Intelligent Feedback For Fast and Effective Autotuning
SESSION: Research Poster Reception
EVENT TYPE: Poster
TIME: 5:15PM - 7:00PM
AUTHOR(S):Santosh Sarangkar, Apan Qasem
ABSTRACT: Fast searches are integral to the success of autotuning systems that perform online tuning. However, finding ways to effectively navigate the vast and complex search spaces has been a major challenge. We present an autotuning system that improves search-time through diagnostic feedback and architecture-sensitive search heuristics. The system incorporates a wide array of performance indicators and collects measurements at loop-level granularity. This information is synthesized and fed into several performance models that guide the search. The main novelty of this approach is that it enables a search algorithm to examine causes - rather than symptoms - of performance bottlenecks and also allows it to react to performance changes and prune away non-favorable regions on-the-fly. The poster will highlight key features of our autotuning system and include experimental results that not only demonstrate the improvement in search times but also provide new insight into search space characteristics.