Mats: MultiCore Adaptive Trace Selection

Dynamically optimizing programs is worthwhile only if the overhead created by the dynamic optimizer is less than the benefit gained from the optimization. Program trace se-lection is one of the most important, yet time consuming, components of many dynamic optimizers. The dynamic ap-plication of monitoring and profiling can often result in an execution slowdown rather than speedup. Achieving sig-nificant performance gain from dynamic optimization has proven to be quite challenging. However, current technolog-ical advances, namely multicore architectures, enable us to design new approaches to meet this challenge. Selecting traces in current dynamic optimizers is typically achieved through the use of instrumentation to collect con-trol flow information from a running application. Using in-strumentation for runtime analysis requires the trace selec-tion algorithms to be light weight, and this limits how so-phisticated these algorithms can be. This is problematic because the quality of the traces can determine the poten-tial benefits that can be gained from optimizing the traces. In many cases, even when using a lightweight approach, the overhead incurred is more than the benefit of the optimiza-tions. In this paper we exploit the multicore architecture to design an aggressive trace selection approach that produces better traces and does not perturb the running application. 1