Memory System Support for Dynamic Cache Line Assembly

The effectiveness of cache-based memory hierarchies depends on the presence of spatial and temporal locality in applications. Memory accesses of many important applications have predictable behavior but poor locality. As a result, the performance of these applications suffers from the increasing gap between processor and memory performance. In this paper, we describe a novel mechanism provided by the Impulse memory controller called Dynamic Cache Line Assembly that can be used by applications to improve memory performance. This mechanism allows applications to gather on-the-fly data spread through memory into contiguous cache lines, which creates spatial data locality where none exists naturally. We have used dynamic cache line assembly to optimize a random access loop and an implementation of Fast Fourier Transform (FFTW). Detailed simulation results show that the use of dynamic cache line assembly improves the performance of these benchmarks by up to a factor of 3.2 and 1.4, respectively.