Book Description

The exponentially increasing gap between processors and off-chip memory, as measured in processor cycles, is rapidly turning memory latency into a major processor performance bottleneck. Traditional solutions, such as employing multiple levels of caches, are expensive and do not work well with some applications. We evaluate a technique, called runahead pre-processing, that can significantly improve processor performance. The instruction and data stream prefetches generated during runahead episodes led to a significant performance improvement for all of the benchmarks we examined. We found that runahead typically led to about a 30% reduction in CPI for the four Spec95 integer benchmarks that we simulated, while runahead was able to reduce CPI by 77% for the STREAM benchmark. This is for a five stage pipeline with two levels of split instruction and data caches: 8KB each of L1, and 1MB each of L2. A significant result is that when the latency to off-chip memory increases, or if the caching performance for a particular benchmark is poor, runahead is especially effective as the processor has more opportunities in which to pre-process instructions. Finally, runahead appears particularly well suited for use with high clock-rate in-order processors that employ relatively inexpensive memory hierarchies.