This document provides a technical description of the IBM SP seaborg.nersc.gov with an emphasis on how the system performs as a production environment for large scale scientific applications. The overall goal is to provide developers with the information they need to write and run such applications. While some of the information presented here may be applicable in a larger context, the focus is on experiences and scaling techniques specific to seaborg.nersc.gov.

In the first few sections we seek to determine how well the theoretical capabilities of this machine may be realized in practice. Most of the measured performance numbers come from small code microkernels or test codes rather than from real user codes.

The microkernel approach described above has value to the user since the most widely quoted performance numbers often reflect the theoretical (peak) performance values rather than those realized in practice. Likewise anecdotes from full blown applications often involve mixed algorithms, hidden constraints or other specifics which make the results hard to generalize. Microkernels and test codes provide a middle ground which is a reasonable best case scenario for real user codes and provide the user with a small kernel of code can serve as a template for the writing or modification of more substantial codes.

In what follows we will present several such examples alongside performance and scaling information. The next section introduces the role of concurrency in HPC. The following sections treat parallel scaling from the machine and application development perspectives.

Suggestions for improving this draft document are welcome. Please contact the author if you have comments, corrections, or questions.