mpi

MPI section

Learning objectives

At the end of the MPI section, the student should be able to

• Explain differences in communicating between processes/threads in a shared memory system vs a distributed memory system
• Describe deadlocking communication patterns and approaches to avoid deadlocks
• Contrast blocking and non-blocking communication
• Write MPI programs in C, C++, or Fortran for:
  • Communicating data between processes
  • Using collective communication calls over a subset of processes
• Compile and run MPI programs in supercomputers
• Start exploring some advanced MPI features relevant for their use case

Demo codes

See demos for demo codes referred to in the slides.

Wednesday exercises

Introduction to MPI

• Hello world

Point-to-point communication

• Message exchange
• Parallel pi with two processes

MPI programming practices

• Message chain
• Heat equation solver: Tasks 1-2
• (Bonus) Parallel pi with any number of processes
• (Bonus) Broadcast and scatter

Collective operations

• Collective operations

Thursday exercises

Debugging

• Debugging

Collective reductions

• Heat equation solver: Task 3

Non-blocking communication

• Non-blocking message chain
• Heat equation solver: Task 4

User-defined communicators

• Communicators and collectives

Further MPI topics

• Cartesian topology
  • Cartesian grid process topology
  • Message chain with Cartesian communicator
  • Heat equation solver: Task 5
• User-defined datatypes
  • User-defined datatypes
  • Modifying extent
  • Communicating struct
• Persistent communication
  • Message chain with persistent communication

Bonus

• Heat equation solver: Remaining tasks