Abstract
In this paper, we propose a PMC-based power modelling methodology that utilizes workload classification. Unlike traditional approaches which use a limited set of benchmarks, our methodology uses a single well-designed micro-benchmark to collect samples of PMCs and power values for training the power estimation model. The micro-benchmark can generate a large variety of representative workloads that are generic in a wide range of applications. Since the micro-benchmark is independent from any applications but includes generic workloads of many applications, our methodology is more widely applicable than the approaches based on a limited set of benchmarks that may have similar workloads. Another novelty of our methodology is that it adopts workload classification. Traditional approaches usually use multi-variable linear regression to correlate PMCs with power for all types of workloads. Since different PMCs may correlate well with power under different workloads, using a single linear multi-variable function to model power is insufficient and ineffective. In our methodology, we classify the workloads and for each workload we use a different, independent linear function to model the relationship between PMCs and power. In this way, the resulting power model is refined and its accuracy of power estimation can be increased. Based on our methodology, we have implemented a power estimation model called W-Classifier. Experimental results show that W-Classifier can estimate power usage well for a larger variety of workload types than the traditional approaches with a single multi-variable linear regression function.