Vulcan: LLM-Driven Heuristics for Systems Optimization

This paper introduces Vulcan, a novel approach to automate the design of system heuristics using Large Language Models (LLMs). It addresses the challenge of manually designing and maintaining performant heuristics in dynamic system environments. The core idea is to leverage LLMs to generate instance-optimal heuristics tailored to specific workloads and hardware. This is a significant contribution because it offers a potential solution to the ongoing problem of adapting system behavior to changing conditions, reducing the need for manual tuning and optimization.

• •Proposes Vulcan, a system that uses LLMs to generate instance-optimal heuristics for resource management.
• •Separates policy and mechanism using LLM-friendly interfaces.
• •Demonstrates performance improvements over state-of-the-art human-designed algorithms in cache eviction and memory tiering tasks.