AutoTunium: An Evolutionary Tuner for General-Purpose Multicore Applications

Today's increasing diversity in multicore hardware challenges programmers when it comes to software performance optimization and portability. As multicore processors are in almost every PC and server, programmers now have to parallelize a larger spectrum of applications, many of which are non-numerical. To obtain good performance, programmers typically try out different software tuning parameter configurations on each platform. However, this manual approach to finding good configurations in the search space is impractical due to combinatorial explosion, but yet it is common practice due to lack of alternatives for general programs. This paper presents a smarter way to tackle this problem algorithmically for a variety of multicore applications, including non-numerical ones. Our work introduces AutoTunium, a novel feedback-directed optimizer that automates the application tuning process with evolutionary search strategies. The software infrastructure is easy to use and integrated in the popular Eclipse environment. It collects run-time information to predict parameter configurations that are likely to lead to good performance in future runs, and configures programs for production runs in the best possible way. We quantify the effectiveness of various tuning strategies on a diverse set of real applications and multicore platforms. The evaluation shows that AutoTunium's evolutionary strategies work well despite the broad scope of applications and perform better in this context than other simplex-based search algorithms. Our insights are derived from model-based analyses as well as from performance analyses with real programs in the PARSEC benchmark suite.