Dead-block prediction & dead-block correlating prefetchers

Abstract

Effective data prefetching requires accurate mechanisms to predict both "which" cache blocks to prefetch and "when" to prefetch

Last-Touch Correlated Data Streaming

“…Figure 2 shows the cumulative distribution of L1 block dead-times-the number of cycles between a last touch to a block until its eventual eviction. 1 The figure corroborates prior results [12,13,26], showing that over 85% of all cache-block dead-times are longer than the memory access latency. Therefore, prefetch requests issued at last touch can complete before the next access to the same cache index, eliminating the entire off-chip miss latency.…”

“…Lai et al proposed Dead-Block Correlated Prefetching [12] to predict the last touch to a cache block prior to that block's eviction, and replace the block by prefetching a subsequently accessed block. By triggering the prefetch at the last touch, DBCP provides maximal lookahead among address-correlating prefetchers.…”

“…In contrast, the delta-correlating GHB prefetcher achieves only 31% performance improvement. DBCP with 2MB on-chip storage [12] achieves only 17% performance improvement. Finally, increasing L2 cache size by a factor of four achieves only 16% performance improvement.…”

“…A recent proposal, the Dead-Block Correlating Prefetcher (DBCP) [12], achieves maximal prefetch lookahead through correlation to "last touch" accesses-the last access to each cache block prior to eviction. However, capacity constraints of DBCP's on-chip correlation table limit its coverage.…”

Fast thread migration via cache working set prediction

“…Dead-block prediction [20] predicts when L1 D-cache blocks are no longer needed before replacement; an additional table is used to predict likely successors for earlyevicted blocks. They use the dead-block predictor and correlations it exposes to prefetch likely misses, and use the freed L1 storage as a prefetch buffer.…”