High-Level Load Balancing for Web Services

A structured approach to high availability and fault tolerance is essential in a production-grade service delivery network, where delays and faults can occur for a multitude of reasons. In this thesis, we consider the high-level scheduling and load sharing properties offered by the Domain Name System, as implemented in popular DNS software packages. At this level, the scheduling mechanism can account for server availability, geographical proximity, time zones, etc. We explore the performance and capabilities of high-level DNS-based load balancing, where we draw special attention to the choice of caching policy (time-to-live) for DNS data. Our findings confirm the high performance of modern DNS server implementations, but question the use of DNS as a suitable load balancing mechanism in itself. Further, we analyse the use of a database-supported DNS service for allowing highly dynamical query responses, and show that this approach has both potentially negative (single point of failure) and positive (improved balancing flexibility) properties