Parallel Program Performance Prediction and Collaborative Filtering Techniques

Hardware trends point to architectures with increasing numbers of cores and heterogeneity [1]. However, software developed today can only be tested in significantly smaller platforms, due to cost and availability of such platforms. Accurate performance prediction mechanisms would provide many benefits, including improved platform provisioning, which is a significant problem in cloud computing, and could also help identify the hardware characteristics that impede performance. This research proposal presents two different approaches to performance predictions: a regression based approach for message-passing applications [2] and a performance modeling toolkit that combines application characteristics and target architecture models to produce cross-architecture performance predictions [3]. Summarizing their advantages and limitations, we conclude that a different approach might be more successful, by using techniques from a different area of computer systems, namely collaborative filtering. Consequently, we present and compare different item-based collaborative filtering recommendation algorithms [4] that can be used to identify similarities given a set of common characteristics and present the opportunities, combining ideas from the aforementioned works, for research in the area of performance predictions using collaborative filtering algorithms.