Machine learning based performance prediction for multi-core simulation

Abstract

Programs co-running on cores share resources on multi-core processor systems. It is now well known that interference between the programs arising from the sharing may result in severe performance degradations. It is the objective of recent research in system scheduling to be aware of shared resource requirements of the running programs (threads). To this end AKULA is a toolset recently developed that provides a platform for experiments and developing thread scheduling algorithms on multi-core processors. In AKULA a bootstrapping module works on the basis of previously collected performance data of programs to simulate program execution on multi-cores. In this paper we describe a different approach where that augments such a bootstrapping module with a model built using machine learning techniques. The proposed model will extend the bootstrapping module's ability to predict degradation in performance due to sharing where previous performance data is not available for pairing/co-scheduling of applications. Also the proposed approach allows greater scalability for variable number of processor cores sharing the resources. © 2011 Springer-Verlag.