely sparse and have such a high degree that dense storage becomes intractable. To address this, several sparse storage formats have been proposed, with most formats being some variation of coordinate-based storage, or COO. These formats are created to optimize the matricized-tensor-times-Khatri-Rao product (MTTKRP) operation, which they often achieve at the expense of insertion complexity. This paper presents a new hash-based format called HaCOO. HaCOO provides the same speed of insertion and access as is found in dense arrays while storing only non-zero values. This improves upon existing sparse formats by providing efficiency in both decomposition and other tensor operations.

Best Poster Finalist (BP): no

Poster: PDF
Poster summary: PDF

Back to Poster Archive Listing