Content-Adaptive Request Distribution Strategy for Internet Server Cluster

Abstract

This paper addresses a distribution strategy for an Internet server cluster where the content-adaptive distribution is performed by each of the front-end node in the cluster. The system architecture considered here is a hybrid one consisting of a set of logical front-end dispatcher nodes and a set of back-end server nodes. Each front-end node in the cluster may service a request locally or forward it to another node based on the request content. This paper suggests a new distribution strategy called CARD (Content-Adaptive Request Distribution) that assigns most frequently used files to be hot which is served locally on each front-end node, while making the rest of the files to be partitioned and served among the back-end nodes. We present and evaluate the optimal configuration and hot size. The approach takes into account the file access patterns and the cluster parameters such as the number of nodes, node memory, TCP handoff overheads, data consistency overheads and disk access overheads. The simulation results show that the CARD achieves a linear speedup with the cluster size and that the CARD outperforms both the traditional centralized and distributed strategies, and outperforms a pure partitioning and replication strategy.