Michal Szymaniak's Thesis Page

Latency-Driven Replication for Globally Distributed Systems

As the Internet becomes the predominant communication medium for our civilization, many Internet services are continuously accessed by vast numbers of clients. Services unable to cope with so many clients increasingly often suffer from long reaction times, which discourages clients from using such services more often, reducing the revenue and impact they can generate.

This thesis proposes to reduce the time (latency) necessary to communicate with an Internet service by means of replication, in which multiple instances (replicas) of that service are deployed in different parts of the Internet. We present techniques for efficient modeling of Internet latencies in globally distributed environments, and demonstrate how our models enable one to quickly choose locations for new replicas such that client-to-service latencies are reduced. We also show how to organize physical machines around the replica locations into virtual hosting platforms for the actual replicas, such that they can continue to service clients despite failures of individual machines. Our experiments indicate that the proposed techniques improve response times and availability of Internet services, and that they can also be applied in other globally distributed systems.