Reproduce “Large Language Models as Realistic Microservice Trace Generators” on OpenResearch

About this paper

Donghyun Kim, Sriram Ravula, Taemin Ha, Alexandros G. Dimakis, Daehyeok Kim, Aditya Akella

The paper introduces TraceLLM, a framework that adapts Large Language Models to generate realistic and diverse synthetic microservice call graphs by explicitly learning a recursive generation process combined with instruction tuning. This approach produces complex, valid traces that accurately reflect real-world distributions and are suitable for training downstream machine learning models.

Key insights

• Large Language Models can effectively generate complex, structured system traces by translating graphical data into a textual representation and employing a recursive generation strategy.
• Explicit instruction tuning, augmented with step-by-step intermediate reasoning, significantly improves an LLM's capability to adhere to intricate structural and domain-specific constraints during synthetic data generation.
• High-fidelity synthetic microservice traces, generated by LLMs with appropriate conditioning, are suitable for training machine learning models for critical system management tasks, performing comparably to real data.

Summary by alphaXiv.