Accelerating the parallel distributed execution of Java HPC applications

In order to speed up the execution of Java applications, JIT compilers compile method bytecodes into native code at run time. Once a method is compiled, subsequent executions of this method will run a generated machine-dependent code, which is generally faster than interpreting bytecodes. In addition, JIT compilers can apply several optimizations to the code to increase the performance even further. When parallelizing the execution of a Java application over a distributed infrastructure such a Grid, a JVM is created on each resource node to run a part of the application. The JIT compiler of each of these JVMs works independently from the others, thus paying the compilation cost and the time to reach the peak performance in every machine. Besides, in some cases, the execution nodes are not dedicated or not directly accessible; this prevents from maintaining a persistent JVM in each node during all the application execution, and gives little opportunity for the JIT compiler to perform mid-to-long term code optimizations. This document proposes a couple of techniques to fully benefit from JIT optimizations when distributing the execution of Java applications. These techniques could be integrated in COMP Superscalar, a framework that allows to run user-selected parts of a Java application in a set of distributed parallel resources.