Improving Prediction Accuracy for Power Consumption in Virtual Environments

Modern processors with multi-cores and several power states make power modeling of servers challenging. This issue becomes more complex in virtualized environments, owing to the presence of hypervisors. Requests and instructions as processed at the guest (i.e Virtual Machine, VM) level cannot be straightforwardly related to instructions processed at the host, and this relation varies with changing virtual environment configuration. However, observing performance at both levels, host and guest, might be helpful in developing realistic performance models. Also, the scale of virtual systems in modern computing environments such as data centers and clouds is very large, and these systems have massive data to manage, in terms of their allocation, scheduling, migration, etc. For such heterogeneous systems and large scale data, machine learning (ML) methodologies can play a vital role. Developing power and performance models using effective performance counters at host and guest level can provide significant features for training data. In this research work correlated performance counters for the running applications are monitored and trained for different models. This novel approach of monitoring performance counters at both levels provides in-depth performance information about individual VMs and servers. Results show that estimating power using these models reduces the prediction error, and hence can help in providing more effective power-aware decisions.