Hyper-heuristics for performance optimization of simultaneous multithreaded processors

In Simultaneous Multi-Threaded processor datapaths, there are many resources that are concurrently shared by multiple threads. A few number of heuristic approaches, which explicitly distribute those resources among threads with the goal of an improved overall performance, have been proposed. A selection hyper-heuristic is a high level search methodology which mixes a predetermined set of heuristics under an iterative framework to exploit their strengths while solving a given problem. In this study, we propose a set of learning selection hyper-heuristics for predicting, choosing and running the best performing heuristic at periodic time intervals that we name epochs. The empirical results show that hyper-heuristics are capable of improving the performance of the studied workloads. The peak performance improvement is observed to be around 25 per cent over a previously proposed Hill Climbing heuristic and around 11 per cent over Adaptive Resource Partitioning Algorithm. Our best hyper-heuristic, HH4, performs better than either of the state-of-the art heuristics on almost 72 per cent of the simulated workloads. HH4 also beats both of the heuristics on around 30 per cent of the simulated workloads.