Design and Implementation of the Lightweight Corefile Browser

One of the main problems in debugging parallel programs is in determining how far a parallel program has executed when it crashes or stops. Even if the program dumps core files for all the parallel processes, the information is at such a low level that is difficult to interpret correctly. Complete parallel core files also occupy a significant amount of disk space. We present the Lightweight Corefile Browser (LCB) tool, a Parallel Tools Consortium project. LCB offers a simple, convenient way of determining the source location and call stack of each processor when a parallel program terminates prematurely. To do this, a minimal amount of locational data is recorded for each processor and input to the LCB analyzer. The command line version of the tool identifies the culprit (i.e., faulting) process, the reason for failure, and the calling sequence that led to failure. The graphical version of the tool presents different views of the program''s dynamic call structure. The user can navigate through these views to explore the global state of program control at the moment of failure.