The International Conference for High Performance Computing, Networking, Storage, and Analysis
The International Conference for High Performance Computing, Networking, Storage, and Analysis
Workshop: 7th International Workshop on Software Correctness for HPC Applications (Correctness '23)
Authors: Simon Schwitanski, Niko Sakic, Joachim Jenke, Felix Tomski, and Marc-André Hermanns (IT Center, RWTH Aachen University)
Abstract: Partitioned communication introduced with MPI 4.0 can improve the communication efficiency of hybrid parallel models. It allows threads on the sender and the receiver side to work on parts of a communication buffer before the communication operation is fully completed. In this presentation, we discuss which kind of erroneous usage patterns are possible with MPI partitioned communication and provide a set of example test cases to be integrated in future classification quality benchmark suites. Further, we explain how we implemented first basic checks specific to partitioned communication, namely argument errors and erroneous partition activation, in our correctness checking tool MUST. The evaluation on the example test cases shows that MUST can correctly detect the errors and can help users to pinpoint bugs in their application.