Fair and adaptive online set-based cache partitioning

In a multi-processing environment extra misses are endured in the last level shared cache because each process is trying to utilize the whole cache space, resulting in high interference between their working sets, and thus a degraded performance. In this paper, we propose a new cache partitioning scheme, which opposed to prior partitioning schemes in the literature, offers both high performance and fair allocation even with increasing number of simultaneously running processes and decreased cache associativity. The proposed scheme takes advantage of the broad number of sets in the cache by partitioning it dynamically set-wise. The partitioning presented is adaptive to accommodate to the changing requirements of the mix of running processes without the need for any offline information and with an eye on fair allocation of cache resources. Performance is evaluated by experimentation on different cache sizes with different associativity and comparison with the Least Recently Used baseline cache and partitioning schemes. We experimented on mixes from the Splash Benchmark Suite. The results shows a performance speedup up to 44% with an average speedup of 15% over all mixes, and also fairness improvement up to 46% in terms of the weighted harmonic mean, with an average of 14% over all mixes.