AMFQ: Approximating Multi-level Feedback Queue Scheduling at Line Rate for Data Center Networks.

Datacenter applications demand a fast response time for short messages to enhance the user experience. Therefore, optimizing the flow completion time (FCT) for short flows is a primary objective for data center networks. Multilevel feedback queue (MLFQ) scheduling has been a widely used algorithm for process scheduling. In this paper, we propose an approximate MLFQ (AMFQ) scheduling algorithm that runs on programmable switches at line rate for data center networks. AMFQ separates incoming network traffic into multiple priority queues, ensuring low queueing delays for short flows. AMFQ adapts the time quantum concept of MLFQ. The number of time quantum will increase proportionally, while the priority of the queue decreases. The number of quantums will decide how long the packets of the flow will be served in the queue. However, implementing AMFQ using a programmable switch with millions of timers to attach to millions of flows is not feasible. Therefore, we propose to group a set of flows to form a version to be scheduled together using only one timer. We evaluated the performance of AMFQ on a P4-based testbed and compared it with existing solutions. The results show that AMFQ achieves better performance while reducing FCT for short flows, making it a promising solution for high-speed data center networks.