Runtime Adaptation of Scientific Service Workflows
Software landscapes are rather subject to change than being complete after having been built. Changes may be caused by a modified customer behavior, the shift to new hardware resources, or otherwise changed requirements. In such situations, several challenges arise. New architectural models have to...
Mathematik und Informatik
|Online Access:||PDF Full Text|
No Tags, Be the first to tag this record!
|Summary:||Software landscapes are rather subject to change than being complete after having been built. Changes may be caused by a modified customer behavior, the shift to new hardware resources, or otherwise changed requirements. In such situations, several challenges arise. New architectural models have to be designed and implemented, existing software has to be integrated, and, finally, the new software has to be deployed, monitored, and, where appropriate, optimized during runtime under realistic usage scenarios. All of these situations often demand manual intervention, which causes them to be error-prone.
This thesis addresses these types of runtime adaptation. Based on service-oriented architectures, an environment is developed that enables the integration of existing software (i.e., the wrapping of legacy software as web services). A workflow modeling tool that aims at an easy-to-use approach by separating the role of the workflow expert and the role of the domain expert. After the development of workflows, tools that observe the executing infrastructure and perform automatic scale-in and scale-out operations are presented. Infrastructure-as-a-Service providers are used to scale the infrastructure in a transparent and cost-efficient way. The deployment of necessary middleware tools is automatically done.
The use of a distributed infrastructure can lead to communication problems. In order to keep workflows robust, these exceptional cases need to treated. But, in this way, the process logic of a workflow gets mixed up and bloated with infrastructural details, which yields an increase in its complexity. In this work, a module is presented that can deal automatically with infrastructural faults and that thereby allows to keep the separation of these two layers.
When services or their components are hosted in a distributed environment, some requirements need to be addressed at each service separately. Although techniques as object-oriented programming or the usage of design patterns like the interceptor pattern ease the adaptation of service behavior or structures. Still, these methods require to modify the configuration or the implementation of each individual service. On the other side, aspect-oriented programming allows to weave functionality into existing code even without having its source. Since the functionality needs to be woven into the code, it depends on the specific implementation. In a service-oriented architecture, where the implementation of a service is unknown, this approach clearly has its limitations. The request/response aspects presented in this thesis overcome this obstacle and provide a SOA-compliant and new methods to weave functionality into the communication layer of web services.
The main contributions of this thesis are the following:
Shifting towards a service-oriented architecture: The generic and extensible Legacy Code Description Language and the corresponding framework allow to wrap existing software, e.g., as web services, which afterwards can be composed into a workflow by SimpleBPEL without overburdening the domain expert with technical details that are indeed handled by a workflow expert.
Runtime adaption: Based on the standardized Business Process Execution Language an automatic scheduling approach is presented that monitors all used resources and is able to automatically provision new machines in case a scale-out becomes necessary. If the resource's load drops, e.g., because of less workflow executions, a scale-in is also automatically performed. The scheduling algorithm takes the data transfer between the services into account in order to prevent scheduling allocations that eventually increase the workflow's makespan due to unnecessary or disadvantageous data transfers. Furthermore, a multi-objective scheduling algorithm that is based on a genetic algorithm is able to additionally consider cost, in a way that a user can define her own preferences rising from optimized execution times of a workflow and minimized costs. Possible communication errors are automatically detected and, according to certain constraints, corrected.
Adaptation of communication: The presented request/response aspects allow to weave functionality into the communication of web services. By defining a pointcut language that only relies on the exchanged documents, the implementation of services must neither be known nor be available. The weaving process itself is modeled using web services. In this way, the concept of request/response aspects is naturally embedded into a service-oriented architecture.|