Appendix
Where to find the data, code and experimental artifacts behind the study.
Replication
The datasets, source code and experimental artifacts supporting the findings of this study are publicly available in the Zenodo repository. The package contains all figures, tables and underlying measurements, allowing readers to inspect results in far greater detail than the printed paper allows.
doi.org/10.5281/zenodo.22114323Cite
To cite this paper, please use the following BibTeX entry:
@misc{matos_extended_2026,
address = {Rochester, NY},
type = {{SSRN} {Scholarly} {Paper}},
title = {The {Extended} {Green} {Side} of the {Lua}},
url = {https://papers.ssrn.com/abstract=7661830},
doi = {10.2139/ssrn.7661830},
language = {en},
urldate = {2026-10-23},
publisher = {Social Science Research Network},
author = {Matos, Diogo and Brandão, André and Guimarães, Miguel and Cunha, Simão and Saraiva, João},
month = oct,
year = {2026},
keywords = {Ahead-of-Time, Benchmarking, Energy efficiency, Green Software, Just-in-Time, Lua},
}Section 5
The credibility and reproducibility of the findings are discussed along four categories.
Each benchmark was executed ten times without a prior statistical power analysis, which might be considered a small sample. However, the low variance observed across runs suggests additional executions would be unlikely to change the main conclusions, and the large number of executions in the battery tests provides strong confidence in those trends.
Background processes, network polling, dynamic frequency scaling and thermal throttling can interfere with measurements. All non-essential services were terminated, Wi-Fi and Bluetooth were disabled, and a cooldown period was enforced between runs. RAPL experiments ran on direct AC power; battery tests ran exclusively on battery under identical ambient conditions.
Energy is captured with Intel RAPL (Package and DRAM), a well-established standard. The CLBG-derived LuaAOT programs are optimized for raw speed rather than idiomatic Lua; LuaFun targets trace-JIT pipelines and the Penlight suites emphasize table-based metaprogramming. Including all five suites balances these characteristics.
All measurements come from a single x86_64 platform under Linux. The performance hierarchy between interpreters, AOT and JIT compilers should hold elsewhere, but exact speedups and energy savings may vary. The projection onto multi-language energy rankings is indicative rather than exact.
Contact
Funding
This work is funded by national funds through FCT — Fundação para a Ciência e a Tecnologia, I.P., under the support UID/50014/2025. Simão Cunha is also funded by FCT grant 2025.04596.BDANA.