Architectures for very large scale workflow management systems

Workflow management systems are distributed computer-based control systems designed to help orchestrate the execution of business processes. As organizations transfer more and more of their paperwork to electronic systems, workflow systems must grow to accommodate much larger numbers of users, workcases, information items, and complexity in general. Workflow problems are surfacing today because traditional workflow products are not able to handle the system growth that is needed as organizations expand rapidly into electronic commerce. Almost all current workflow products are based upon the server-client architecture in which control and data for business processing are centralized in a workflow server machine. This server becomes a bottleneck as organizations attempt to scale up. Very little attention has been given to this scalability problem in the literature of workflow research and development. It is frequently assumed that scalability can be easily obtained by replication of the workflow server.;This thesis proposes a new framework and taxonomy of workflow architectures that allows comparative scalability analysis of various existing and new architectures. A useful formal abstraction is defined that connects various workflow models, such as the information control net, to various workflow implementation architectures. I introduce a family of scalable architectures called active-model workflow management systems (AM workflow). The AM workflow architectures are categorized into class-active workflow and instance-active workflow according to the patterns of process structures.;In the latter part of the thesis, we analyze the AM workflow architecture via LQM (layered queuing model) analysis using the modified MOL (method of layers) algorithm for closed queuing network models. Our basic hypothesis is that AM workflow is significantly more scalable and workable than traditional server-client workflow in environments that must process thousands or millions of workcases simultaneously. The analyses and the data that support this hypothesis are explained in the thesis. I conclude that the server-client architecture is woefully inadequate for future very large scale workflow. I find that this result also holds for server-client architectures that attempt to scale up by replicating the workflow server. Along the way, I point out that there are a number of very promising hybrid workflow architectures. I note that among the hundreds of workflow products on the market, none of them seem to implement AM workflow or its hybrids. I hope to see these types of architectures emerging in products of the future; successful utilization will be the ultimate validation of our thesis hypothesis.