Optimized execution of deterministic blocks in Java PathFinder

Abstract. Java PathFinder (JPF) is an explicit-state model checker for Java programs. It explores all executions that a given program can have due to dierent thread interleavings and nondeterministic choices. JPF implements a backtracking Java Virtual Machine (JVM) that executes bytecodes using a special representation of JVM states. This special rep-resentation enables JPF to quickly store, restore, and compare states; it is crucial for making the overall state exploration ecient. However, this special representation creates overhead for each execution, even execu-tion of deterministic blocks that have no thread interleavings or nonde-terministic choices. We propose mixed execution, a technique that improves execution time of deterministic blocks in JPF. Our technique leverages the fact that JPF is written in Java: JPF is a special JVM that runs on top of a regular, host JVM. Mixed execution works by translating the state between the special JPF representation and the host JVM representation. We also present lazy translation, an optimization that speeds up mixed execution by translating only the parts of the state that an execution dynamically depends on. We evaluate mixed execution on six subject programs that use JPF for generating tests for data structures and on one case study for verifying a network protocol. The experimental results show that mixed execution can improve the overall state exploration time up to 36.98%, while improving the execution time of deterministic blocks up to 69.15%.