Depiction and playout of multi-threaded program executions

Abstract

Execution of a shared memory multi-threaded program is non-deterministic even for a fixed input. Consequently, a limited amount of the program behavior should be traced and recorded during run-time. However, if the tracing overheads are too high, we have the risk of slowing down the program considerably and even distorting the program behavior. In this paper, we propose to collect and store only the synchronization dependencies during run-time. These dependences are visualized as a message sequence chart (MSC). We do not record the data dependences across threads resulting from unsynchronized reads and writes of a shared variable. Instead all possible orderings of unsynchronized reads/writes are analyzed post-mortem. To describe all these behaviors, we use an important extension of message sequence charts called live sequence charts (LSC). Our MSC/LSC based description of a multi-threaded program execution can be simulated in an automated manner. This can help in understanding program behavior.