Workflow management systems (WFMSs) that are geared for the orchestration of enterprise-wide or even "virtual-enterprise"-style business processes across multiple organizations are complex distributed systems. They consist of multiple workflow engines, application servers, and ORB-style communication servers. Thus, deriving a suitable configuration of an entire distributed WFMS for a given application workload is a difficult task. This paper presents a mathematically based method for configuring a distributed WFMS such that the application's demands regarding performance and availability can be met while aiming to minimize the total system costs. The major degree of freedom that the configuration method considers is the replication of the underlying software components, workflow engines and application servers of different types as well as the communication server, on multiple computers for load partitioning and enhanced availability. The mathematical core of the method consists of Markov-chain models, derived from the application's workflow specifications, that allow assessing the overall system's performance, availability, and also its performability in the degraded mode when some server replicas are offline, for given degrees of replication. By iterating over the space of feasible system configurations and assessing the quality of candidate configurations, the developed method determines a configuration with near-minimum costs.

Workflow management systems (WFMS) that are geared for the orchestration of business processes across multiple organizations are complex distributed systems: they consist of multiple workflow engines, application servers, and communication middleware servers such as ORBs, where each of these server types can be replicated on multiple computers for scalability and availability. Finding an appropriate system configuration with guaranteed application-specific quality of service in terms of throughput, response time, and tolerable downtime is a major challenge for human system administrators. This paper presents a tool that largely automates the task of configuring a distributed WFMS. Based on a suite of mathematical models, the tool derives the necessary degrees of replication for the various server types in order to meet specified goals for performance and availability as well as "performability" when service is degraded due to outages of individual servers. The paper describes the configuration tool, with emphasis on how to capture the load behavior of workflows in a realistic manner. We also present extensive experiments that evaluate the accuracy of the tool's underlying models and demonstrate the practical feasibility of automating the task of configuring a distributed WFMS. The experiments use a detailed simulation which in turn has been validated through measurements with the Mentor-lite prototype system.

Workflow management systems (WFMS) are a cornerstone of mission-criticial, possibly cross-organizational business processes. For largescale applications both their performance and availability are crucial factors, and the system needs to be properly configured to meet the application demands. Despite ample work on scalable system architectures for workflow management, the literature has neglected the important issues of how to systematically measure the performance of a given system configuration and how to determine viable configurations without resorting to expensive trial-anderror or guesswork. This paper proposes a synthetic benchmark for workflow management systems; based on the TPC-C order-entry benchmark, a complete e-commerce workflow is specified in a system-independent form. This workflow benchmark, which stresses all major components of a workflow system and is parameterized in a flexible manner, has been applied to two operational systems, the commercial system Staffware97 and our own prototype system Mentor-lite. The paper reports performance results from our measurements and discusses lessons learned. In particular, the results demonstrate the scalability of the Mentor-lite architecture. The measurements also underline the need for configuring systems intelligently, and the paper outlines an auto-configuration tool that we have been building to this end. 1 Introduction 1.1 Problem Statement Workflow technology has penetrated into mission-critical, enterprise-wide or even crossorganizational, business applications. Typical examples are insurance claim processing, cargo shipping, or healt-care tracking and planning, and workflow technology is also embedded in many e-commerce services. Following the terminology of WfMC [32] (see also [5, 6, 8, 15, 18]), a workflow is a set of activities that belong together in order to achieve a certain business goal. Activities can be completely automated or based on interaction with a human user and intellectual decision-making. In particular, an activity can spawn requests to an arbitrary "invoked application" that is provided by some server independently of the current workflow. Workflow management systems (WFMS) orchestrate the control and data flow between a workflow's activities, based on a high-level specification of the intended behavior (e.g., using Petri-net variants, state charts, or some script language) with some leeway for exception handling and run-time improvisation (as needed, e.g., in medical applications). Despite their business success, most WFMS products exhibit specific idiosyncracies and, by and large, significant deficiencies and limitations in terms of their performance. The current situation is probably comparable to that of relational database systems in the eighties. Also and similarly to database technology, configuring and tuning a WFMS for satisfactory performance falls more in the realm of black art (i.e., guesswork or expensive trial-and-error experimentation) and sorely lacks scientific foundations. Even such mundane basics such as sy...

Workflow management systems (WFMS) that are geared for the orchestration of business processes across multiple organizations are complex distributed systems: they consist of multiple workflow engines, application servers, and communication middleware servers such as ORBs, where each of these server types can be replicated on multiple computers for scalability and availability. Finding an appropriate system configuration with guaranteed application-specific quality of service in terms of throughput, response time, and tolerable downtime is a major challenge for human system administrators. This paper presents a tool that largely automates the task of configuring a distributed WFMS. Based on a suite of mathematical models, the tool derives the necessary degrees of replication for the various server types in order to meet specified goals for performance and availability as well as "performability" when service is degraded due to outages of individual servers. The paper describes the configuration tool, with emphasis on how to capture the load behavior of workflows in a realistic manner. We also present extensive experiments that evaluate the accuracy of the tool's underlying models and demonstrate the practical feasibility of automating the task of configuring a distributed WFMS. The experiments use a detailed simulation which in turn has been validated through measurements with the Mentor-lite prototype system.

Enterprise-spanning workflows require workflow management systems that can be tailored to specific application needs, as well as enhanced support for interoperability between different workflow management systems. In virtual enterprises, the interoperability problem is not limited to workflow execution, but also entails facilities like worklist management and history management to be interoperable. We present a lightweight system architecture, consisting of a small system kernel on top of which extensions like history management and worklist management are implemented as workflows themselves. The functionality of the kernel such as distributed workflow execution and interoperability interfaces is available for all extensions. We show the feasibility of our approach by presenting the implementation of history management in our workflow specification language, based on state and activity charts, on top of our lightweight kernel, coined Mentor-lite.