Data Local Iterative Methods For The Efficient Solution of Partial Differential Equations

A cooperation
between

and
.

Funded by
.

h048z: Preliminary performance results on a Hitatchi SR8000

Overview:

In order to mitigate the effect of the growing gap between CPU speed and main memory performance current computer architectures use hierarchical memory designs comprising one or even more levels of cache memory. This is especially true for the Hitachi SR8000-F1, which is based on a Power3-like processor using a 128 kByte on-chip L1 cache.

The general goal of our project h048z is to investigate how numerical algorithms can be designed and/or restructured such that the hierarchical memory structure of the underlying architecture is respected. In particular we focus on the processing of algorithms on regular meshes.

Our research efforts concerning code performance optimizations on the Hitachi architecture yield techniques which are immediately applied by the KONWIHR projects gridlib as well as ParEXPDE.

Summary of results:

• Performance results for blocked Gauss-Seidel code on regular grids
• Performance results for hierarchical hybrid grids
• Porting the DiMEPACK library

A detailed description of the results in Postscript format can be found here.
For more information on the DiME project you may visit our Project homepage