Energy Profiling vs. Runtime Profiling

Context

In present-day times of climate change all industry sectors should bring down their green house gas emissions (carbon footprint). However, the IT-sector’s emissions increase unchecked. The research area for this seminar thesis is the field of “green” software engineering. In this research area one aims at finding ways to lower the energy consumption of a piece of source code when run as a program. Using specialized registers (so called model specific registers) that deliver the cumulated energy consumption of the CPU one can track back energy consumption to different parts of a program.

Motivation

In theory, when CPU voltage and frequency are fixed the energy consumption of source code is proportional to its time “consumption”. We want to give experimental evidence for this theoretical fact. We want to find out if energy profiling of source code when CPU voltage and frequency are fixed is equivalent to runtime profiling.

Goal

• The student will read one or more research papers about this topic.
• We aim at profiling benchmark functions/projects. The student will write Shell and MATLAB scripts for automating the measurements.
• Say, we have two functions f1 and f2. Further we call X% the ratio of their runtimes and Y% the ratio of their energy consumptions. We’re interested in whether the equation X = Y holds - under the constraint that the CPU voltage and frequency are fixed.

Requirements

• Linux: user skills
• MATLAB: basic to advanced skills

Pointers

• Dynamic voltage and frequency scaling (Wikipedia):  https://en.wikipedia.org/wiki/Dynamic_frequency_scaling
• CMOS (Complementary metal–oxide–semiconductor) (Wikipedia):  https://en.wikipedia.org/wiki/CMOS
• CPU Power Consumption Experiments and Results Analysis:  http://async.org.uk/tech-reports/NCL-EEE-MICRO-TR-2015-197.pdf