Sparsity-Aware Distributed Tensor Decomposition

Student: Zheng Miao (Clemson University)
Supervisor: Jiajia Li (Pacific Northwest National Laboratory)

Abstract: Tensors are used by a wide range of applications as data structures to model multi-dimensional data. Tensor decomposition is a class of methods for latent data analytics. This work presents a sparsity-aware tensor decomposition on a distributed memory system. We optimize the CANDECOMP/PARAFAC decomposition, a popular low-rank tensor decomposition used in a wide variety of applications. We first throughly investigate and gain some insights of a state-of-the-art implementation, which guide our optimization direction. To solve these problems, we propose three optimization strategies: predicting the optimal grid configuration, tensor dimensions-oriented data distribution, and overlapping computation and communication. Our proposed sparsity-aware distributed CANDECOMP/PARAFAC decomposition, outperforms the state-of-the-art distributed SPLATT library by up to 2.3 X on 64 CPU nodes.

ACM-SRC Semi-Finalist: no

Poster: PDF
Poster Summary: PDF


Back to Poster Archive Listing