Fast IPC estimation for performance projections using proxy suites and decision trees

摘要

Accurate IPC estimates are critical for generating performance projections of key workloads on future designs. However, the need to respond to projections requests in a timely manner in the face of rapidly evolving applications and software stacks and tight schedule constraints, often preclude design teams from executing detailed workload analysis, sampling and simulation flows for such purposes. We address this problem by taking advantage of the large amount of data that performance modeling teams commonly generate as part of architectural studies across thousands of workload scenarios. We propose two methods for exploiting these datasets: one that builds proxy suites, and another that builds decision-tree based classifiers. Both methods can generate IPC estimates for a target workload without collecting new workload samples, or running a single additional simulation. We discuss our experience using these techniques to estimate the IPC of numerous commercial workloads on four industrial x86 processor designs. The resulting IPC estimates were on average, within 2% of those obtained via measurements or detailed cycle-accurate simulations Importantly, using these methods, we were able to generate IPC estimates for a target workload in a matter of hours to 1-2 days, compared to several weeks using conventional approaches.