This is a sample of the traces from multiple LLM inference services in Azure, collected on May 10th-19th 2024. This dataset is the data described and analyzed in the HPCA 2025 paper 'DynamoLLM: Designing LLM Inference Clusters for Performance and Energy Efficiency'.
The dataset comprises this description and a Jupyter Notebook with the plots in the HPCA paper.
The data is made available and licensed under a CC-BY Attribution License. By downloading it or using them, you agree to the terms of this license.
If you use this data for a publication or project, please cite the accompanying paper:
Jovan Stojkovic, Chaojie Zhang, Íñigo Goiri, Josep Torrellas, Esha Choukse. "DynamoLLM: Designing LLM Inference Clusters for Performance and Energy Efficiency", in Proceedings of the International Symposium on High-Performance Computer Architecture (HPCA 2025). IEEE, Las Vegas, USA, 2025.
Lastly, if you have any questions, comments, or concerns, or if you would like to share tools for working with the traces, please contact us at [email protected]
You can download the datasets here:
| Field | Description |
|---|---|
| TIMESTAMP | Invocation time |
| ContextTokens | Number of context tokens |
| GeneratedTokens | Number of generated tokens |
Due to customer privacy requirements (e.g., GDPR), we do not have visibility into the content of the prompts. We instead use the production traces to guide the input and output sizes, where we send the input prompt with the required number of tokens, and force the model to generate the corresponding number of output tokens for each request. Note that the text of the inputs prompts does not impact the performance metrics that we benchmark, since they depend only on the input and output sizes.
This data is the sample data used in the HPCA paper mentioned above.