| Lecturer Profile |
Kai Chen is an Assistant Professor with CSE department at HKUST. He received his BS/MS from University of Science and Technology of China in 2004/2007 and PhD from Northwestern University in 2012, all in Computer Science. His research interests include networked systems design and implementation, data center networking, data centric networking, and cloud and big data systems. His work has been published in various top conferences and journals such as SIGCOMM, NSDI, IEEE/ACM TON and so on. One of his papers was a SIGCOMM’10 best paper candidate, fast-tracked to IEEE/ACM TON. More information can be found at: http://www.cse.ust.hk/~kaichen/ |
| Lecture Abstract |
Our group has been working on datacenter networks for years. In this talk, I will first overview a lab-scale 100-server datacenter cluster we built and research works we conducted based on it in the last 4 years. Then, I will pick one of them called PIAS, Practical Information-Agnostic flow Scheduling, and discuss the details. PIAS was motivated by the observation that many prior datacenter flow scheduling schemes minimize flow completion times (FCT) assuming prior knowledge of flows and custom switch functions. This makes them superior in performance, but hard to use in practice. By PIAS, we present a scheme to minimize FCT with no prior knowledge and existing commodity switch functions. At its heart, PIAS leverages multiple priority queues available in existing commodity switches to implement a Multiple Level Feedback Queue (MLFQ), in which a PIAS flow is gradually demoted from higher-priority queues to lower-priority queues based on the amount of bytes it has sent. Consequently, short flows are likely to be completed in the first few high-priority queues and thus be prioritized over long flows in general, which enables PIAS to emulate SJF without knowing flow sizes in advance. PIAS is readily implementable with commodity switches and backward compatible with legacy TCP/IP stacks. We have shown its performance through both testbed experiments and ns2 simulations. To make our work easy to reproduce, we made the implementation code online at: http://sing.cse.ust.hk/projects/PIAS. |
