Improved Computation-Communication Trade-Off for Coded Distributed Computing using Linear Dependence of Intermediate Values

Abstract

In large scale distributed computing systems, communication overhead is one of the major bottlenecks. In the map-shuffle-reduce framework, which is one of the major distributed computing frameworks, the communication load among servers can be reduced by increasing the computation load of each server, that is, there is a trade-off between computation load and communication load. Recently, it has been shown that coded distributed computing (CDC) improves this trade-off relationship by letting servers encode their intermediate computation results. The original CDC scheme does not assume any special structures on the functions that servers compute. However, in actual problems, these functions often have some structures, and the trade-off relation may be further improved by using that structures. In this paper, we propose a new scheme that further improves the trade-off relationship by utilizing the linear dependency structure of the intermediate computation results. The intermediate values computed in the map phase can be considered as vectors on $\mathbb{F}_{2}$. In some applications, these intermediate values have a linear dependency and in such cases, it is sufficient for each server to send a basis of the linear subspace and linear combination coefficients. As a result, the proposed approach improves over the best-known computation-communication overhead trade-off in some applications.