Characterizing the Effect of Microarchitecture Design Parameters on Workload Dynamic Behavior

Program runtime behavior exhibits significant variations across multiple scales. The increasing design complexity and technology scaling make microprocessor performance and efficiency increasingly depend on runtime workload dynamics. Therefore understanding the effect of design parameters on workload dynamics at early, microarchitecture exploration stage is crucial for high-performance and complexity-efficient designs. In this study, we apply wavelet-based analysis to decompose workload dynamics into a series of wavelet coefficients, which represent program behavior ranging from low-resolution approximation to high-resolution detail. We then construct error-bounded linear regression models that relate microarchitecture design parameters to various wavelet coefficients that capture workload dynamics at multiresolution levels. The most significant factors affecting program dynamics at different scales are obtained. To our knowledge, this paper presents the first work on microarchitecture design space exploration focusing on workload dynamics.