BEGIN:VCALENDAR
PRODID:-//Microsoft Corporation//Outlook MIMEDIR//EN
VERSION:2.0
BEGIN:VEVENT
DTSTART:20151119T220000Z
DTEND:20151119T223000Z
LOCATION:18CD
DESCRIPTION;ENCODING=QUOTED-PRINTABLE:ABSTRACT: We investigate efficient parallelization of the most common iterative sparse tensor decomposition algorithms on distributed memory systems. A key operation in each iteration of these algorithms is the matricized tensor times Khatri-Rao product (MTTKRP), which amounts to element-wise vector multiplication and reduction depending on the sparsity of the tensor. We investigate fine and coarse-grain task definitions for this operation, and propose hypergraph partitioning-based methods for these to achieve load balance as well as reduce the communication requirements. We also design distributed memory sparse tensor library, HyperTensor, which implements a well-known algorithm for the CANDECOMP/PARAFAC(CP) decomposition utilizing these task definitions and partitions. We use this library to test the scalability of the proposed implementation of MTTKRP in CP decomposition context upto 1024 MPI ranks. We observed up to 194 fold speedups using 512 MPI processes on a real world data, and significantly better scalability than a state of the art implementation.
SUMMARY:Scalable Sparse Tensor Decompositions in Distributed Memory Systems
PRIORITY:3
END:VEVENT
END:VCALENDAR