Abstract
This paper investigates the problem of partitioning the last-level shared cache of multicore architectures. Contention for such a shared resource has been shown to severely degrade performance when running multiple applications. As architectures incorporate more cores, multiple application workloads become increasingly attractive, further exacerbating contention at the last-level cache. Today, cache replacement policies, extensively studied for uniprocessor systems, are being employed within new multicore architectures with little, if any, adaptation. However the parameters in these new systems are likely to be different. The least recently used (LRU) policy, for example, which is widely accepted as the best replacement policy in uniprocessor caches, often results in poor resource sharing in a multicore system, signalling the importance of reevaluating the effectiveness of these policies in the new architectures.
This paper proposes Adaptive Bloom Filter Cache Partitioning., (ABFCP), a low-cost, dynamic cache partitioning mechanism capable of better resource sharing at the last-level cache than LRU, improving the performance of an eight-core system on average by 5.92% over the LRU policy. Moreover, the proposed scheme provides the equivalent performance benefits that could be gained from almost a 50% increase in the last-level cache and shows increasing benefit as the number of cores rises.