Temporal Streaming of Shared Memory

Abstract

Abstract

An analysis of address and branch patterns with PatternFinder

“…To reduce the simulation time, we eliminate overlapping occurrences of the pattern within a node during the construction of the suffix tree for a particular iteration. For example, in Figure 1, at Level 2 (i.e., the pattern length of 2), for pattern aa (3,6,7), the second and third occurrences of it overlap, therefore, its third occurrence is eliminated to update the occurrences list to be nonoverlapping as aa (3,6). This eliminates significant number of patterns from the suffix tree, which in turn improves processing times.…”

“…For example, CLIENT02 and SERVER01 have very long non-overlapping patterns with very good coverage, 80% and 35%, respectively. A mechanism similar to memory streaming proposals [3,4,[31][32][33][34][35] can be used for also branch outcome streaming. 0 100 200 300 400 500 600 700 800 900 1000 INT01 INT02 INT03 INT04 INT05 INT06 CLIENT01 CLIENT02 CLIENT03 CLIENT04 CLIENT05 CLIENT06 intervals (WARI) (columns 3 and 4), where the coverage of a pattern is used as its weight, so hotter patterns have a greater influence on the reported average value.…”

“…This was the first study to quantify the observation that extended memory access sequences recur using a hierarchical compression algorithm, called SEQUITUR [2]. Several researchers then proposed ways to exploit this behavior [3,4]. Surprisingly, after a decade, their analysis has remained one of the most detailed for quantifying hot memory streams.…”

Semantics-Aware, Timely Prefetching of Linked Data Structure

“…Many hardware based data prefetching methods either regularity-based [30,22,14,32,3,10,13] or correlationbased [29,28,34,4,12,2] have been given in Section I. The Global History Buffer [23] proposes a general FIFO structure for recording and identifying nearby missing address patterns.…”

“…This approach, however, cannot handle irregular missing address patterns. Correlation-based prefetchers, such as correlated [2], Markov [12], hot-stream [4], temporal streaming [35,34], spatialstreaming [28] and a hybrid temporal/spatial streaming [29] prefetchers, on the other hand, record the history of nearby missing addresses to trigger prefetches assuming such miss correlations will be repeated. To be effective, the correlated prefetchers must record a long cache miss history and incur significant storage overhead.…”