Nonlinear Wave Equation Solver

Originally tuned for Cray-YMP

Loop intensive, but shorter average vector length

Three versions of the code:

• V1 - Most significant loops optimized for tasking and vectorization, but memory bank conflicts exist
• V2 - Bank conflicts eliminated
• V3 - Remaining serial loops optimized for tasking/streaming

Traditional development cycle would go through V1 and V2 to get highest single processor performance

Third step necessary to maximize performance on MSP

Previous slide

Next slide

Back to first slide

View graphic version

• -Now let us look at an application that was originally tuned for the Y-MP and see what additional tuning techniques will be needed to migrate to MSPs on the SV1

• -I will reference 3 version of the code that relate to the levels of optimization:

• 1) V1 - vectorization

• 2) V2 - minimize memory bank conflicts

• 3) V3 - eliminate serial dependencies