A Hybrid Adaptive Feedback Based Prefetcher
Santhosh Verma, Lu Peng · 2010
We present a hybrid adaptive prefetching scheme in this paper. The scheme consists of a hybrid stride/sequential prefetching mechanism which adapts its aggressiveness based on feedback metrics collected dynamically during program execution. Metrics such as prefetching accuracy and prefetching lateness are used to adjust aggressiveness in terms of prefetch distance (how far ahead of the current miss it fetches) and prefetch degree (the number of prefetches issued). The scheme is applied separately at both the L1 and L2 cache levels. For a set of 15 SPEC CPU 2006 benchmarks, our scheme achieves a geometric mean CPI improvement of 29 % compared to no prefetching and 6 % compared to a stride prefetcher. The performance improvement over a set of memory intensive benchmarks is 45 % over no prefetching and 9 % over stride. 1.