Automatic Partitioning of Object-Oriented Programs with Multiple Distribution Objectives

There are many domains in which it would be very useful if we could take a monolithic program and automatically convert it into a distributed program according to some partitioning objective, such as minimizing total execution time or minimizing energy consumption on one particular distribution node. We present two novel techniques that work together to provide automatic multi-objective partitioning in the specific context of object-oriented programs. First, our method constructs an object relation graph (ORG) using a combination of static analysis and offline profiling information. Instead of directly partitioning this ORG, we then transform it into a target graph (TG) to abstract from concrete distribution infrastructures and objectives. By applying this two-layer graph modeling, we achieve a unified partitioning stragegy for different goals. We have built a framework that implements our techniques for Java, and explore its performance in the context of workload offloading for resource-constrained mobile devices. Our system takes standard Java bytecode programs as its input and automatically creates distributed Java bytecode programs as its output. Preliminary benchmark results are highly promising, with an average speedup factor of almost 1.5 and an average energy savings of 83.5% for the beneficial benchmarks.