Improving Continuation-Powered Method-Level Speculation for JVM Applications

Most applications running on the Java Virtual Machine (JVM) make extensive use of dynamic object-oriented programming features such as inheritance, polymorphism, and encapsulation. This makes them very hard or even impossible to analyze statically, defeating most of the automatic parallelization research done so far for traditional computeheavy scientific applications. In this paper, we propose and evaluate multiple extensions to the JaSPExMLS framework, a speculative parallelization framework that is aimed at irregular applications. This framework works atop a modified JVM and employs Method-Level Speculation (MLS), a task-identification technique that is better suited for irregular applications. Our custom JVM is a modified version of the OpenJDK Hotspot VM that was extended with support for first-class continuations, while still inheriting Hotspot’s high-performance features such as just-in-time compilation, adaptive optimization, state-of-the-art garbage collection, and support for the latest Java versions. JaSPEx-MLS automatically modifies applications to use Software Transactional Memory (STM) and to allow the spawn and synchronization of speculative tasks in a scheme similar to Fork/Join parallelism. Speculative execution is supported by our novel relaxed STM model, which is tightly coupled with our framework and includes support for integrating with Futures. We present novel techniques for improving MLS runtime task extraction and coordination, describe our implementation of those techniques onto JaSPEx-MLS, and present experimental results showing their impact on both reducing speculative execution overheads and extracting further parallelism from sequential applications.