Ditto – Deterministic Execution Replay for the Java Virtual Machine on Multi-processors

Alongside the rise of multi-processor machines in the last decade, concurrent programming models have grown to near ubiquity. Programs built using these models are prone to bugs with rare pre-conditions, arising from unanticipated interactions between parallel tasks. Moreover, conventional debugging methodologies are not well suited to deal with non-deterministic faults, leading to inefficient debugging efforts in which most resources are consumed in reproduction attempts. Deterministic replay tackles this problem by recording faulty executions and using the traces to generate equivalent ones. Replayers can be efficient on uni-processor machines, but struggle with unreasonable overhead on multi-processors. We present Ditto, a deterministic replayer for concurrent JVM applications executed on multi-processor machines. By integrating state-of-the-art and novel techniques it manages to consistently out-perform previous deterministic replayers targeted at Java programs, in terms of recording overhead, replaying overhead and trace file size. The main contribution of Ditto is a novel pair of recording and replaying algorithms that (a) leverage the semantic differences between load and store memory accesses, (b) serialize memory accesses at the instance field level, (c) employ partial transitive reduction and program-order pruning on-the-fly, and (d) take advantage of TLO static analysis, escape analysis and JVM compiler optimizations to identify thread-local accesses. Palavras Chave Keywords