harvardMadsys/2026_prefixcache

LLM Prefix-Cache Block Traces

Two LLM inference traces at KV-block granularity, for studying prefix-cache eviction. Each request lists the KV blocks of its prompt in prefix order, so a simulator can replay which blocks a request reuses from earlier requests. No text is released: a block is an opaque integer id.

The dataset accompanies the paper When Fancy Eviction Fails: Rethinking Cache Replacement For LLM Prefix Reuse (arXiv:2609.28870).

Both traces are tokenized with the Qwen3-Coder-30B-A3B tokenizer and cut into 16-token blocks, so 1k blocks hold 1.5 GiB of that model's bf16 KV cache. Requests whose input plus output exceed the model's 262,144-token context are removed, so no request is longer than 16,374 blocks.

File	Size	Workload	Requests	Block accesses	Distinct blocks	Sessions	Span
freeinference.lcp.zst	76 MB	Coding and assistant agents on the FreeInference gateway, 2026-07-11 to 07-17	327,548	656,240,179	39,578,367	222,934	7.0 days
chutes.lcp.zst	232 MB	Chat and roleplay on the Chutes inference proxy, 2026-03-10 to 07-19	515,817	603,642,254	139,518,804	402,221	130.9 days

The request-level traces of these platforms are harvardMadsys/freeinference_agentic_trace and harvardMadsys/chutes_metadata.

Format

The files are in lcp, libCacheSim's binary format for prefix-cache traces, compressed with zstd. Each request is a record followed by its block entries:

• Request: arrival and completion time (microseconds, relative), session and user id (chutes has no user ids), turn, and input and output token counts. A request's blocks become reusable when it completes, and requests are ordered by completion time.
• Block entry: the block id, and the position of the block's next access, so offline-optimal policies such as Belady run in one forward pass.

The full specification is doc/lcp_format.md.

Usage

See libCacheSim for how to build and use the tools. In short:

hf download harvardMadsys/2026_prefixcache --repo-type dataset --local-dir 2026_prefixcache

git clone https://github.com/1a1a11a/libCacheSim
cd libCacheSim
cmake -G Ninja -B _build -DCMAKE_BUILD_TYPE=Release
ninja -C _build

# Replay a 16k-block (24 GiB) cache under three eviction policies.
./_build/bin/prefixsim --trace ../2026_prefixcache/freeinference.lcp.zst \
    --cache-size 16k --algo lru,s3fifo,belady_compute \
    --cost-model qwen3coder30b_blksz_16

# Print a trace's statistics, or convert it to JSONL.
./_build/bin/tracePrint ../2026_prefixcache/chutes.lcp.zst lcp --print-stat 1
./_build/bin/traceConvLLM --input ../2026_prefixcache/chutes.lcp.zst --output chutes.jsonl

prefixsim streams the compressed file: an LRU run over freeinference.lcp.zst takes about 2 minutes and 150 MiB of memory. Its quickstart explains the options and the output, and its design document explains the cache model.

Integrity

Each file is a single zstd frame with zstd's built-in content checksum, so zstd -t <file> checks it. SHA-256:

File	SHA-256
freeinference.lcp.zst	39f75bbba1ce912546517802719ef7742d0db7fa53dcbc66681bba090aa5d1f7
chutes.lcp.zst	da57f89c74db817c121d24d3d6e4bc415eb7dd54aa72548c6971a991bd99970b

License

This dataset is released under the Creative Commons Attribution 4.0 International License (CC BY 4.0).

If you use this dataset, please cite the paper below.

Citation

@misc{liu2026fancyeviction,
  title         = {When Fancy Eviction Fails: Rethinking Cache Replacement For LLM Prefix Reuse},
  author        = {Liu, Yiyu and Yu, Minlan and Yang, Juncheng},
  year          = {2026},
  eprint        = {2609.28870},
  archivePrefix = {arXiv},
  primaryClass  = {cs.DC},
  url           = {https://arxiv.org/abs/2609.28870}
}