Resource Centric Characterization and Classification of Applications Using KMeans for Multicores

Abstract

The knowledge on the behavior of an application program towards consumption of shared resources in multicore systems could assist in characterizing and classifying these programs. Further categorizing applications assists in predicting optimal coschedules for multicores, which eventually leads to lower contention and enhance performance. The proposed work characterizes applications on the basis of variations in IPC due to various resource allocations. Further classification is done based on parameters of cache memory and Dram bandwidth utilization obtained using hardware counters. A statistical approach is used for classifying the applications. The variance values obtained for an application's behavior towards different resource allocations is considered to build training and test set and KMeans learning algorithm is applied to classify the workloads. The accuracy obtained with the proposed method is 85.71%.