Computer Science Thesis Oral

Wednesday, May 3, 2017 - 2:00pm to 3:30pm


Traffic21 Classroom 6501 Gates Hillman Centers


TIMOTHY ZHU, Ph.D. Student

In today's datacenters, storage and network resources are shared to lower costs and achieve better utilization, but how does one meet tail latency performance goals in these shared environments? Workloads exhibit different behaviors (e.g., burstiness, load) and can have different latency requirements. Furthermore, storage and networks each exhibit unique performance characteristics that make it hard to provide latency guarantees. For example, in SSD storage devices, writes are slower than reads, while in networks, packets traverse a series of queues and congest with different workloads at each queue. In this talk, I will describe my thesis work on meeting tail latency goals when sharing storage and networks. Our work looks at the problem from the perspective of scheduling policies, admission control, and workload placement. In particular, we implement new policies and mechanisms for controlling the congestion between multiple workloads of varying burstiness. Our system is unique in that it incorporates mathematically grounded techniques such as Deterministic Network Calculus (DNC) and Stochastic Network Calculus (SNC) to guarantee tail latency. In fact, we are the first to implement SNC in a computer system. Thesis Committee: Mor Harchol-Balter (Chair) Gregory R. Ganger David G. Andersen Michael A. Kozuch (Intel Labs) Arif Merchant (Google)

For More Information, Contact:


Thesis Oral