BurstGPT: ChatGPT and GPT-4 Workload Traces for LLM Serving Optimization

Burstgpt provides workload traces for ChatGPT (GPT-3.5) and GPT-4, released by HPMLL to facilitate the optimization of Large Language Model (LLM) serving systems. The data captures request patterns and arrival characteristics from production-scale models as of early 2024. It is designed to help researchers model the 'bursty' nature of inference traffic in high-performance computing environments.

Use Cases

• Simulating request arrival rates to optimize KV cache eviction policies
• Benchmarking inference engine throughput using GPT-4 burst patterns
• Designing auto-scaling logic for LLM clusters based on GPT-3.5 workload fluctuations

Strengths

• Contains traces for both GPT-3.5 and GPT-4 models
• Released under CC-BY-4.0 for open academic and commercial research
• Specifically targets 'bursty' traffic patterns in LLM serving

Limitations

• Lack of specific column documentation in the provided metadata
• Does not include raw prompt or response text due to privacy constraints
• Unknown total record count or file size

Provenance

Source

HPMLL (High Performance Machine Learning Laboratory)

Collection Method

Logged workload traces from LLM serving environments

Freshness

Last updated February 2026, reflecting modern LLM usage patterns.