Detecting Idle-Power Regressions in Linux

BBB-Link: https://bbb.tu-dresden.de/b/mat-xin-oyh-xzn

Presentation Language: English

Optimizing idle power is critical for modern computers because it substantially dictates their lifetime energy consumption and thus has a significant impact on their operational costs as well as environmental footprint. However, sustaining these efficiencies remains a significant challenge due to complex, regression-prone operating system mechanisms that are notoriously difficult to accurately measure and maintain. Previous work focused on pinpointing such regressions in a limited kernel history. And while pinpointing was described, we currently still know little about the requirements to detect (unknown) idle-power regressions.
In this project we investigated the requirements for a broader kernel scan for idle-power regressions, such as the limits and restrictions existing. The knowledge gained was used to create a system for broader scans of the Linux kernel history. This was done by tailoring artificial kernel idle-power regressions and statistically evaluating of such to then implement a system allowing performant kernel scans. In this process we were able to identify the key limitations of such wide kernel scans and were able to identify a potentially previously unknown idle-power regression.