Entropy-based profile characterization and classification for automatic profile management

The recent adoption of pre-JIT compilation for the JVM and .NET platforms allows the exploitation of continuous profile collection and management at user sites. To support efficient pre-JIT type of compilation, this paper proposes and studies an entropy-based profile characterization and classification method. This paper first shows that highly accurate profiles can be obtained by merging a number of profiles collected over repeated executions with relatively low sampling frequency for the SPEC CPU2000 benchmarks. It also shows that simple characterization of the profile with information entropy can be used to guide sampling frequency of the profiler in an autonomous fashion. On the SPECjbb2000 benchmark, our adaptive profiler obtains a very accurate profile (94.5% similar to the baseline profile) with only 8.7% of the samples that would normally be collected using a 1M instructions sampling interval. Furthermore, we show that entropy could also be used for classifying different program behaviors based on different input sets.