Using automatically derived load thresholds to manage compute resources on-demand

Dynamic computing environments that support a changing application set and load are now a reality. Within these environments the decision to add or remove resources, must be made in a timely manner, while at the same time they must avoid excessive resource rebalancing. When the overhead incurred by the resource reallocation process is significant, we should be confident that additional resources are necessary before initiating an allocation process. This process is complicated by the instability of application content and frequent changes in average request processing time. Load and response time thresholds need to be dynamically and automatically adjusted if they are to remain effective. We investigate automated methods for selecting and setting thresholds, estimating load, and analyzing response time. Specifically, these methods estimate the load on a set of application servers, and set appropriate resource allocation thresholds based on the relationship between projected site response times and server load. We describe a solution to the problem of determining the maximum load over multiple servers under changing conditions, which include both changing traffic and application characteristics.