IMPROVING MULTIBANK MEMORY ACCESS PARALLELISM ON SIMT ARCHITECTURES

Memory mapping has traditionally been an important optimization problem for high-performance parallel systems. Today, these issues are increasingly affecting a much wider range of platforms. Several techniques have been presented to solve bank conflicts and reduce memory access latency but none of them turns out to be generally applicable to different application contexts. One of the ambitious goals of this Thesis is to contribute to modelling the problem of the memory mapping in order to find an approach that generalizes on existing conflict-avoiding techniques, supporting a systematic exploration of feasible mapping schemes.