Scalable Web server cluster design with workload-aware request distribution strategy WARD

We consider a Web cluster in which the content-aware distribution is performed by each of the nodes in a Web cluster. Each server in the cluster may forward a request to another node based on the requested content. We propose a new workload-aware request distribution strategy WARD, that assigns a small set of most frequent files, called core, to be served locally, by any server in a cluster, while partitioning the rest of the files to be served by different cluster nodes. We propose an algorithm, called ward-analysis, to compute the nearly optimal core size. The algorithm takes into account workload access patterns and cluster parameters such as number of nodes, node RAM, TCP handoff overhead, and disk access overhead. Our simulations driven by a realistic workload show that WARD achieves super-linear speedup with increased cluster size. It shows superior performance compared with traditional round-robin strategy (up to 260% increased throughput for a cluster of 16 nodes), and outperforms a pure partitioning strategy based on a cache-affinity requests distribution (up to 50% increased throughput for a cluster of 16 nodes)