Detecting Stragglers in Programmable Data Plane

Flow scheduling mechanisms in modern datacenters aim to reduce flow completion time (FCT). However, scheduling mechanisms that operate without prior knowledge, such as PIAS (NSDI 2015), or with imprecise flow information like QClimb (NSDI 2024), can inadvertently introduce stragglers–packets within a flow that experience significantly higher queueing delays than others. These stragglers can lead to prolonged FCT, undermining the goals of flow scheduling. While existing network monitoring tools focus on root causes of performance bottlenecks, they lack mechanisms for detecting ”victims” of such issues. In this paper, we present STRAGFLOW, a data-plane tool for straggler detection. STRAGFLOW monitors queueing delays at line rate, identifies stragglers in realtime, and reports them to the control plane. We evaluate STRAGFLOW using real-world network traces and demonstrate that it can effectively detect stragglers across different scheduling schemes and various link conditions. Our results show that STRAGFLOW can provide valuable insights into straggler distribution, helping operators diagnose and mitigate flow scheduling issues to improve overall network performance.