Characterizing Software Aging in GPU-Based LLM Serving Systems

This paper presents a novel methodology to investigate software aging in GPU-based large language model (LLM) serving systems, contrasting with traditional CPU-centric approaches. Through a 216-hour empirical study across six deployments, the authors observed significant memory aging, with leak rates varying based on serving runtime and deployment configurations. This research establishes a reproducible framework for further exploration of software aging and rejuvenation in the context of LLM serving, highlighting the need for practitioners to consider these dynamics in system design and maintenance.