Defining relevant distances between server workloads

Abstract

Computer system design studies traditionally involve only small collections of benchmarks. Detailed benchmark analysis is extremely time consuming and requires a large amount of human and machine resources. Therefore, it is essential that the benchmark collection be representative of the customer workloads for which an architecture is developed. In recent work, interworkload distances have been proposed as a way of characterizing workload similarity. These distances are based on measurable/computable program characteristics, such as instruction mix or dependence distance. In the literature, these characteristics enter the distances symmetrically. We observe that the program behavior impact of different characteristics varies significantly. We propose a method of estimating the program behavior impact via a regression model. Its components then enter the distance definition directly, thus emphasizing high-impact characteristics. We also propose a data collection methodology that can be deployed at a customer site without requiring code instrumentation and/or a detailed simulation setup. We build a dataset consisting of 84 program characteristics for each of the 106 workloads and apply the proposed distance methodology to it.