Modelling of parameters in supercomputer workloads
Baiyi Song
, Carsten Ernemann
and Ramin Yahyapour
Abstract
Evaluation methods for parallel computers often require the availability of relevant workload information. To this end, workload traces recorded on real installations are frequently used. Alternatively, workload models are applied. However, often not all necessary information are available for a specific workload. In this paper, a model is presented to recover an estimated job execution time when this information is not available. The quality of the modelled estimated runtime is evaluated by comparing different workload traces for which this information is available.
Full Text: PDF