The Great Divide: Old School vs. New School in HPC Scheduling

This presentation explores the fundamental philosophical and technical divide between traditional HPC scheduling approaches and emerging cloud-native methodologies in high-performance computing. The "old school" philosophy centers on finite, managed capacity with maximum utilization in static environments, while the "new school" embraces elastic, virtually infinite capacity with dynamic scaling and cost-per-cycle optimization. As traditional HPC users resist cloud-native tools and cloud users struggle with HPC workload nuances, this cultural and technical gap creates significant challenges for organizations seeking to bridge on-premises and cloud deployments. The talk examines how new hybrid schedulers and approaches — like the many Slurm-on-Kubernetes solutions — are emerging to address these divides, while cloud-native projects evolve to support topology awareness, multi-cluster operations, and AI/ML training workloads. Beyond scheduling, the presentation addresses the critical "data gravity problem," highlighting how data proximity remains vital even as cloud scalability improves, leading to hybrid-cloud deployments that balance data locality with computational flexibility. The session concludes by discussing emerging trends toward unified scheduling standards across on-premises and cloud environments, and increasing collaboration between traditional HPC and cloud-native scheduling ecosystems.