Predictable execution of scientific workflows using advance resource reservations

Scientific Workflows are long-running and data intensive, and may encompass operations provided by multiple physically distributed service providers. The traditional approach to execute such workflows is to employ a single workflow engine which orchestrates the entire execution of a workflow instance, while being mostly agnostic about the state of the infrastructure it operates in (e.g., host or network load). Therefore, such centralized best-effort execution may use resources inefficiently -- for instance, repeatedly shipping large data volumes over slow network connections -- and cannot provide Quality of Service (QoS) guarantees. In particular, independent parallel executions might cause an overload of some resources, resulting in a performance degradation affecting all involved parties.
In order to provide predictable behavior, we propose an approach where resources are managed proactively (i.e., reserved before being used), and where workflow execution is handled by multiple distributed and cooperating workflow engines. This allows to efficiently use the existing resources (for instance, using the most suitable provider for operations, and considering network locality for large data transfers) without overloading them, while at the same time providing predictability -- in terms of resource usage, execution timing, and cost -- for both service providers and customers.
The contributions of this thesis are as follows. First, we present a system model which defines the concepts and operations required to formally represent a system where service providers are aware of the resource requirements of the operations they make available, and where (planned) workflow executions are adapted to the state of the infrastructure.
Second, we describe our prototypical implementation of such a system, where a workflow execution comprises two main phases. In the planning phase, the resources to reserve for an upcoming workflow execution must be determined; this is realized using a Genetic Algorithm. We present conceptual and implementation details of the chromosome layout, and the fitness functions employed to plan executions according to one or more user-defined optimization goals. During the execution phase, the system must ensure that the actual resource usages abide to the reservations made. We present details on how such enforcement can be performed for various resource types.
Third, we describe how these parts work together, and how the entire prototype system is deployed on an infrastructure based on WSDL/SOAP Web Services, UDDI Registries, and Glassfish Application Servers. Finally, we discuss the results of various evaluations, encompassing both the planning and runtime enforcement.