Adaptive workload-dependent scheduling for large-scale content delivery systems

Content delivery has become an important application in the Internet. \Content" in this context can be a range of objects from movies to web pages to software distribution. A streaming content delivery server should provide nearly immediate and continuous service by provisioning suucient server and local network resources for the duration of playout. Because of the resource implications of delivering many large les simultaneously, scalability is an important requirement. Good scalability can be achieved by using a single channel to serve multiple users waiting for the same object (referred to as batching). Batching is especially useful during high load periods. Typical strategies in use today for allocating channels use a greedy, allocate-as-needed policy with little consideration for anything other than satisfying the current request or maximizing the number of batched requests. Macroscopic system characteristics, like request arrival patterns, have stable long-term averages, but can vary unpredictably across shorter intervals. This variability can cause scheduling algorithms to suuer poor and unpredictable short-term performance. In this paper, we propose a set of rate-based allocation algorithms to solve these limitations. We present our work in developing a set of workloads with variable request rates, quantify the drawbacks of traditional greedy channel allocation algorithms, and quantify the advantages of the rate-based algorithms. We also generalize the content delivery model and discuss when rate-based algorithms might be eeective other kinds of systems.