ExanaDBT

Abstract

In this paper, we present a dynamic compilation system called ExanaDBT for transparently optimizing and parallelizing binaries at runtime based on the polyhedral model. Starting from hot spot detection of the execution, ExanaDBT dynamically estimates gains for optimization, translates the target region into highly optimized code, and switches the execution of original code to optimized one. To realize advanced loop-level optimizations beyond trace- or instruction-level, ExanaDBT uses a polyhedral optimizer and performs loop transformation for rewarding sustainable performance gain on systems with deeper memory hierarchy. Especially for successful optimizations, we reveal that a simple conversion from the original binaries to LLVM IR will not enough for representing the code in polyhedral model, and then investigate a feasible way to lift binaries to the IR capable of polyhedral optimizations. We implement a proof-of-concept design of ExanaDBT and evaluate it. From the evaluation results, we confirm that ExanaDBT realizes dynamic optimization in a fully automated fashion. The results also show that ExanaDBT can contribute to speeding up the execution in average 3.2 times from unoptimized serial code in single thread execution and 11.9 times in 16 thread parallel execution.