Distilling structure in scientific workflows

Motivation and Objectives Scientific workflows management systems, (e.g., (Missier et al., 2010; Ludaesher et al., 2006; Goeck et al. 2011)) are increasingly used to specify and manage bioinformatics experiments. An experiment is then represented by a workflow in which a large number of bioinformatics tasks are linked to each other. A workflow specification is a framework for the execution of workflows. It specifies the order to be observed between the different tasks and their relationships with the workflow inputs and workflow outputs. According to the input data given to the workflow specification and assignments of values to the task parameters, different workflow runs are then obtained and may lead to different intermediate and final output data. Both workflow specifications and runs are represented by graphs. Faced with the increasing complexity of runs and the need for reproducibility of results, provenance has become an important research topic. A significant number of tools for managing vast amounts of data provenance have been designed to assist the storage of provenance data (e.g., indexing), query the data (e.g., difference between executions, search for patterns), visualize the workflow provenance or (re)schedule executions... (See (Cohen-Boulakia and Leser, 2011) for a review on that topic). These tools all make intrinsically complex operations on graph structures (search for subgraphs in a graph, comparing graphs, ...), which, if carried out on Directed Acyclic Graphs (DAGs), with no other restriction of structure, lead to NP-hard problems. Instead, these problems can be solved in polynomial time when specific restrictions are imposed on graphs, such as considering series-parallel (SP) structures (Bein et al., 1992). Some provenance management approaches such as (Bao et al., 2009; Callahan et al., 2006) have therefore chosen to restrict workflow graphs to SP structures. As in general, workflows obtained using workflow systems are DAGs with any structure, graphs transformation approaches such as (Escribano et al., 2009) can be exploited to transform workflow graphs into SP graphs. (Cohen-Boulakia et al, 2012) has recently designed SPFlow, the first algorithm able to rewrite any scientific workflow graph structure into an SP workflow structure while preserving provenance information. As expected, such an approach has a cost in that nodes and/or edges have to be duplicated in the rewritten workflow. Determining the reasons why some workflows have non SP structures may help users to directly design workflows having a structure closer to SP structures. The rewriting process may then be used on less complex, distilled, workflows. The aim of this paper is to present the results obtained on the study that we have conducted on the set of Taverna workflows (Missier et al., 2010) available on myExperiment (De Roure et al, 2009) to analyze the reasons why workflows have non SP structures.