PrefixBench-H100: Artifact for IISWC 2026 submission #164

PrefixBench-H100 is a reproducible benchmark and measurement framework for
characterizing prompt-prefix reuse (KV-cache reuse) in large language model
serving on a single NVIDIA H100 NVL GPU.

The benchmark drives two production serving runtimes -- vLLM 0.21.0 and
TensorRT-LLM 1.2.1 -- through identical OpenAI-compatible workloads from a
single load generator, so that measured differences are attributable to the
runtimes rather than to differences in how each was benchmarked. It sweeps
shared-prefix length, suffix diversity, request concurrency, arrival pattern
(burst and Poisson), output length, KV-cache block size, KV-cache capacity,
and model scale, collecting time-to-first-token, end-to-end latency,
throughput, prefix-cache hit statistics, queueing and prefill time, and GPU
memory usage.

CONTENTS

- Workload generators: synthetic repeated-prefix traces with exact token
  counts, and RAG-style template traces.
- Matched load driver (asyncio/aiohttp) supporting burst and Poisson arrivals.
- Runner scripts for both runtimes, including flags for selecting the
  TensorRT-LLM backend (PyTorch or TensorRT engine) and for capping KV-cache
  capacity to induce cache pressure.
- 102 raw per-cell measurement results (JSON) and a merged CSV.
- Analysis and plotting scripts, plus a Jupyter notebook (analysis.ipynb)
  that regenerates every figure in the paper.
- Reproducibility documentation pinning all component versions.

REPRODUCTION

Every figure and table can be regenerated from the included raw results with
no GPU access: install the Python analysis dependencies and run
analysis.ipynb. Re-running the full measurement campaign requires a single
H100 NVL (94 GB) and takes approximately three GPU-hours; scripts are
provided for environment setup, model preparation, GPU settings, and sweep
orchestration.

This artifact accompanies a paper submitted to the IEEE International
Symposium on Workload Characterization (IISWC) 2026. Author information is
omitted while the paper is under double-blind review and will be added to
this record after publication.