valeriol29/mooncake-traces

Dataset Card for Dataset Name

Mooncake online traces with share prefixes.

Dataset Details

Dataset Description

The dataset is processed from the Mooncake official repository.

• Curated by: MoonshotAI
• License: Apache 2.0

Uses

This is mainly for inferencing system benchmarking workload.

• Timestamp: The timestamp field indicates the relative arrival times of requests, ranging from 0 to 3,600,000, in milliseconds.

• Input & Output Length: For privacy protection, our trace does not include actual text or tokens. Instead, it uses input_length and output_length, representing the number of input and output tokens, similar to Splitwise.

• Hash ID: The hash_ids field describes prefix caching relationships. It is generated by hashing token blocks (with a block size of 512) into prefix hash values that include both the current and all preceding blocks. The resulting hash values are then mapped to globally unique IDs. Identical hash IDs indicate that a block of tokens, along with preceding tokens, are the same, thus allowing reuse within the corresponding KVCache. For example, in the provided samples, the first 12 hash IDs are identical, indicating they can share prefix caching for the first 12×512=6,144 tokens.

[
  {
    'hash_ids': [46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 2111, 2112],
    'input_length': 6955,
    'output_length': 52,
    'timestamp': 27000
  },
  {
    'hash_ids': [46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 2124],
    'input_length': 6472,
    'output_length': 26,
    'timestamp': 30000
  }
]

Dataset Card Authors and Contact

• Jiacheng Li: jiacheng.li@epfl.ch

Citation

BibTeX:

@article{qin2024mooncake,
  title        = {Mooncake: A KVCache-centric Disaggregated Architecture for LLM Serving},
  author       = {Ruoyu Qin, Zheming Li, Weiran He, Mingxing Zhang, Yongwei Wu, Weimin Zheng, and Xinran Xu},
  year         = {2024},
  url          = {https://arxiv.org/abs/2407.00079}
}

@inproceedings {qin2025mooncake,
  author       = {Ruoyu Qin and Zheming Li and Weiran He and Jialei Cui and Feng Ren and Mingxing Zhang and Yongwei Wu and Weimin Zheng and Xinran Xu},
  title        = {Mooncake: Trading More Storage for Less Computation {\textemdash} A {KVCache-centric} Architecture for Serving {LLM} Chatbot},
  booktitle    = {23rd USENIX Conference on File and Storage Technologies (FAST 25)},
  year         = {2025},
  isbn         = {978-1-939133-45-8},
  address      = {Santa Clara, CA},
  pages        = {155--170},
  url          = {https://www.usenix.org/conference/fast25/presentation/qin},
  publisher    = {USENIX Association},
  month        = feb
}