Clustering of RCE Workflow Graphs

Abstract

RCE is an integration environment which allows to create automated workflows orchestrating multi-disciplinary simulation tools in a distributed manner. A workflow consists of components representing tools and connections between these components. The components can be grouped by users within the GUI by creating colored labels. This requires specialist knowledge and is a fully manual task. We investigate the feasibility of automating this task by applying graph clustering methods on such workflows. To this end, we model graphs based on workflows by adopting components as vertices and connections as edges whereby we transfer connection properties to edge weights. We examine three different hierarchical clustering algorithms: edge betweenness, spectral bisection and agglomerative clustering. Additionally, we apply four different metrics to stop the algorithms when a cluster is found: cluster density, global clustering coefficient, average local clustering coefficient and modularity. We examine different mappings of edge weights in combination with the mentioned algorithms and metrics. As groups in workflows have no canonical definition we evaluate our approach qualitatively. We consider 27 results of 1008 parameter combinations as useful. The most expedient approach across multiple workflows is the edge betweenness algorithm with the modularity metric with an undirected graph representation. The scores for the metrics and the mapping vary across workflows and do not enable us to draw general conclusions. We show that our approach is feasible, whereas we remark that a quantitative study is necessary to validate our results in general.