This February 2025 paper introduces ELMo-Tune-V2, a novel framework that leverages Large Language Models (LLMs) to fully automate the optimization of Log-Structured Merge-tree-based Key-Value Stores (LSM-KVS). Unlike previous methods that rely on human experts or limited automated tuning, ELMo-Tune-V2 integrates LLMs for self-navigated workload characterization, automatic tuning across a broad parameter space, and real-time dynamic configuration adjustments. The framework demonstrates significant performance improvements for popular LSM-KVS systems like RocksDB, addressing the complex interplay of hardware, resource limits, and evolving workloads. ELMo-Tune-V2 achieves this through innovations in LLM-based workload synthesis, feedback-driven iterative fine-tuning, and real-time adaptive tuning, showcasing the potential of LLMs in solving complex data system optimization challenges.

Source: https://arxiv.org/pdf/2502.17606