Experimental investigation of memory-related software aging in LLM systems

Large Language Models (LLMs) have been increasingly adopted in a wide range of applications, many of which require long-running inference processes. However, these systems may be subject to software aging phenomena, leading to progressive performance degradation and potential failures. In this work, we experimentally investigate memory-related software aging in LLM inference. We performed 48-hour experiments with three open-source models (Pythia, OPT, and GPT-Neo) under low, medium, and high workloads, monitoring memory consumption at both system and process levels. Using the Mann–Kendall test and Sen’s slope estimator, we observed monotonic growth in RAM usage across all models on Central Processing Units (CPUs), with OPT presenting the steepest slopes. Process-level analysis further revealed that LLM processes were the primary contributors to memory growth, along with background services. Additionally, we conducted identical experiments on Graphics Processing Units (GPUs). Unlike the experiments without a GPU, GPU-based experiments revealed bounded oscillations and abrupt resets likely due to driver-level memory management, while host RAM and process-level monitoring still revealed clear symptoms of aging. These findings demonstrate that software aging manifests differently across execution environments, reinforcing the need for environment-specific monitoring approaches.