On application-level load balancing in FastReplica

In the paper, we consider the problem of distributing large-size content to a fixed set of nodes. In contrast with the most existing end-system solutions to this problem, FastReplica [L. Cherkasova, J. Lee, FastReplica: efficient large file distribution within content delivery network, in: 4th USENIX Symposium on Internet Technologies and Systems (USITS), 2003] does not attempt to build a ‘good’ overlay structure, but simply uses a fixed mesh overlay structure. This can significantly reduces the overheads incurred in probing, building and maintaining the overlay structure, otherwise. However, FastReplica is oblivious to heterogeneous and dynamic environments. To remedy this problem, we propose an application-level load balancing idea: putting more data on ‘good’ paths and less on ‘bad’ ones. Our goal is to study (1) how to make FastReplica adaptive to dynamic environments and (2) how much performance gain can be achieved by exploring the application-level load balancing idea in FastReplica.Toward this end, we provide a theoretical analysis of a simplified model, which provides the insights serving as a basis to develop an implementation of this concept. Then, we present a performance evaluation on a wide-area testbed with a prototype implementation, showing that addition of application-level load balancing in FastReplica can achieve significant speedups by exploiting heterogeneous paths and dynamically adapting to bursty traffic.