CDES: An approach to HPC Workload Modelling

Computational science and complex system administration relies on being able to model user interactions. When it comes to managing HPC, HTC and grid systems user workloads - their job submission behaviour, is an important metric when designing systems or scheduling algorithms. Most simulators are either inflexible or tied in to proprietary scheduling systems. For system administrators being able to model how a scheduling algorithm behaves or how modifying system configurations can affect the job completion rates is critical. Within computer science research many algorithms are presented with no real description or verification of behaviour. In this paper we are presenting the Cluster Discrete Event Simulator (CDES) as an strong candidate for HPC workload simulation. Built around an open framework, CDES can take system definitions, multi-platform real usage logs and can be interfaced with any scheduling algorithm through the use of an API. CDES has been tested against 3 years of usage logs from a production level HPC system and verified to a greater than 95% accuracy.