Viraj Thakkar

Log-Structured-Merge tree-based Key-Value Stores (LSM-KVSs) are important data storage building blocks in modern IT infrastructure. However, tuning their performance involves configuring over 100 parameters, a task typically done manually or with limited parameters in auto-tuning mechanisms. This paper explores and answers the following question: can we leverage LLM’s understanding of the system and LSM-KVS components for unrestricted parameter-pool tuning of LSM-KVS?LLMs are trained on readily available LSM-KVS source code, research papers, and open materials enabling the machines to have human-like understanding. We investigate integrating Large-Language Models (LLMs) into an automated tuning framework for LSM-KVS to enhance the tuning capability and interactivity. Our framework utilizes LLMs to recommend tailored configurations with calibrated prompts based on hardware, system, and workload information. Initial results demonstrate upto 3X throughput improvements and an upto 9X reduction in p99 latency across various hardware and workloads compared to the out-of-box configuration for the LSM-KVS.