The following is a list of MPI features categorized by which Cornell Virtual Workshop roadmap covers the concepts. If you complete all five CVW MPI roadmaps, then all of the features listed will be covered in detail. For now, this list will give you some idea of what MPI was designed to do.

Introduction to MPI (this roadmap)

Deterministic assignment of tasks to machines

Specify where each code section will run, in order to efficiently execute jobs that do equal work (you can do the load balancing)

Simple, unique rank for each process

Identify output from each of many instances with unique file names

Separate name and address spaces

Work with distributed data without having to change variable names

Guaranteed in-order delivery of messages

Safely assume format and content of received messages

High performance

Establish communication patterns at the outset -- sockets are set at MPI_INIT, so connections are already established

MPI Point-to-Point Communication

Blocking communication

Halt a program's execution until a message has been sent or received (Ssend)

Non-blocking communication

Perform other, unrelated work after a message has been queued (Isend)

MPI Collective Communication

Collective communication functions

Make use of Broadcast, Scatter/gather, Scan/Reduction, All-to-all, Barrier

MPI One-Sided Communications

One-Sided Calls

Allow a single MPI process to initiate sending or receiving in another process, potentially reducing synchronization overhead.

Remote Direct Memory Access

Use RMA to directly access the memory of other MPI processes through low-latency communication paths like Intel Omni-Path or InfiniBand.

MPI Advanced Topics

Virtual topologies

Define communicators on a subset of processes, or define private communicators for use by libraries

Standard interface to utility functions

Assure portability through a standard interface to, e.g., timer, hostname

Error handling

Provide error-handling functions for communication failures

Persistent communications

Pre-set, then use and reuse communication paths, all implicitly non-blocking