Extensible and Versatile Energy Measurement Framework for CPUs

Minimizing energy consumption is crucial, from mobile devices to datacenters. An important processor feature in this regard are “sleep states”. They allow to temporarily deactivate idle cores, lowering their energy consumption. Reactivating a core from a sleep state takes time, however, which can impact performance. Modern processors usually support a variety of sleep states, each taking different measures to save energy. Sleep states taking more drastic measures save more energy, but need more time for reactivation. The operating system has to navigate this energy-performance trade-off when selecting the most appropriate sleep state once a core becomes idle. To make an optimal decision, it needs accurate data on the energy consumption and wake-up latency of each sleep state. The data commonly used for this task is of questionable quality, however.

This thesis presents a framework capable of measuring these sleep state characteristics, allowing to gather first-hand, trustworthy data instead. The framework builds upon an already existing implementation which so far measured the energy consumption of sleep states on two Intel processors. In this thesis, I add wake-up latency measurements. Additionally, I extend the framework to work on AMD processors. I also add support for ARM, for which I introduce external energy measurements to the framework. I test the implementation on four different systems, using an Intel Core i7-4790, an Intel Core i7-6700K, an AMD Ryzen 5 5600G and an ARM Cortex-A76, respectively. Results show that the framework works well across manufacturers and architectures, providing precise and reproducible values for both sleep state energy consumption and wake-up latencies.

The talk will be held in German.

(Master Thesis)