Automatic Request Characterization in Internet Services

Modeling system performance and workload characteristics has become essential for efficiently provisioning Internet services and for accurately predicting future resource requirements on anticipated workloads. The accuracy of these models benefits substantially by differentiating among categories of requests based on their resource usage characteristics. However, categorizing requests and their resource demands often requires significantly more monitoring infrastructure. In this paper, we describe a method to automatically differentiate and categorize requests without requiring sophisticated monitoring techniques. Using machine learning, our method requires only aggregate measures such as total number of requests and the total CPU and network demands, and does not assume prior knowledge of request categories or their individual resource demands. We explore the feasibility of our method on the .Net PetShop 4.0 benchmark application, and show that it works well while being lightweight, generic, and easily deployable.