Memory system support for image processing

Processor speeds are increasing rapidly, but memory speeds are not keeping pace. Image processing is an important application domain that is particularly impacted by this growing performance gap. Image processing algorithms tend to have poor memory locality because they access their data in a non-sequential fashion and reuse that data infrequently. As a result, they often exhibit poor cache and TLB hit rates on conventional memory systems, which limits overall performance. Most current approaches to addressing the memory bottleneck focus on modifying cache organizations or introducing processor-based prefetching. The Impulse memory system takes a different approach: allowing application software to control how, when, and where data are loaded into a conventional processor cache. Impulse does this by letting software configure how the memory controller interprets the physical addresses exported by a processor. Introducing an extra level of address translation in the memory. Data that is sparse in memory can be accessed densely, which improves both cache and TLB utilization, and Impulse hides memory latency by prefectching data within the memory controller. We describe how Impulse improves the performance of three image processing algorithms: an Impulse memory system yields speedups of 40% to 226% over an otherwise identical machine with a conventional memory system.