2019/07/12 by Adarsh M. Subramaniam, Subramaniam, Adarsh M., Anoosheh Heidarzadeh +3 · 1 citation
Computer Science · Engineering · #Cooperative Communication and Network Coding #Stochastic Gradient Optimization Techniques #Sparse and Compressive Sensing Techniques
paper · pdf · doi:10.48550/arxiv.1907.05965
We propose a class of codes called random Khatri-Rao-Product (RKRP) codes for\ndistributed matrix multiplication in the presence of stragglers. The main\nadvantage of the proposed codes is that decoding of RKRP codes is highly\nnumerically stable in comparison to decoding of Polynomial codes and decoding\nof the recently proposed OrthoPoly codes. We show that RKRP codes are maximum\ndistance separable with probability 1. The communication cost and encoding\ncomplexity for RKRP codes are identical to that of OrthoPoly codes and\nPolynomial codes and the average decoding complexity of RKRP codes is lower\nthan that of OrthoPoly codes. Numerical results show that the average relative\nL2-norm of the reconstruction error for RKRP codes is substantially better\nthan that of OrthoPoly codes.\n