The TranSend service runs on a cluster of high performance workstations. Client requests are load balanced across machines in the cluster in order to maximize request throughput and minimize the end-to-end latency of each request through the system [15]. As observed in the Home IP traces, the load presented to TranSend is quite bursty on time scales on the order of seconds. Fortunately, the service time for an individual request is on the order of milliseconds; if a burst of traffic arrives at the system, it takes only a few seconds for the backlog associated with that burst to be cleared from the system.
Over longer time scales, we have indeed observed relatively stable, non-bursty load. Certain real-world events (such as the publication of an article in the campus newspaper about the service) did trigger temporary load bursts that persisted for hours, however these bursts were extremely rare (they have only occurred two or three times during the 4 months that TranSend has been active). Because of this long-term smoothness, we were able to allocate a fixed number of cluster nodes to TranSend. To handle the infrequent long-term bursts of activity, we designed TranSend to easily recruit ``overflow nodes'' in times of need.