Skip to main content

Artifact Evaluation

Each LLMServingSim paper that ships an artifact lives on its own branch, frozen at the state submitted to the artifact-evaluation committee. This page is the entry point for reviewers and readers who want to reproduce the published figures end to end.

Heads up: the artifact branches are frozen for reproducibility. Don't open PRs against them; new development goes to main. See For Contributors → PR workflow.

Available artifacts​

PaperVenueBranchReproduces
LLMServingSim 2.0ISPASS 2026ispass26-artifactFigures 5–10
LLMServingSimIISWC 2024(released artifact, see Zenodo DOI)Original paper figures

The CAL 2025 entry shares the ISPASS 2026 codebase and does not have its own artifact branch.

ISPASS 2026 — ispass26-artifact​

Cho, Choi, Heo, Park. "LLMServingSim 2.0: A Unified Simulator for Heterogeneous and Disaggregated LLM Serving Infrastructure", ISPASS 2026. Zenodo DOI.

The branch reproduces Figures 5 through 10 of the paper, plus the supporting throughput / power / memory / latency parsers under evaluation/.

The artifact pre-dates the v1.1.0 directory restructure and the vLLM-based profiler rewrite, so on ispass26-artifact you'll see the older layout (cluster_config/, dataset/, output/, inference_serving/, main.py) instead of the serving/ / configs/ / workloads/ / outputs/ paths that the rest of this site documents. Follow the branch's own README, not this site's Getting Started, while you're inside the artifact.

1. Switch to the artifact branch​

git clone --recurse-submodules https://github.com/casys-kaist/LLMServingSim.git
cd LLMServingSim
git checkout ispass26-artifact

If you already cloned, just git checkout ispass26-artifact and git submodule update --init --recursive to pick up the pinned ASTRA-Sim submodule.

2. Set up the environment​

The artifact ships its own Docker launcher and build script (rather than the two-container split on main):

./docker.sh        # launches the artifact's simulator container
./compile.sh       # builds ASTRA-Sim + Chakra inside the container

docker.sh mounts the repo at /app/LLMServingSim. Run all subsequent commands from that working directory inside the container.

3. Reproduce a single figure​

Each figure has its own driver script under evaluation/:

cd evaluation

bash figure_5.sh        # Hardware coverage (A6000, H100)
bash figure_6.sh        # Multi-instance + P/D disaggregation
bash figure_7.sh        # MoE expert parallelism + offloading
bash figure_8.sh        # Prefix caching across CPU / CXL pools
bash figure_9.sh        # CXL memory expansion
bash figure_10.sh       # Power and energy modeling

Each script writes intermediate logs to evaluation/figure_X/logs/, parsed numbers to evaluation/figure_X/parsed/, and the final PDF next to the script.

4. Reproduce everything​

cd evaluation
bash run_all.sh

This is the same as running all six figure_*.sh scripts in sequence. Expect this to take a few hours on a single workstation; each figure runs many simulator invocations.

5. Compare against the preserved snapshot​

Frozen reference outputs live under evaluation/artifacts/. To compare your generated parsed output against those snapshots:

# Compare every figure
bash compare.sh

# Compare one figure
bash compare.sh 5

# Compare a subset
bash compare.sh 5 7 9

For visual confirmation, diff the regenerated figure_X.pdf against the committed figure_X_ref.pdf (or figure_Xa_ref.pdf for multi-panel figures) in each folder.

Per-figure details​

Each evaluation/figure_X/ folder has its own README.md with the figure's goal, axis definitions, reference inputs, expected TSV files, and the PDF naming convention. Start there if a figure fails to reproduce or numbers drift outside the comparison tolerance.

The umbrella reference is evaluation/README.md, which lists the parsers, fonts, and folder layout used across all figures.

When reproduction fails​

A few common cases:

  1. compile.sh errors on the submodule: rerun git submodule update --init --recursive from the host and try again. The submodule pin is part of the artifact.
  2. figure_X.sh runs but the parsed output doesn't match: check the corresponding evaluation/figure_X/README.md for the tolerance band the artifact was certified at; small drift on the exact wattage or latency value is expected as long as the qualitative trend matches the reference PDF.
  3. A specific simulator command fails on the branch but works on main: that's expected. The artifact is frozen at the paper's submission state; bug fixes and new features that landed on main afterward are not back-ported.
  4. You need to extend the artifact (e.g., add a new GPU to Figure 5): we recommend doing the work on main instead and citing the new result separately. The artifact branch should stay reproducible against the paper.

Reaching the artifact authors​

For artifact-specific questions (reproduction failures, environment setup, requesting a missing reference output), email the main contributors:

CC both whenever possible. See the Contact page for the full list of channels.