Availability and performance oriented availability modeling of webserver cache hierarchies

Abstract

Large scale Webserver configurations are widely implemented by Internet portals, and hence have evolved to rely upon numerous performance and high availability system mechanisms. The principle objective of this paper is to provide a method to compare key hierarchical features of Webserver configurations with respect to availability. Towards this goal, we model their structure and behavioral characteristics using stochastic reward nets. Several performance and availability measures are also introduced to compare two configurations: (1) a single cluster with multiple servers each featuring a percentage of its workload cached in main memory and (2) a two cluster site, a minimal star architecture providing benefits of geographic hierarchy. These are differentiated by the varying rates of cache service afforded by in-memory and on-disk caching, and the varying rates of cache service imposed by network latency in a distributed hierarchy. Both rely behaviorally upon the cache array routing protocol (CARP) to provide parent-child hierarchical caching, distributed parallel processing, hash based load balancing and dynamic repair and failover to peer systems. Principally, our results show that using a local cache hierarchy of both in-memory and disk caching significantly offsets the disadvantage of increased network latency occasioned by a geographically centralized site. Specifically, a geographic hierarchy alone is shown to be a much less efficient means of accommodating a heavy read workload.