data center network schedulers

Abstract: The growing scale of data and models in machine learning calls for efficient optimization methods. Distributed optimization addresses this challenge, with compression mechanisms helping to reduce communication costs. A classical tool for dimensionality reduction is the Johnson-Lindenstrauss (JL) transform, which is simple to implement and offers strong geometric guarantees. In particular, JL transforms approximately preserve inner products and pairwise distances between vectors with high probability. In this paper, we exploit this property in Byzantine-robust optimization, where worker messages are weighted according to their alignment with a trusted reference gradient. We further support our theoretical findings with experimental results.

Abstract: This paper studies coded distributed computing (CDC) in a dynamic system in which workers may depart and new clusters may join. The caches of the surviving workers and their existing Reduce assignments stay untouched, while the storage brought by arriving clusters is put to use. In contrast to elastic computing, which targets linear functions, and to dynamic coded caching, which requires placement across all users, the model considered here accommodates general MapReduce tasks under a placement that is partly fixed and partly new. The proposed scheme builds cluster-wise placement delivery arrays (PDAs) and couples them through a Cartesian product. A delivery-aware rule reassigns the abandoned Reduce functions, and a generalized communication PDA allows even workers that hold no Reduce function to serve as coded transmitters. For arbitrary feasible disconnections, the exact load is derived. A file-wise converse for instantly decodable XOR multicasts shows that, in the absence of disconnections, the scheme lies within a factor of two of the best scheme in this multicast class when new clusters arrive, and within a factor of four otherwise, even when the benchmark optimizes over all feasible uncoded placements.