Transparent Offload

Code and measurements for the paper "Fine-Grained Computation Offload for Off-the-Shelf Servers in Tens of Lines" (Bojie Li, Pine AI).

What should the CPU do while it waits for a fine-grained accelerator offload (microseconds to a few milliseconds)? Blocking pays a context switch comparable to the offload itself; busy-waiting burns the core. The answer is to overlap the offload with other requests' work — and this repo measures how much an existing server must change to do that, across ten off-the-shelf servers and a zero-modification LD_PRELOAD fiber runtime, all on real hardware (NVIDIA RTX PRO 6000 Blackwell GPU; a real RSA-2048 TCP signer for the remote-offload class).

arXiv: https://arxiv.org/abs/2607.02630  ·  Project page: https://01.me/research/transparent-offload  ·  Paper PDF: paper/paper.pdf

Layout

Directory	Contents
paper/	LaTeX source, figures (figs/gen_figs.py), and the compiled paper.pdf
transparent-runtime/	The LD_PRELOAD M:N fiber runtime, its results (TRANSPARENCY_RESULTS.md), and the fiberizability classifier (sweep/)
transparent-runtime/apps/	The ten minimal-edit server integrations (Redis, nginx, memcached, Node.js, Python, Apache, Go, PostgreSQL, MariaDB, HAProxy) and MINIMAL_EDIT_RESULTS.md
runtime/	Coroutine-vs-block-vs-busy microbenchmarks and the open-loop latency study (openloop.csv)
detector/, conflict-measurement/	The page-protection conflict detector and the shared-state correctness study
gpu/	GPU AES kernels and latency characterization
transparency/	Offload-boundary interposition study

Start with INDEX.md, which maps every claim in the paper to its evidence file, and EXPERIMENT_PLAN.md for the full experimental plan. Build notes (CUDA version, scheduling privileges) are at the bottom of INDEX.md.

Citation

If you use this work, please cite the paper:

@misc{transparentoffload,
  author        = {Li, Bojie},
  title         = {Fine-Grained Computation Offload for Off-the-Shelf Servers in Tens of Lines},
  year          = {2026},
  eprint        = {2607.02630},
  archivePrefix = {arXiv},
  url           = {https://arxiv.org/abs/2607.02630},
  note          = {Code and data: https://github.com/19PINE-AI/transparent-offload},
}

License

MIT — see LICENSE.