Towards Performance Predictive Application-Dependent Workload Characterization

Workload characterization is important for users, designers, and those specifying future machine acquisitions. If the characterization method is carefully crafted to be comprehensive and consistent across platforms, it can be used to specify characteristics and components that comprise an optimal micro-architecture for the workload or application. Prior work has traditionally focused on two primary objectives: explaining application performance on a particular architecture through bottleneck identification and studying application similarity. This work defines an efficient characterization methodology that enables performance prediction in the context of architecture resources in addition to understanding application performance and similarity. We use four different and relatively new benchmark suites; two of which have not been characterized before. We apply this technique on two distinct micro-architectures to show that the characterization is consistent across platforms and can be used to accurately and optimally map applications to a machine in a testbed of available platforms.