Εκμετάλλευση Των Multi-core Επεξεργαστών Για Αριθμητικούς Κώδικες Περιορισμένους Από Μνήμη Με Χρήση Τεχνικών Prefetching

Abstract

Numerical codes occur in almost all science and engineering disciplines; e.g., computational fluid dynamics, take effect on applications such as aeronautics, automotive, power generation, chemical and petrochemical process. They consist of small kernels, each carrying out a large number of Floating Point operations on proportionally big data sets. Moreover, such codes introduce significant portion of iterations. If we suppose that the data sets resided entirely in the main memory and every time that data were needed for computation were fetched from the main memory, then numerical codes would be confounded to very poor execution times, because of the existing barrier for quick memory accesses. Happily, memory hierarchy and optimizing techniques that can be applied to numerical codes as well have existed since a long time, bringing the execution times of numerical codes up to satisfactory levels.From the existing optimization techniques, the technique of data prefetch could be proved as particularly beneficial when speaking of numerical codes that act on large data portions, especially if a multi-core processor would be utilized as a prefetch engine. The potential of improvement of the execution times from the utilization of this technique and the further investigation of such a possible gain comprise the subject of this Thesis.