Probabilistic Job History Conversion and Performance Model Generation for Malleable Scheduling Simulations

Malleability support in supercomputing requires several updates to system software stacks. In addition to this, updates to applications, libraries and the runtime systems of distributed memory programming models are also necessary. Because of this, there are relatively few applications that have been extended or developed with malleability support. As a consequence, there are no job histories from production systems that include sufficient malleable job submissions for scheduling research. In this paper, we propose a solution: a probabilistic job history conversion. This conversion allows us to evaluate malleable scheduling heuristics via simulations based on existing job histories. Based on a configurable probability, job arrivals are converted into malleable versions, and assigned a malleable performance model. This model is used by the simulator to evaluate its changes at runtime, as an effect of malleable operations being applied to it.