Debugging distributed shared memory applications

A debugger is a crucial part of any programming system, and is especially crucial for those supporting a parallel programming paradigm, like OpenMP. A parallel, relaxed-consistency, distributed shared memory (DSM) system presents unique challenges to a debugger for several reasons: 1) the local copies of a given variable are not always consistent between distributed machines, so directly accessing the variable in the local memory by the debugger won't always work as expected; 2) if the DSM and debugger both modify page protections, they will likely interfere with each other; and 3) since a large number of SIGSEGVs occur as part of the normal operation of a DSM program, a program error producing a SIGSEGV may be missed. In this paper, we discuss these problems and propose solutions.