Characterizing Inter-Execution and Inter-Application Optimization Persistence

Dynamic translation and optimization systems employ code caches to improve performance and to support reuse of dynamically generated code sequences within a single run of an application. However, these intra-application caching techniques are ineffective at amortizing runtime costs on shortrunning applications or scripts that repeatedly invoke the same application. For these situations, dynamic translation systems have successfully used persistent code caching—saving the code cache during one run to prime later executions—as a means for amortizing runtime translation costs. The success of persistent code caching in dynamic optimizers will depend heavily on the amount of interexecution and inter-application optimization persistence that can be found in software applications. Our experiments with the DynamoRIO dynamic optimization system demonstrate that many of the most heavily executed code traces in SPECCPU2000 are identically optimized during successive executions. Our results indicate that there is a significant opportunity for leveraging interexecution persistence, and there is even a small opportunity for inter-application persistence for small programs.