A Methodology for Co-Location Aware Application Performance Modeling in Multicore Computing

As multicore processor architectures are now prevalent in server nodes of parallel and distributed computing systems, it has become important to characterize the performance of applications run on these architectures. This study investigates the performance degradation an application experiences from memory interference due to other applications colocated on cores of the same multicore processor. We propose a methodology for designing models that are capable of utilizing varying amounts of information relating to an application and its co-located applications to predict the application´s execution time performance degradation due to co-location. We evaluate the models sing several application co-location scenarios based on real world test data from two scientific benchmark suites on two server class Intel Xeon multicore processors.