Exploration of the Influence of Program Inputs on CMP Co-scheduling

Recent studies have showed the effectiveness of job co-scheduling in alleviating shared-cache contention on Chip Multiprocessors. Although program inputs affect cache usage and thus cache contention significantly, their influence on co-scheduling remains unexplored. In this work, we measure that influence and show that the ability to adapt to program inputs is important for a co-scheduler to work effectively on Chip Multiprocessors. We then conduct an exploration in addressing the influence by constructing cross-input predictive models for some memory behaviors that are critical for a recently proposed co-scheduler. The exploration compares the effectiveness of both linear and non-linear regression techniques in the model building. Finally, we conduct a systematic measurement of the sensitivity of co-scheduling on the errors of the predictive behavior models. The results demonstrate the potential of the predictive models in guiding contention-aware co-scheduling.