Computer-aided engineering design uses simulations to explore a design space and identify promising regions. The hierarchical structure of the engineering design process suggests distinctive workload patterns that we believe are common in industry, yet have been little captured by previous characterisations. Selecting a scheduling policy that is ill-suited to the workload it serves may lead to poor performance. This paper characterises the workload run on the private grid of a large aircraft manufacturer over a period of 30 months. Cycles of load at daily and weekly scales are observed, and there are extended periods where the grid operates at saturation and work must queue. A method of creating workloads containing such cycles at a given total load percentage by adjusting inter-arrival times is given. Task execution times are demonstrated to be distributed in a log-uniform way, and an algorithm is given to generate execution times following this distribution. Graphs of dependencies between tasks are shown to have variation in node degree greater than that of random graphs, and a process of constructing dependency graphs following the observed distribution is described.

BibTex Entry

@inproceedings{Burkimsher2014,
 author = {A. Burkimsher and l. Bate and L. S. Indrusiak},
 booktitle = {22nd High Performance Computing Symposium (HPC 2014)},
 organization = {Society for Modeling and Simulation International (SCS)},
 pages = {639--646},
 title = {A Characterisation of the Workload on an Engineering Design Grid},
 year = {2014}
}