ScaMPI – Design and Implementation

Abstract

MPI (Message Passing Interface) [7] is an established and de-facto standard for information exchange based on the message-passing paradigm. MPI was standardized by the MPI Forum in 1996. Academia, industry, and vendors of high-performance computers drove the effort.

The MPI has a rapidly growing community as a standard user API (application programming interface) for parallel programming. Today, MPI is the preferred API for portable, parallel programs, and the success of the standard can be illustrated by applications running on both shared and distributed memory systems. Examples of the former are systems from SGI and Sun, whereas IBM, Cray (now SGI), and Scali deliver systems adhering to the latter category. Applications written using MPI are deemed very portable, and they can easily be ported between shared and distributed memory systems. Thus, seen from an ISV (independent software vendor), a message passing application is more portable than an application using the shared memory paradigm, since the message passing application can run on either shared or distributed memory systems.

In the rest of the chapter the design and implementation of ScaMPI, Scali’s high performance MPI implementation is presented. Key technical achievements of ScaMPI are low latency, high bandwidth and flexibility of transport medium as well as options for speeding up application performance within SMPs by allowing the use of threads. The programming environment for ScaMPI provides various built-in options for debugging and tuning. In addition ScaMPI is integrated with powerful third party software. Performance of important ScaMPI primitives is discussed in light of recent performance measurements. These measurements also document excellent scalability of ScaMPI for up to 96 inexpensive dual CPU nodes.