In large-scale application scenarios, performance is a critical issue. The execution of enterprise-wide and cross-organizational workflows may generate a very high load, which does not only affect workflow servers but also the underlying communication network. Due to the high amount of communication between workflow server(s) and  clients, the network may become a bottleneck, especially if a large amount of "long-distance" communication occurs. In order to avoid such bottlenecks we aim to reduce the network load by partitioning workflow schemes and by migrating the control of corresponding workflow instances from one server to another during run-time if favorable; i.e., a workflow instance may no longer be controlled by only one workflow server. When performing such a migration, at least a description of the instance state (including workflow relevant data) and of previously applied instance changes has to be transmitted to the target server. In order to avoid unnecessary, subnet-spanning communication between servers, however, parallel branches of a workflow instance should be controlled independently from each other – at least as no synchronization due to other reasons (e.g., a dynamic workflow change) becomes necessary.

The overall project goal was to develop a comprehensive approach for achieving scalability in (distributed) workflow management systems. In particular, we address the following challenges: 

• The load caused by the execution of a large number of workflow instances must be adequately distributed to servers such that a single server does not become a bottleneck.

• The assignment of servers to single activities (i.e., the partitioning of the respective workflow schema) must be done in a way that minimizes the overall communication load.

• Necessary migrations between workflow servers must be efficiently performed; i.e., they should only require a minimal amount of data to be transmitted.

• The assignment of workflow servers to single activities must be done in a flexible way. For example, to achieve high locality between the workflow server controlling a particular activity and the potential actors working on it, variable server assignments must be supported.

• Even for workflow instances with partitioned schema and distributed control the support of dynamic changes is indispensable. We aim at an approach that provides the full change functionality as in the central case while, at the same time, achieving an extremely favorable behavior with respect to communication costs.

The developed concepts were implemented in the ADEPT1 workflow management systems. Tool presentations and demonstrations on distributed workflow management were given at several conferences and industrial fairs, including EDBT’2000, BIS’2000, and CeBIT’2000.