Predictable Performance and Low Cost for Geo-Distributed Applications

Applications that run in the cloud must be geo-distributed to achieve high reliability, minimize delays, and follow data localization laws. However, geo-distributed applications face many challenges in achieving these goals due to a key bottleneck: the wide-area network (WAN). Without enough, cheaply-available WAN bandwidth, applications are forced to degrade the service they provide or shut down altogether. Additionally, the latency between data centers limits how quickly an application can serve user requests. My dissertation makes two contributions to better serve providers of geo-distributed applications. First, I present a new architecture, HEYP, for sharing a private WAN across many tenants. State-of-the-art WANs maximize efficiency by trying make use of every bit of spare capacity on the network. However, in doing so, they risk introducing interference between tenants. In contrast, HEYP offers strong isolation guarantees between tenants, but without sacrificing the efficiency of existing shared WANs. Next, I characterize the impact that latency has on read and write operations when manipulating data that is stored in multiple data centers, and how there exists a three-way tradeoff between optimizing for read latency, write latency, and cost. I study existing approaches and identify inefficiencies that cause them to require significantly higher cost than necessary to meet a set of latency targets. I then describe the design of Pando, a system that I have developed to offer a near-optimal tradeoff between read latency, write latency, and cost. Both Pando and HEYP offer substantially improved tradeoffs between performance and cost that can benefit geo-distributed applications. More generally, however, my work shows that current geo-distributed systems offer tradeoffs that are suboptimal in important ways, and that focusing on the constraints of WANs will be a key part of improving them.PHDComputer Science & EngineeringUniversity of Michigan, Horace H. Rackham School of Graduate Studieshttp://deepblue.lib.umich.edu/bitstream/2027.42/174596/1/uluyol_1.pd