Paper notes – Energy efficiency across programming languages
Which languages are the fastest and most energy efficient? The simple answer is: C, C++, Rust. The accurate answer is: it depends.
Papers
Pereira, R., Couto, M., Ribeiro, F., Rua, R., Cunha, J., Fernandes, J.P., and Saraiva, J. (2017). Energy efficiency across programming languages: how do energy, time, and memory relate? In Proceedings of the 10th ACM SIGPLAN International Conference on Software Language Engineering (ACM), pp. 256–267. https://doi.org/10.1145/3136014.3136031
And
Pereira, R., Couto, M., Ribeiro, F., Rua, R., Cunha, J., Fernandes, J.P., and Saraiva, J. (2021). Ranking programming languages by energy efficiency. Science of Computer Programming 205, 102609. http://dx.doi.org/10.1016/j.scico.2021.102609
Notes
5 years is a long time in technology, so it’s a good thing the same authors have a followup paper published in 2021 to take another look. The 2021 paper validates the 2017 results.
There is a general assumption that energy consumption is related to execution time because Energy = Power * Time. However, this paper shows that both Power and Time can vary across language.
Benchmarks are always a risky business – they can be easily set up to show what you want them to show. This paper uses the The Computer Language Benchmarks Game (CLBG) as a well understood suite of language benchmarks to attempt to provide reproducible test scenarios.
The 2021 paper adds a more “real world” analysis on a codebase that better represents day-to-day programming problems. This does not show any major differences that the CLBG misses and validates the results from 2017.
The tests were performed using Intel’s RAPL tool, which provides energy consumption metrics from Intel CPUs. This was a good choice at the time, but in 2022 ARM chips are of major relevance. Not only are they in billions of mobile devices – where energy efficiency is really important – but Apple is pushing their entire line of computers to ARM. Energy efficiency is just as important in laptops.
Tests were done on an Ubuntu desktop. This might make the test setup easier, but how relevant is it as representative of real world environments? Using server hardware would be interesting, particularly as ARM is also growing on the server – Google Cloud just announced ARM-based instances so we now have all three cloud providers offering ARM in the data center.
OS-level optimization would be another factor to consider. Mobile devices have done this for a while, but Apple talks about the different types of cores in its systems. High-efficiency vs high performance cores will express different characteristics:
An app may execute threads on both P and E cores over a period of time. The OS places threads on P or E cores based on the following criteria: Information your app provides, Observation of the app’s workload, Observation of the system as whole. On Apple Silicon Macs, the system observes applications and daemons separately from each other. This allows the system to execute them with individual efficiency and performance characteristics. As an example, an app running in the background may have its threads placed on E cores to optimize battery life while the foreground app is taking advantage of P cores.
Optimize for Apple Silicon with performance and efficiency cores, Apple.
The results show that C is the fastest and most energy efficient programming language. The top 5 languages were all fairly consistent across all three considerations – energy efficiency vs execution time vs memory usage.
Compiled languages are almost always fastest and most energy efficient
The most energy efficient language is almost always the fastest language, but there is no one language that consistently beats all other languages in every benchmark.
The methodology ran each benchmark 10 times and removed outliers, then calculated median, mean, standard deviation, min, and max values. The authors state this allowed them to remove variability caused by system processes. It would be interesting to run these tests in serverless environments because cold-starts are an important characteristic that needs to be handled and may have an impact on the overall results in cloud environments.
Conclusions
Which languages are the fastest and most efficient?
The simple answer is: C, C++, Rust.
The accurate answer is: it depends.
To quote from the 2017 paper:
The most common performance characteristics of software languages used to evaluate and choose them are execution time and memory usage. If we consider these two characteristics in our evaluation, C, Pascal, and Go are equivalent. However, if we consider energy and time, C is the best solution since it is dominant in both single objectives. If we prefer energy and memory, C and Pascal constitute the Pareto optimal set. Finally, analyzing all three characteristics, this scenario is very similar as for time and memory.
If the developer is only concerned with execution time and energy consumption, then yes, it is almost always possible to choose the best language. Unfortunately, if memory is also a concern, it is no longer possible to automatically decide for a single language.