Dynamic Resource Management and Job Scheduling for High Performance Computing = Dynamisches Ressourcenmanagement und Job-Scheduling für das Hochleistungsrechnen

Job scheduling and resource management plays an essential role in high-performance computing. Supercomputing resources are usually managed by a batch system, which is responsible for the effective mapping of jobs onto resources (i.e., compute nodes). From the system perspective, a batch system must ensure high system utilization and throughput, while from the user perspective it must ensure fast response times and fairness when allocating resources across jobs. Parallel jobs can be divided into four categories rigid, moldable, malleable, and evolving. While rigid jobs have fixed resource requirements over their entire life cycle, moldable jobs allow batch systems to deviate from the requested number of resources before job start. In contrast, malleable and evolving jobs can adapt to changing resource allocations at runtime. While batch systems can expand or shrink a malleable job’s resource allocation at any point of time, expanding and shrinking an evolving job occurs only in response to a request made by the application itself. Traditional batch systems support only rigid and moldable jobs, that is, they perform static resource management. However, this is not sufficient as supercomputing enters a new era. Scientific applications are becoming much more complex and now often exhibit unpredictably changing resource requirements. Programming models are also becoming more adaptive in nature to support malleability for energy efficiency and fault tolerance. Therefore, scheduling evolving and malleable jobs (i.e., dynamic resource management) will be indispensable, especially on future large-scale systems. This dissertation therefore proposes novel dynamic resource management and scheduling techniques for cluster systems, making multiple contributions in the areas of dynamic resource (de)allocation mechanisms, efficient adaptive job scheduling, and resiliency. As the first contribution, this thesis presents dynamic scheduling methods for evolving jobs. A fairness scheme is proposed to ensure the fair allocation of resources between static and dynamic resource requests. The evaluation with a workload containing both rigid and evolving jobs shows that high resource utilization and throughput can be achieved, while maintaining the fair dynamic assignment of resources. It is also demonstrated how these methods can be beneficially employed in heterogeneous architectures with network-attached accelerators. The second contribution presents a unique scheduling technique for malleable jobs and an algorithm for the combined scheduling of all four types of jobs in a cluster environment. We introduce the Dependency-based Expand/Shrink (DBES) algorithm, which rests on a two-phase malleable job expand/shrink strategy. The batch system is evaluated with a mixed workload and our strategy achieves consistently superior performance in comparison to state-of-the-art malleable job scheduling strategies. Finally, as the last contribution, we present a scheduling algorithm for dynamic node replacement, which improves the resiliency of cluster systems. The algorithm uses the unique features of the four job types and can provide replacement nodes instantly to jobs affected by node failures. Among current fault tolerance mechanisms, our technique causes the smallest loss of throughput.