MPI Derived Datatypes and Data Communication Analysis in Meteorological Applications

摘要

Current trends in architecture are increasingly towards larger number of processing elements on a chip, and the speed of cores as a whole still improves as before. Communication bandwidth and storage capacity become the bottleneck of high performance computing currently, it's essential to explore the effects that the way data communication and storage bring in the era of MPI-2. Before that, users need to pack noncontiguous data manually before sending the data to another point, and unpack the data after receiving from others. Thus, we need to allocate a sufficient buffer to storage the packed data that is ready for sending, which introduces some additional communication overhead. MPI Derived datatypes (DDTs) were added as a novel feature to MPI-2 and provide a flexible mechanism for working with arbitrary noncontiguous layouts of data in memory. In this paper, we mainly focus on the datatypes and data communication pattern in meteorological applications. Based on the behavior of data communication of a meteorological application, we compare the performance of different communication routines, and investigate how the usage of derived data types affects performance in comparison with the manual packing of the send buffer in those routines. Most of the meteorological models were written in the era of MPI-1, and they were constrained to a few of datatypes. The work would also contribute to improving the communication process of those models.