Energy Efficiency in Programming Languages

Nicolas van Kempen,
Hyuk-Je Kwon,
Dung Tuan Nguyen,
Emery Berger.

Read the paper!

Documentation

Requirements

Running the measurement tool requires:

• an Intel processor with RAPL support.
• Linux (tested on Ubuntu 22.04).

Docker

The easiest way to run these benchmarks is using Docker:

% sudo modprobe msr # Enable msr kernel module.
% sudo python3 -m scripts.build_docker_image
% sudo docker run --privileged -v [OUTPUT_DIRECTORY]:/root/data energy-languages [OPTIONS]

Here is an example running all languages/benchmarks pairs:

% sudo docker run -it --rm --privileged -v `pwd`/data/`hostname -s`/docker-default:/root/data energy-languages \
    --languages C C++ Rust Go Java C\# JavaScript TypeScript PHP Python PyPy Lua LuaJIT \
    --warmup 1 \
    --iterations 21 \
    --timeout 10000

Running additional experiments:

% sudo docker run -it --rm --privileged -v `pwd`/data/`hostname -s`/docker-default:/root/data energy-languages \
    --benchmark-root experiments \
    --languages "C as C++" "Go-no-GC" "JavaScript as TypeScript" \
    --warmup 1 \
    --iterations 21 \
    --timeout 10000

Running Java-N experiments:

% sudo ./scripts/docker-java-n.sh Java docker-default

Processors can be set to their minimum frequency with the following commands:

# Use frequency-info to check the defaults for your machine.
% cpupower frequency-info

# Pin to lowest possible frequency.
% echo 1 | sudo tee /sys/devices/system/cpu/intel_pstate/no_turbo
% sudo cpupower frequency-set --governor powersave
% sudo cpupower frequency-set --max 800Mhz

# Restore defaults after running experiments.
% echo 0 | sudo tee /sys/devices/system/cpu/intel_pstate/no_turbo
% sudo cpupower frequency-set --governor powersave
% sudo cpupower frequency-set --max 3.4GHz

Finally, experiments can be pinned to a single code by using Docker's
--cpuset-cpus option, specifying a core number. For example:

% sudo docker run -it --rm --privileged --cpuset-cpus=13 [...]

Figures

Average power draw as a function of the average number of cores used

% python3 -m scripts.normalize_cores \
    --data-root data/`hostname -s`/docker-default \
    --languages C C++ Rust Go Java C\# JavaScript TypeScript PHP Python PyPy Lua LuaJIT

Average power draw as a function of memory activity

% python3 -m scripts.dram \
    --data-root data/`hostname -s`/docker-default \
    --languages C C++ Rust Go Java C\# JavaScript TypeScript PHP Python PyPy Lua LuaJIT

Average power draw controlling for external factors

% python3 -m scripts.fixed_time \
    --data-root data/`hostname -s`/docker-min-freq-cpuset-13 \
    --languages C C++ Rust Go Java C\# JavaScript TypeScript PHP Python PyPy Lua LuaJIT

Other

A few other scripts can be useful to generate figures and results.
Their options are self-explanatory and/or documented in the --help message.
Notably:

• scripts.table: Build two tables comparing runtime and energy
  consumption for a list of languages.
• scripts.barchart: Compare two implementations of the same programming
  language.
• scripts.java_n: Generate a figure for a given benchmark of time per
  iteration as a function of number of iterations.

License

The original benchmark suite, the
Computer Language Benchmark Game,
is under
BSD-3-Clause.

Code from
the repository this one originally forked from
is under
MIT.

Any other code in this repository is under Apache-2.0.