Modeling Dwarfs for Workload Characterization

Workload characterization is an important concept in performance tuning and efficiency improvement of high performance computing systems. Not only does it allow a system to dynamically adjust resources, it also helps improve energy efficiency resulting in lower cost for computation. Nevertheless, lack of quantitative methods to describe workload characteristics prevents the system to work at its full potential. In this work, we proposed two discrete-time Markov chains, which can describe workload characteristics of a set of computational kernels, called Berkeley\´s "dwarfs". These discrete-time models were derived from the relationship between stalling in processor pipeline and resource usage behaviors of the workload. The models can quantitatively describe behaviors of each dwarf. Moreover, the model enables the concept of combining each dwarf to extend performance analysis over any arbitrary processes.