High-Performance Computing: Optimising Workload Mapping and Scheduling Strategies

A review of 66 recent studies on high-performance computing (HPC) scheduling demonstrates the inadequacy of traditional job shop formulations for modelling complex, heterogeneous data centres. HPC scheduling remains computationally intractable – classified as NP-hard – leading to prevalent use of heuristic and meta-heuristic algorithms. Analysis indicates potential for improvement through hybrid optimisation strategies integrating heuristics, machine learning, and novel techniques, tailored to specific workload characteristics, to enhance scalability and efficiency in diverse HPC environments.

The efficient allocation of computational tasks to diverse high-performance computing (HPC) systems represents a persistent challenge in scientific computing. As HPC facilities increasingly adopt heterogeneous architectures – combining CPUs, GPUs, and specialised accelerators – the complexity of optimising workload placement and scheduling intensifies. A systematic investigation of current approaches to this problem is presented by Sharma and Kunkel, of Georg-August-Universität Göttingen, in their review, ‘A Review of Tools and Techniques for Optimization of Workload Mapping and Scheduling in Heterogeneous HPC System’. Their work analyses 66 papers published between 2017 and 2024, identifying limitations in conventional scheduling methods and advocating for integrated optimisation strategies combining heuristic algorithms with machine learning techniques to improve scalability and efficiency.

Mapping and Scheduling Strategies in High-Performance Computing: A Review

This paper presents a systematic review of current strategies for mapping and scheduling workloads within high-performance computing (HPC) environments, with a focus on heterogeneous systems. The authors establish a prototype workflow to define core concepts in workload characterization and resource allocation, then analyse 66 research papers published between 2017 and 2024 to evaluate contemporary tools and techniques. The review confirms that traditional Job Shop scheduling formulations frequently lack the necessary expressiveness to accurately model the complexity of modern HPC data centres, necessitating more adaptable approaches.

The analysis reaffirms the NP-hard nature of HPC scheduling problems, stemming from their combinatorial complexity and the multitude of system and workload constraints, which presents significant challenges for optimization. Consequently, researchers predominantly employ heuristic and meta-heuristic strategies, encompassing nature-inspired algorithms, evolutionary methods, sorting techniques, and search algorithms, to navigate this complexity. These approaches attempt to find acceptable, though not necessarily optimal, solutions within a reasonable timeframe, given the computational intractability of finding the absolute best solution.

The study identifies a gap between current methodologies and the demands of increasingly complex HPC environments, prompting a need for innovative solutions. To address this, the research advocates for hybrid optimisation approaches, strategically combining the strengths of heuristics, meta-heuristics, machine learning, and emerging techniques. Such integration, when specifically tailored to particular problem domains, offers potential for substantial improvements in scalability, efficiency, and adaptability of workload optimisation.

Researchers establish a prototype workflow to define foundational concepts in workload characterisation and resource allocation, providing a common framework for analysing the selected papers. This ensures a consistent evaluation of different approaches and facilitates the identification of key trends and gaps in the field, ultimately driving progress in HPC scheduling. The analysis also reveals a need for more expressive modelling of HPC data centres, acknowledging that traditional scheduling formulations often fail to capture the complexities of modern systems.

The review confirms that traditional Job Shop scheduling formulations frequently lack the necessary expressiveness to accurately model the complexity of modern HPC data centres, hindering effective resource allocation. Researchers consistently employ heuristic and meta-heuristic strategies, including Modified TOPSIS, Simulated Annealing, and various implementations of Genetic Algorithms (EA), to address this limitation. The analysis also incorporates more recent approaches such as Deep Q learning and reinforcement learning algorithms, demonstrating a shift towards more sophisticated techniques.

The study identifies a trend towards hybrid optimisation approaches, integrating diverse methodologies to overcome the limitations of single-method approaches. These methods strategically combine heuristics, meta-heuristics, machine learning, and emerging techniques, offering a pathway to improved scalability, efficiency, and adaptability. Algorithms such as CLARA framework, SMLCM, and PMHEFT exemplify this trend, combining different approaches to enhance performance and optimise resource utilisation.

The review provides a comprehensive overview of the field, demonstrating a clear shift towards more sophisticated and adaptable scheduling strategies. By synthesising the findings from a large body of recent research, the authors present a valuable resource for researchers and practitioners seeking to optimise workload management in complex HPC systems. This synthesis highlights the importance of tailoring scheduling strategies to specific HPC architectures and workloads for maximum effectiveness.

Researchers consistently employ heuristic and meta-heuristic strategies, encompassing nature-inspired algorithms, evolutionary computation, sorting methods, and various search algorithms, to navigate the complexities of HPC scheduling. These approaches attempt to find acceptable, though not necessarily optimal, solutions within a reasonable timeframe, given the computational intractability of finding the absolute best solution. The analysis also incorporates more recent approaches such as Deep Q learning and reinforcement learning algorithms, demonstrating a growing interest in machine learning-based scheduling techniques.