Exploiting Host Availability in Distributed Systems.

As distributed systems become more decentralized, fluctuating host availability is an increasingly disruptive phenomenon. Older systems such as AFS used a small number of well-maintained, highly available machines to coordinate access to shared client state; server uptime (and thus service availability) were expected to be high. Newer services scale to larger number of clients by increasing the number of servers. In these systems, the responsibility for maintaining the service abstraction is spread amongst thousands of machines. In the extreme, each client is also a server who must respond to requests from its peers, and each host can opt in or out of the system at any time. In these operating environments, a non-trivial fraction of servers will be unavailable at any give time. This diffusion of responsibility from a few dedicated hosts to many unreliable ones has a dramatic impact on distributed system design, since it is difficult to build robust applications atop a partially available, potentially untrusted substrate. This dissertation explores one aspect of this challenge: how can a distributed system measure the fluctuating availability of its constituent hosts, and how can it use an understanding of this churn to improve performance and security? This dissertation extends the previous literature in three ways. First, it introduces new analytical techniques for characterizing availability data, applying these techniques to several real networks and explaining the distinct uptime patterns found within. Second, this dissertation introduces new methods for predicting future availability, both at the granularity of individual hosts and clusters of hosts. Third, my dissertation describes how to use these new techniques to improve the performance and security of distributed systems.PhDComputer Science & EngineeringUniversity of Michigan, Horace H. Rackham School of Graduate Studieshttp://deepblue.lib.umich.edu/bitstream/2027.42/58445/1/jmickens_1.pd