Decentralized Stochastic Optimization and Gossip Algorithms with Compressed Communication

TL;DR

This paper introduces novel gossip-based algorithms for decentralized stochastic optimization that incorporate compressed communication, achieving efficient convergence rates and significantly reducing communication costs in distributed machine learning tasks.

Contribution

The paper proposes CHOCO-SGD and CHOCO-GOSSIP algorithms that support compressed communication with theoretical convergence guarantees, advancing decentralized optimization methods.

Findings

CHOCO-SGD achieves convergence rate similar to centralized methods despite compression.

CHOCO-GOSSIP converges linearly with arbitrary compressed messages.

Algorithms outperform state-of-the-art baselines and reduce communication by over 100 times.

Abstract

We consider decentralized stochastic optimization with the objective function (e.g. data samples for machine learning task) being distributed over $n$ machines that can only communicate to their neighbors on a fixed communication graph. To reduce the communication bottleneck, the nodes compress (e.g. quantize or sparsify) their model updates. We cover both unbiased and biased compression operators with quality denoted by $\omega \leq 1$ ($\omega=1$ meaning no compression). We (i) propose a novel gossip-based stochastic gradient descent algorithm, CHOCO-SGD, that converges at rate $\mathcal{O}\left(1/(nT) + 1/(T \delta^2 \omega)^2\right)$ for strongly convex objectives, where $T$ denotes the number of iterations and $\delta$ the eigengap of the connectivity matrix. Despite compression quality and network connectivity affecting the higher order terms, the first term in the rate, $\mathcal{O}(1/(nT))$, is the same as for the centralized baseline with exact communication. We (ii) present a novel gossip algorithm, CHOCO-GOSSIP, for the average consensus problem that converges in time $\mathcal{O}(1/(\delta^2\omega) \log (1/\epsilon))$ for accuracy $\epsilon > 0$. This is (up to our knowledge) the first gossip algorithm that supports arbitrary compressed messages for $\omega > 0$ and still exhibits linear convergence. We (iii) show in experiments that both of our algorithms do outperform the respective state-of-the-art baselines and CHOCO-SGD can reduce communication by at least two orders of magnitudes.