Description: Derived datatypes let you describe non-contiguous or structured memory layouts so MPI can send them directly, which reduces data movement overhead. Virtual topologies let you create a processor mapping that gives structure to a communicator effectively reducing the algorithmic and network overhead of figuring out who talks to whom. Together they make MPI coded faster, cleaner, and more scalable.

The aims in this seminar are two-fold:

1. Discuss and demonstrate how to develop derived types to describe data layouts to automate the packing/unpacking of non-continuous data or nonuniform data types defined by data structures.

2. Discuss and demonstrate to develop virtual topologies that better align with an application’s data communication pattern to simplify and optimize the communication logic between processes.