Transparent Resource Adaptivity for Dynamic and Irregular Task-Based Applications on HPC Clusters

• Applicant: Professor Dr. Jonas Posner
• Subject Area: Computer Architecture, Embedded and Massively Parallel Systems
• Term: since 2025
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 558599020

The static resource management of today's supercomputers, where jobs request a fixed number of nodes and keep them for the entire execution, has significant limitations. Resource managers are inflexible in scheduling jobs, and programs cannot adjust their resources to meet changing needs at runtime. This results in reduced performance for both the supercomputer and individual programs, especially for irregular and dynamic workloads that have unpredictable resource needs at runtime. Resource adaptivity can overcome these limitations by allowing jobs to add or remove nodes at runtime, either initiated by the resource manager (malleable) or by the jobs themselves (evolving). Jobs that can do both are called adaptive. However, resource adaptivity is not widely adopted due to a significantly increased programming complexity required to enable it. Additionally, widely used resource managers, such as Slurm, and programming models, such as MPI, offer limited support. Asynchronous Many-Task (AMT) programming is a promising alternative to MPI, enhancing programmer productivity and automatically balancing dynamic and irregular workloads. AMT splits computations into many tasks that are dynamically assigned to computational resources by the runtime system. AMT's transparent resource management is ideal for adaptivity, allowing the runtime system to redistribute tasks and data to enable resource changes without additional programmer effort. However, AMTs with built-in adaptivity are still lacking. This project aims to unlock the potential of adaptivity through AMT. We will develop adaptivity techniques for an existing AMT system with load-balancing capabilities. New programming constructs will allow for expressing explicit resource needs. Heuristics will enable the AMT system to autonomously identify computationally intensive phases to request additional nodes, while proactively releasing nodes during phases of low computational intensity. We will develop algorithms for scheduling a mix of rigid and adaptive jobs, ensuring fairness but incentivizing adaptive jobs. The resource manager will make decisions based on a global system state, balancing overall supercomputer performance and individual program efficiency. To do this, jobs will periodically report their status (impact of node changes). These algorithms will be evaluated with real-world experiments and simulations, using performance models derived from application experiments and supercomputer logs. Finally, we will enable communication between our adaptive AMT system and a prototype resource manager. Experiments with real-world applications will validate the feasibility and benefits of low-overhead resource changes. Methodologically, this project involves designing innovative algorithms, implementing them in prototypes, and evaluating them through experiments and simulations. The knowledge gained could pave the way for adaptivity to significantly improve the efficiency of supercomputers.

• DFG Programme: Research Grants