The Semantics of a Resource Request: Moving Beyond Cores and Gigabytes

Abstract

This article challenges traditional resource request models in cluster computing that rely on abstract CPU core and memory specifications, particularly for modern AI/ML workloads. Current coarse-grained abstractions mask essential hardware characteristics and workload requirements, causing performance unpredictability, resource inefficiency, and service level violations. By examining the limitations of current models, the paper exposes performance non-fungibility issues and contention on unmanaged resources like memory bandwidth and cache hierarchies. A new multi-dimensional taxonomy of resource semantics encompasses hardware attributes, workload behaviors, and operational policies. Through analysis of schedulers from HPC, hyperscale, and open-source domains, the article shows how richer semantics are gradually being adopted. The research addresses opportunities and challenges of this semantic shift and explores future directions, including standardized Resource Description Languages and the application of Large Language Models for intelligent scheduling. The article advocates for semantically-aware schedulers capable of intelligent, topology-aware resource matching to enhance performance and efficiency in large-scale systems.