Towards Realistic Performance Bounds for Implicit CFD Codes

The performance of scientic computing applications often achieves a small fraction of peak performance [7,17]. In this paper, we discuss two causes of performance problems| insucient memory bandwidth and a suboptimal instruction mix|in the context of a complete, parallel, unstructured mesh implicit CFD code. These results show that the performance of our code and of similar implicit codes is limited by the memory bandwidth of RISC-based processor nodes to as little as 10 % of peak performance for some critical computational kernels. Limits on the number of basic operations that can be performed in a single clock cycle also limit the performance of \cache-friendly " parts of the code. 1