Is Correlation Ranking Really Reliable for the Performance Counter Selection Conducted for Power Estimation?

Abstract

Power estimation is of great value to power-aware adaptation and power provisioning in computing platforms. Performance counter measurements are widely used as main input of the power estimation. However, the number of hardware counters that can be simultaneously measured is relatively small, and the number of OS-level counters is usually far from endurable. As a result, how to pick up the most significant counters becomes an important pre-step of performance counters based power consumption modeling. As the most important criteria of counter selection, correlation between counter measurement and power dissipation has been widely used for several decades, and thus correlation ranking becomes a typical technique for counter selection. However, few works discuss its accuracy and the question why it should be the general first priority. This paper investigates the weightiness of the correlation ranking in counter selection for power estimation. Through comparing correlation ranking with several popular feature selection approaches on tremendous workloads on single and multiple core platforms, we obtain that correlation ranking is not optimal on a large number of benchmarks. Therefore, it can be concluded that correlation ranking should not be generally considered as the first priority to select performance counters. This paper provides an indication for potential researchers to be aware of these issues when estimating power dissipation.