MPL*: Efficient record/replay of nondeterministic features of message passing libraries

Abstract

A major source of problems when debugging message passing programs is the nondeterministic behavior of the promiscuous receive and nonblocking test operations. This prohibits the use of cyclic debugging techniques because the intrusion caused by a debugger is often large enough to change the order in which processes interact. This paper describes the solutions we propose to efficiently record and replay the nondeterministic features of message passing libraries (MPL) like MPI or PVM. It turns out that for promiscuous receive operations it is suf-ficient to keep track of the sender of the message, and for nonblocking test-operations to keep track of the number of failed tests. The proposed solutions have been implemented for an existing MPI-library, and performance measurements reveal that the time overhead of both record and replay executions is very low with respect to the (nondeterministic) original execution while the size of the log files remains very small.