Leveraging Power Caps to Save Energy Across Programming Languages

Energy efficiency of software is crucial in minimizing environmental impact and reducing operational costs of ICT systems. Energy efficiency is therefore a key area of focus in contemporary software language engineering research. A recurrent discussion that excites our community is whether runtime performance is always a proxy for energy efficiency. While a generalized intuition seems to suggest this is the case, this intuition does not seem to align with the fact that energy is the accumulation of power over time; hence, time is only one of the factors in this accumulation. We focus on the other factor, power, and the impact that capping it has on the energy efficiency of running software.

We conduct an extensive investigation comparing regular and power-capped executions of 9 benchmark programs obtained from The Computer Language Benchmarks Game across 20 distinct programming languages. Our results show that employing power caps can be used to trade running time, which is degraded, for energy efficiency, which is improved, in all the programming languages and in all benchmarks that were considered. Our findings include demonstrating overall energy savings of almost 14% across the 20 programming languages, with notable savings of 27% in Haskell. This saving, however, comes at the cost of an overall increase of the program’s execution time of 91% in average.

We were furthermore able to draw similar observations using language specific benchmark frameworks for programming languages of different paradigms and with different execution models. This was achieved analyzing a wide range of benchmark programs from the nofib Benchmark Suite of Haskell Programs, DaCapo Benchmark Suite for Java, and the Python Performance Benchmark Suite. Our findings reveal energy savings of approximately 8% to 21% across the test suites, with execution time increases ranging from 21% to 46%. Notably, the DaCapo suite exhibits the most significant values, with 20.84% energy savings and a 45.58% increase in execution time.

Our results have the potential to drive significant energy savings in the context of computational tasks for which runtime is not critical. These include Batch Processing Systems, Background Data Processing and Automated Backups and Maintenance, for example.