Intermediate MPI Programming

This section covers some more advanced aspects of MPI including non-blocking communcations (universally used to avoid deadlock), collective communication patterns, virtual topologies and derived datatypes. The relevant exercises are numbers 5, 6, 7 and 8; exercise 9 is entirely optional, but hopefully fun to attempt. This block also includes simple example solutions to the basic exercises.

Learning Objectives

After completing this section the student will:

gain familiarity with intermediate aspects of MPI such as non-blocking and collective communication, virtual topologies and derived datatypes
have gained practical experience by solving various exercises

Non-Blocking Communication

This subsection covers non-blocking communication functions (e.g. MPI_Isend and MPI_Irecv) that can be used to avoid deadlock (slides).

Collectives

This section covers collective communication patterns such as MPI_Barrier, MPI_Bcast, MPI_Scatter, MPI_Gather, MPI_Scan, MPI_Reduce and MPI_Allreduce (slides).

Virtual Topologies

This subsection covers creation and use of virtual topologies in MPI to simplify structured communication (slides).

Derived Datatypes

This subsection covers derive datatypes, which facilitate communication of composite data structures and non-contiguous data (slides).

Simple example solutions using MPI in C and Fortran can be downloaded here.