Myth or Reality: Binary-Level Dynamic Optimization
 

Shukang Zhou        Bruce R. Childers        Mary Lou Soffa

In submission
If you are interested in getting a copy of the paper, please send Sukang an email.

Abstract

Dynamic optimization is often considered a promising complement to static optimization because it can utilize run-time information. Despite the promise, current implementations of dynamic optimizers that operate at the binary-level typically provide minimal benefit. Thus, the question becomes how much of the claimed benefit due to dynamic optimization is just an unfullfilled promise? To answer this question, in this work, we performed a limit study to discover the potential performance improvement from dynamic optimization. This limit study identifies important factors for dynamic optimization (trace quality, memory disambiguation, overhead, and hardware effects) and makes them ideal to understand their influence. Our results show that binary-level dynamic optimization does indeed have promise under ideal circumstances with a significant speedup of up to 37.3%. To understand why current implementations fail to reach this potential, we experimentally investigated the profitability of individual factors with realistic optimization techniques. Our results demonstrate that if these factors are improved, each one may help improve program speedup by up to 42.9% over current implementations. From our study, researchers are better positioned to identify what is important for dynamic optimization and to develop new, more powerful techniques.

Shukang Zhou, 07/14/2008