Computing-Kernels Performance Prediction Using Dataflow Analysis and Microbenchmarking
Résumé
On modern multi-core processors, the growing gap between memory size, bandwidth and latency compared to computing capability makes the memory hierarchy predominant for performance. The Micro- kernel-Description-Language based Performance Evaluation Framework, MDL-PEF, accurately predicts optimized inner-loops performance de- pending on the loop's data access. The MDL-PEF approach revolves around a data flow description language, MDL. A static analysis step extracts the data flow structures of the assembly code. Then the pre- dictor uses pattern matching against an MDL-Microkernel database for predicting performance. Finally, MDL-PEF provides an automatic tool to initialize a pattern matching database for the target architecture. The overall system can predict the kernel performance on different plat- forms and optimizations, helping the user choose the best architecture for a given kernel. Preliminary experiments, with a 56 elements database, predict the innermost loop throughput of 636 binary loops of the NAS benchmarks with an average 10% of relative error. The performance pre- dictor is part of the Modular Assembly Quality Analyzer and Optimizer (MAQAO) performance tool framework. Future works will extend MDL- PEF to other architecture paradigms and more complex control flows such as outer loops.
Origine : Fichiers produits par l'(les) auteur(s)
Loading...