Impact of data transformations on memory bank locality

Abstract

High-energy consumption presents a problem for sustainable clock frequency and deliverable performance. In particular, memory energy consumption of array-intensive applications can be overwhelming due to poor cache locality. One option for reducing memory energy is to adopt a banked memory architecture, where memory space is divided into banks and each bank can be powered down if it is not in active use. An important issue here is the bank access pattern, which determines opportunities for saving energy. In this paper, we present a compiler-based data layout transformation strategy for increasing the effectiveness of a banked memory architecture. The idea is to transform the array layouts in memory in such a way that two loop iterations executed one after another access the data in the same bank as much as possible; the remaining banks can be placed into a low-power mode. Our simulation-based experiments with nine array-intensive applications show significant savings in memory energy consumption.