IBM SP Parallel Scaling Overview

Conclusions

Implementing parallel scientific codes that scale well can be challenging. This document has tried to address some of the impediments to concurrency faced by researchers deploying HPC codes on NERSC hardware. The goal of such an endeavor is not scaling itself but rather the advancement of scientific research through computation. The choices in the level of parallelism that most effectively drives that advancement can be hopefully be expanded by such an examination.

If you have questions about parallel code scaling on seaborg not answered in this document, feel free to contact the author.

References

Useful Documents

• Scientific Applications in RS/6000 SP Environments
• RS/6000 Scientific and Technical Computing: POWER3 Introduction and Tuning Guide
• RS/6000 SP: Practical MPI Programming
• Understanding and Using the SP Switch
• NERSC IBM SP Memory Page
• NERSC MPI-I/O Page
• RS/6000 SP 375MHz POWER3 SMP High Node Overview , August 29, 2000
• On seaborg see, /usr/lpp/LoadL/READMES/LoadL.README and /usr/lpp/LoadL//READMES/poe.README

• triad.f
• mpimem.c
• p2p.c
• comm.c