Optimal Cache Partition-Sharing

Abstract

When a cache is shared by multiple cores, its space may be allocated either by sharing, partitioning, or both. We call the last case partition-sharing. This paper studies partition-sharing as a general solution, and presents a theory an technique for optimizing partition-sharing. We present a theory and a technique to optimize partition sharing. The theory shows that the problem of partition-sharing is reducible to the problem of partitioning. The technique uses dynamic programming to optimize partitioning for overall miss ratio, and for two different kinds of fairness.Finally, the paper evaluates the effect of optimal cache sharing and compares it with conventional solutions for thousands of 4-program co-run groups, with nearly 180 million different ways to share the cache by each co-run group. Optimal partition-sharing is on average 26% better than freefor-all sharing, and 98% better than equal partitioning. We also demonstrate the trade-off between optimal partitioning and fair partitioning.

Cache Exclusivity and Sharing

“…The access rate was used earlier in StatStack developed by Eklov et al [16]. Brock et al used the term stretched footprint [4] and called co-run logical clock common logical time [53]. A similar use was also given by Hu et al [24] in computing the AET in shared cache.…”

“…The sharing of L1 is solved by HOTL (Section 2). In particular, we use the recent result by Brock et al [4], who showed that HOTL implies the existence of a cache partition, called the natural cache partition (NCP), whose performance equals to that of cache sharing. The natural partition of a program is its effective occupancy in shared cache.…”

“…Consider the two-program group G = {a, b} that share L1 and L2, where L2 is the victim cache of L1. According to Brock et al [4], the effect of L1 sharing is equivalent to a partition of L1 into two parts of size…”

Rethinking Memory Management in Modern Operating System: Horizontal, Vertical or Random?

“…This result indicates that the performance of Curve-VP relative to the cache partitioning in these cases (3,4 or more LLCT applications) is limited due to the interference among LLCT applications and the resource constraint (more LLCT applications are mapped into Bank Group A in Figure 11). Although the performance gains on the cache partitioning and Curve-VP are similar with high number of LLCT applications, segregating LLCT applications brings an opportunity to trade the performance among applications.…”

“…Many efforts [4,10,13,17,24,36,41] classify workloads based on hardware profiling, and then choose appropriate scheduling policies for different classifications or create performance model for analysis. The latest work in [3] proposes a promising cache partitioning and sharing approach based on a recent advance in locality theory in [37]. OS-level approaches for memory utilization monitoring [7,8,39,40] have also been studied to assist resource management.…”