Squeezed Random Khatri-Rao Product Codes

Authors

Image provided by Ruowan Ji
Ruowan
Ji
Texas A&M University
Profile
Asit
Pradhan
Texas A&M University
Profile picture for user AnooshehHeidarzadeh
Anoosheh
Heidarzadeh
Texas A&M University
Profile
Krishna
Narayanan
Texas A&M University

Abstract

We introduce a class of codes, called Squeezed Random Khatri-Rao Product (RKRP) codes, for coded matrix multiplication when each worker node can perform multiple submatrix products. The proposed codes are a generalization of RKRP codes in [1] and are built on the idea of squeezed polynomial codes in [2]. We show that squeezed RKRP codes are maximum distance separable with probability 1. They have the same communication cost as that of squeezed polynomial codes while offering better numerical stability.

Paper Manuscript