Price of Precision in Coded Distributed Matrix Multiplication: A Dimensional Analysis

TL;DR

This paper analyzes the impact of precision levels on coded distributed matrix multiplication schemes, revealing that approximate codes may underperform compared to simple replication when considering precision constraints.

Contribution

The paper introduces a dimensional analysis framework for AMD codes, highlighting limitations in their efficiency when accounting for precision levels, compared to trivial replication schemes.

Findings

AMD codes fall short of trivial replication in efficiency when precision is considered

Trivial replication achieves better recovery threshold and lower complexity

Precision levels significantly influence the performance of coded matrix multiplication schemes

Abstract

Coded distributed matrix multiplication (CDMM) schemes, such as MatDot codes, seek efficient ways to distribute matrix multiplication task(s) to a set of $N$ distributed servers so that the answers returned from any $R$ servers are sufficient to recover the desired product(s). For example, to compute the product of matrices ${\bf U, V}$, MatDot codes partition each matrix into $p>1$ sub-matrices to create smaller coded computation tasks that reduce the upload/storage at each server by $1/p$, such that ${\bf UV}$ can be recovered from the answers returned by any $R=2p-1$ servers. An important concern in CDMM is to reduce the recovery threshold $R$ for a given storage/upload constraint. Recently, Jeong et al. introduced Approximate MatDot (AMD) codes that are shown to improve the recovery threshold by a factor of nearly $2$, from $2p-1$ to $p$. A key observation that motivates our work is that the storage/upload required for approximate computing depends not only on the dimensions of the (coded) sub-matrices that are assigned to each server, but also on their precision levels -- a critical aspect that is not explored by Jeong et al. Our main contribution is a dimensional analysis of AMD codes inspired by the Generalized Degrees of Freedom (GDoF) framework previously developed for wireless networks, which indicates that for the same upload/storage, once the precision levels of the task assignments are accounted for, AMD codes surprisingly fall short in all aspects to even the trivial replication scheme which assigns the full computation task to every server. Indeed, the trivial replication scheme has a much better recovery threshold of $1$, better download cost, better computation cost, and much better encoding/decoding (none required) complexity than AMD codes. The dimensional analysis is supported by simple numerical experiments.