Processors have evolved to the now de-facto standard multicore architecture. The continuous advances in technology allow for increased component density, thus resulting in a larger number of cores on the chip. This, in turn, places pressure on the off-chip and pin bandwidth. Large Last-Level Caches (LLC), which are shared among all cores, have been used as a way to control the out-of-chip requests. In this work we focus on analyzing the memory behavior of a modern demanding application, a graph-based database workload, which is representative of future workloads. We analyze the performance of this application for different cache configurations in terms of: memory access time, bandwidth requirements, and power consumption. The experimental results show that the bandwidth requirements reduce as the number of clusters reduces and the LLC per cluster increases. This configuration is also the most power efficient. If on the other hand, memory latency is the dominant factor, assuming bandwidth is not a limitation, then the best configuration is the one with more clusters and smaller LLCs.

Peer Reviewed