Runtime Checking of MPI Applications with MARMOT

The Message Passing Interface (MPI) is widely used to write parallel programs using message passing, but it does not guarantee portability between different MPI implementations. When an application runs without any problems on one platform but crashes or gives wrong results on another platform, developers tend to blame the compiler/architecture/MPI implementation. In many cases the problem is a subtle programming error in the application undetected on the platforms used previously. Finding this bug can be a very strenuous and difficult task. This paper presents MARMOT, an automated tool designed to check the correctness of MPI applications during runtime. Examples of such violations are the introduction of irreproducibility, deadlocks, incorrect management of resources such as communicators, groups, datatypes etc. or the use of non-portable constructs.