Please note: This master’s research paper presentation will take place online.
Brian Song, Master’s candidate
David R. Cheriton School of Computer Science
Supervisor: Professor Martin Karsten
Modern high performance network services, such as Memcached, typically aim to minimize tail latency while maintaining a high request rate. Historically, developers often choose between two main architectures: the event-driven model and the thread-per-session model. The former tends to maximize concurrency but comes with increased programming complexity. The latter provides developers with more intuitive synchronization logic, but suffers from scalability limitations due to substantial kernel overhead introduced by the standard Linux scheduler.
This study presents a systematic analysis of the latency life cycle of short-lived, I/O bound workloads. Observations indicate that even after optimizing network interrupt affinity and socket configurations, the generic Linux scheduler continues to be a significant source of latency. In particular, the default Earliest Eligible Virtual Deadline First (EEVDF) scheduler incurs a computational overhead that is not proportional to the task execution time. To address this problem, this study introduces SCX_RR, a specialized scheduler implemented on top of the sched_ext framework. SCX_RR adopts a push-model round-robin scheduling policy combined with a lock-free work-stealing mechanism to minimize the scheduling overhead. Experimental evaluation with Memcached shows the thread-per-session model with sched_ext scheduling can maintain the advantages of the synchronous programming model, while reducing 99th-percentile queuing latency by about 70% compared to EEVDF, achieving throughput comparable to event-driven systems.
Attend this master’s research paper presentation virtually on Zoom.