Scheduling Coflows in Multi-Core OCS Networks with Performance Guarantee

Scheduling Coflows in Multi-Core OCS Networks with Performance Guarantee (arXiv:2604.08242v5) is a research paper by Xin Wang, Hong Shen, Hui Tian, and Dong Wang that addresses coflow scheduling in modern data center networks using multiple independent optical circuit switching (OCS) cores. Published on September 23, 2026, the study fills a gap in existing literature, which primarily focused on single-core OCS or multi-core electrical packet switching (EPS) networks.

The authors propose an approximation algorithm that jointly integrates cross-core flow assignment and per-core circuit scheduling to minimize the total weighted coflow completion time (CCT). This approach operates under the not-all-stop reconfiguration model, where reconfiguring one circuit does not interrupt others, addressing challenges like cross-core coupling and port exclusivity. The paper provides a provable worst-case performance guarantee and demonstrates through trace-driven simulations using real Facebook workloads that the algorithm effectively reduces both total weighted and tail CCT. The same framework can also be applied to multi-core EPS networks with corresponding approximation guarantees.

Problem and setting. The paper studies coflow scheduling in data center networks built from multiple independent optical circuit switching (OCS) cores. Coflows model application-level communication jobs composed of many parallel flows, and scheduling them as a unit can reduce job completion times relative to per-flow scheduling. In a multi-core OCS fabric, bandwidth is provisioned through coarse-grained optical circuits, with each core providing a separate set of lightpaths. This creates a scheduling problem that is more constrained than conventional packet-switched coflow scheduling: the scheduler must decide which coflows run when, how their flows are mapped to cores, and how circuit setup/teardown timing affects overall job completion.

Key contributions and insights. The central contribution is a scheduling framework for coflows in multi-core OCS networks with performance guarantees. Rather than treating the optical fabric as a single monolithic network, the work accounts for the parallelism and discreteness of multiple OCS cores, where circuit-level allocation, core contention, and reconfiguration timing jointly determine achievable completion times. The paper’s insight is that predictable coflow performance in such networks requires coordinated scheduling across cores, not merely routing or bandwidth allocation within one fabric. By providing provable performance bounds, the work moves beyond heuristic throughput optimization toward schedulers that can guarantee bounded completion-time behavior under network constraints.

Why it matters. As data center networks adopt OCS and multi-plane architectures to meet large bandwidth demands with lower power and latency, scheduling must increasingly operate at the application-job level. Coflows are a natural abstraction for collective and job-scoped communication, and tail completion time is often the critical metric in real workloads. A scheduling approach with guarantees for multi-core OCS networks is therefore relevant both to network designers evaluating optical architectures and to systems researchers seeking principled methods for delivering predictable performance in reconfigurable, circuit-based data center fabrics.