Constructing Resilient Region in Dynamic Optimization Systems via Dynamic Adjustment of Bias Thresholds

Abstract

Performance of dynamic optimization system is strongly affected by the region it selects to optimize. Larger region has more opportunities to exploit the potential of optimization techniques, but it also has the risk of flushing and re-execution of correct-path instructions caused by the internal mis-speculations. To address this problem, we propose a method for constructing resilient regions by constructing them across multiple branches only when the branches exhibit a biased behavior. To minimize the re-execution of correct-path instructions as well as to provide meaningful performance gain, we also propose to monitor the behavior of dynamic branches and adaptively adjust the merging threshold with respect to application phase changes. We first analyze the divergence pattern of dynamic branches, and then describe proposed region construction techniques (static and dynamic methods). Our evaluation results show that the proposed design with dynamic adjustment scheme shows the best coverage (over 85%) with negligible performance degradation.

Ipoom Jeong

Assistant Professor

My research interests include CPU/GPU microarchitectures, memory/storage system designs, and smart-I/O devices