CS 575 Supercomputing for the Sciences
Ch. 14 Message-Passing Environments (focus on MPI)
November 26, 2003

Dr. Kris Stewart (stewart@rohan.sdsu.edu)
San Diego State University

MPI on Rohan [10Nov03]

This URL is http://www.stewart.cs.sdsu.edu/cs575/lecs/ch14.html

Quoting our text, p. 299, '...Message passing is the "assembly language" of parallel processing.'

PVM - Parallel Virtual Machine (we will not cover)

MPI - Message Passing Interface (our goal for using SDSC's Blue Horizon)

Our Focus is MPI

Fig. 14-1: A two-dimensional plate with four constant heat sources
© O'Reilly Publishers (Used with permission)

This provides a physical description of a common problem from science, the 2-dimensional heat flow problem, with given edge conditions (boundary is zero degrees) and four heat sources in the interior of the plate. Consider this 2D region and dividing it into separate grid elements, as we did earlier in this course to motive the heat diffusion problem. Also consider the case of using a machine such as Rohan, which has a total of 8 CPUs. We will never have dedicated access to this computer since it is the general purpose computer to serve the entire campus. We might consider how to partition our model to use 4 processors (therefore half of the computing power of our main campus machine).

Fig. 10-1: A Shared-memory multiprocessor

© O'Reilly Publishers (Used with permission)

NPACI Blue Horizon IBM SP2

Our goal in this class is to use the Blue Horizon which is a distributed memory processor with

• 1152 processors
• IBM 375MHz Power3 Nodes
• 4GB memory per 8-processor node
• AIX 5.1 L Operating System
• As evidenced by the 1.7 TFLOPs peak performance
• NPACI Resources

The online documentation from the Maui High Performance Computing Center are linked below. Today's lecture will go directly to some of the most important points in the following list. Please examine the full presentation for a better perspective. The instructor (Kris Stewart) favorite book on parallel programming is "Parallel Programming: Techniques and Applications" by Barry Wilkinson and Michael Allen, Prentice Hall, 1999.

What is Parallelism?

Parallel Programming

Shared Memory Parallel Programming Paradigm (Rohan)

Note: we see the bottleneck since we have no way to increase memory bandwidth.

Simple Heat Equation

Several illustrative figures

Data Paraellism to illustrate how MPI communicates

Especially useful are the digrams showing how the 2D spatial grid of the array of unknowns is distributed to four (4) separate nodes. We see how grid point that are interior can communicate with neighbors while grid points on the boundary must communicate with storage from another node.