Demand-Aware Distributed Scheduling With Adaptive Buffer Control in Reconfigurable Data Center Networks

초록

Reconfigurable data center networks (RDCNs), integrating the electrical packet switch (EPS) with the optical circuit switch (OCS), improve network adaptability by enabling high-throughput connections between top-of-rack (ToR) pairs. However, existing RDCN scheduling schemes face challenges in responsiveness, particularly during traffic bursts. In this article, we propose a novel demand-aware distributed scheduling framework called P4-DADS, utilizing P4-based programmable ToR switches (P4ToR). To prevent conflicts arising from simultaneous OCS port allocations, P4-DADS employs a token-ring-based distributed reservation algorithm, enhanced with an adaptive buffer control (ABC) mechanism. By formulating a Markov decision process (MDP) problem, the optimal ABC policy is obtained through a value iteration algorithm, ensuring that packets are immediately ready for transmission during sudden demand surges. P4-DADS improves network responsiveness and scalability, as evidenced by a 145.95% increase in throughput and a 87.31% reduction in flow completion time. These improvements demonstrate the potential of P4-DADS as a scalable and efficient solution for resource management in RDCN.

키워드