Coflow Scheduling in Hybrid-Switched Data Center Networks under Not-All-Stop Reconfiguration

Abstract

Scheduling parallel data flows (a coflow) across two computation stages of a job over data center networks (DCNs) is crucial to the completion of the job. To meet the growing demands of data-intensive applications, the hybrid-switched design combining an optical circuit switch (OCS) and an electrical packet switch (EPS) has emerged as a promising solution. The architecture enables high-throughput OCS transmission for large-volume traffic and utilizes EPS to handle small-volume residual demand. The key challenge is to minimize the coflow completion time (CCT) while accounting for the OCS reconfiguration delay and the EPS bandwidth constraint. Existing solutions are typically either heuristic or provide performance guarantees only under the all-stop (synchronous) reconfiguration setting, which requires pausing all circuits during any reconfiguration. This paper considers the more flexible and realistic setting of not-all-stop (asynchronous) reconfiguration, in which only affected ports are interrupted during reconfiguration. We present the first approximation algorithm with an \(\mathcal {O}(\tau _D)\) performance guarantee for single-coflow scheduling in an N × N hybrid-switched DCN under the not-all-stop setting, where τD ≤ N is the maximum number of nonzero entries in any row or column of the coflow demand matrix D. We then extend our study to the multiple-coflow setting and propose a heuristic algorithm that converts specific pure-EPS schedulers into a hybrid OCS-EPS scheduler. Experimental results on real-world Facebook data traces demonstrate the superiority of our coflow scheduling algorithms over state-of-the-art schemes, achieving shorter (total weighted) CCT.