Improving the scalability of shared cache multi-core systems

Abstract

Resource sharing in multi-core processors leads to performance variability due to cross-core interference between the cores. To minimize the effects of cross-core interference, dynamic cache partitioning can be applied at run-time. Cache partitioning can be applied to improve the utilization of the shared cache. However, this scheme must be scalable as the number of cores increases. The Lotka-Volterra competition model is a popular model in biology which models population growth based on species' interactions. This paper uses a mathematical model of the shared-cache multi-core processor based on the Lotka-Volterra (LV) competition model and investigates the effectiveness of the dynamic cache partitioning scheme as the number of cores sharing the cache increase. Based on the results of the investigations, the paper proposes an improved cache topology which increases the overall system throughput with partitioning.