Fractal Matrix Multiplication: A Case Study on Portability of Cache Performance
WAE '01 Proceedings of the 5th International Workshop on Algorithm Engineering
JuliusC: a practical approach for the analysis of divide-and-conquer algorithms
LCPC'04 Proceedings of the 17th international conference on Languages and Compilers for High Performance Computing
| Hi-index | 0.00 |
Data Locality exploitation and Performance Evaluation are important issues in compiler and algorithms so that modern architectures can be fully utilized. Multiprocessor platforms, with RISC processor and multiple level of caches, are becoming very common platforms. In spite their computational capability, they are not easy to be modeled and therefore application performance is not easy to be estimated. Indeed, it is often the case that developers code applications that can obtain very different performance on different platforms. This almost magic but unpleasant behavior is due to many factors. One of them is Data Locality. A smart reutilization of data, stored at the different levels of the memory hierarchy, may improves performance. In this investigation we are looking for a machine independent definition of data locality, inherent to the application and a general technique to exploit it automatically at source code level: reorganizing the schedule of instructions or, sometime, rewriting the algorithm. Indeed, we are proposing a machine independent method to estimate data locality, Access Complexity, we are proposing an heuristic to schedule instructions and manage memory, then we apply our ideas on a case study, matrix matrix multiplication.