RP-DFC: Responsive Probes and Dynamic Flow Classification based load balancing in datacenter networks

Abstract

Datacenter networks achieve high bandwidth by establishing multiple accessible paths between hosts. This necessitates a load-balancing approach to effectively select optimal paths for data flows, minimizing transmission delays and path congestion. The current congestion awareness and load-balancing methods that rely on active detection encounter issues with substantial bandwidth overhead from probes and overlook the distribution characteristics of network traffic. This paper introduces RP-DFC as a distributed load-balancing approach that utilizes responsive probes and flow classification within the data plane. RP-DFC employs in-band network telemetry and active detection to create a responsive probe congestion awareness mechanism. This mechanism can adaptively adjust the detection frequency according to the network congestion status, significantly decreasing the bandwidth overhead of active detection. RP-DFC further enhances network performance in high-load scenarios by employing advanced large-flow and small-flow classification techniques tailored to data centers’ unique traffic distribution characteristics. This strategic implementation at the network edge optimizes traffic management, significantly outperforming existing methods. RP-DFC exhibits a substantial 30% performance improvement over HULA under high-load conditions, concurrently reducing probe overhead by an impressive 98%. Moreover, when benchmarked against alternative methods like W-ECMP, RP-DFC achieves a notable 29% enhancement in performance, highlighting its effectiveness in optimizing data center network operations.