Going the distance for TLB prefetching: an application-driven study

Abstract

Abstract

CoLT: Coalesced Large-Reach TLBs

“…In response, researchers have considered techniques to improve TLB structure, lookup, and placement [9,13]. More sophisticated techniques such as TLB prefetching and mechanisms to accelerate page walks have also been considered [5,11,19,27]. Our goal is to propose techniques, orthogonal to past work, to further boost TLB performance.…”

“…In the first regime, the OS does not succeed in mapping contiguous physical pages to contiguous virtual pages. In these cases, only traditional techniques such as changing TLB organizations or prefetching are likely to be successful in increasing hit rates [6,11,19]. At the other end of the spectrum, there exists a regime of extremely high contiguity, when the OS decides that the overheads of superpage construction are well worth the effort.…”

“…Translation Lookaside Buffers (TLBs) are crucial to system performance due to their long miss penalties [5,11,14,19,22,26]. Past work has shown that TLB misses degrade performance by 5% to 14% for even nominally-sized applications.…”

Data Cache Prefetching Using a Global History Buffer

“…When a cache miss occurs, the miss address indexes the correlation Distance prefetching is a generalization of Markov prefetching. 5 Originally proposed for prefetching TLB entries, distance prefetching can easily be adapted to prefetching cache lines. This adaptation uses an address delta, the distance between two consecutive miss addresses, as the prefetch key into the correlation table.…”

“…[3][4][5][6] Figure 1a illustrates conventional We propose an alternative structure, shown in Figure 1b, for holding prefetch history. In this structure, a fixed-length FIFO table, the global history buffer (GHB) holds cache miss addresses.…”