Evaluation of last level set-based cache partitioning techniques in terms of performance, power and fairness

Last Level Cache is an important resource that is shared among many cores in multi-core processors. In an unmanaged system, this resource is allocated on demand, which may cause performance problems when there is not sufficient cache space for all running applications. In order to achieve higher system performance, smart partitioning mechanisms must be utilized. This thesis describes a scalable, fine-grain and high-performance set-based partitioning mechanism, along with a tailor-made allocation policy. Evaluations yield that the proposed mechanism can provide 5% improvement on average, in terms of throughput and fairness, compared to Vantage partitioner, state-of-the-art in the literature. Additionally, set-based partitioning is investigated as a potential energy-saving technique. Average energy consumed per instruction by the cache is reduced by 28%, while losing less than 1% throughput on average, with minor modificiations in the allocation policy and enforcement scheme of the proposed set-based partitioner.