On Gradient Coding with Partial Recovery

TL;DR

This paper introduces a generalized gradient coding framework that allows for partial gradient recovery, providing bounds and schemes to optimize computation and communication loads in distributed systems.

Contribution

It derives lower bounds on computation load for partial gradient recovery and proposes schemes that achieve these bounds with improved communication efficiency.

Findings

Lower bounds on computation load for partial gradient recovery.

Schemes achieving optimal bounds with high communication load.

Practical schemes with balanced computation and communication loads.

Abstract

We consider a generalization of the gradient coding framework where a dataset is divided across $n$ workers and each worker transmits to a master node one or more linear combinations of the gradients over its assigned data subsets. Unlike the conventional framework which requires the master node to recover the sum of the gradients over all the data subsets in the presence of straggler workers, we relax the goal to computing the sum of at least some $\alpha$ fraction of the gradients. We begin by deriving a lower bound on the computation load of any scheme and also propose two strategies which achieve this lower bound, albeit at the cost of high communication load and a number of data partitions which can be polynomial in $n$. We then propose schemes based on cyclic assignment which utilize $n$ data partitions and have a lower communication load. When each worker transmits a single linear combination, we prove lower bounds on the computation load of any scheme using $n$ data partitions. Finally, we describe a class of schemes which achieve different intermediate operating points for the computation and communication load and provide simulation results to demonstrate the empirical performance of our schemes.