Straggler-Resilient Asynchronous Decentralized ADMM for Consensus Optimization.

For its ability to combine the decomposability of dual ascent methods with superior convergence properties, the alternating direction method of multiplier (ADMM) has attracted substantial research interests and applications in various fields, e.g. distributed systems. Nevertheless, the problem of stragglers (slow-response nodes) and single-point-of-failures (a single node causing the failure of the entire system) remains. Therefore, we propose a novel asynchronous decentralized ADMM algorithm to address this. Through a series of simulations, our algorithm outperforms several state-of-the-art alternatives by converging multiple times faster in all settings. The results also show the resilience of our algorithm against stragglers and single points of failure as well as its ability to achieve superior convergence speed without sacrificing accuracy.