Using A Multiscale Approach to Characterize Workload Dynamics

Workload characterization has been essential ingredients for the design of computer systems. As computer architecture becomes adaptive and reconfigurable, its efficiency increasingly depends on the dynamic behavior of workloads. Program execution manifests wildly varied changes at run time. As software becomes large and sophisticated, such variation can span a wide range of time scales. As current workload characterization methodologies have paid less attention to the scaling behavior of program, we are often at a loss as to how to interpret and forecast the changing of workload dynamics over time. To address this issue, we advocate a multiscale workload characterization methodology that can capture the dynamic nature of workloads across different time scales. We apply this technique to study the scaling properties of the SPEC2000 integer benchmarks. The on-line program scaling estimator proposed in this paper allows one to capture both short-term and long-term execution characteristics of large program in-flight.