Implementing Fast and Reusable Datatype Processing

Methods for describing structured data are a key aid in application development. The MPI standard defines a system for creating “MPI types” at run time and using these types when passing messages, performing RMA operations, and accessing data in files. Similar capabilities are available in other middleware. Unfortunately many implementations perform poorly when processing these structured data types. This situation leads application developers to avoid these components entirely, instead performing any necessary data processing by hand. In this paper we describe an internal representation of types and a system for processing this representation that helps maintain the highest possible performance during processing. The performance of this system, used in the MPICH2 implementation, is compared to well-written manual processing routines and other available MPI implementations. We show that performance for most tested types is comparable to manual processing. We identify additional opportunities for optimization and other software where this implementation can be leveraged.