Workflow skeletons: Improving scientific workflow execution through service migration

Planning the execution of a long-running scientific workflow orchestrating a huge number of services in a cloud infrastructure is a hard thing to do, because anticipating the current infrastructure situation at a given point in time is far from easy. This paper describes a means for adjusting ""workflow skeletons""-based scientific applications to this dynamically changing situation by using service migration allowing for moving services to different hosts, even if they already started execution. From a global workflow perspective, there may be several reasons for deciding at runtime to move services to other hosts than originally planned. For example, sometimes it is not easy to predict in which phase of the workflow big data will be produced. In such situations, moving services to the data's location instead of staging data to the services will save time and network bandwidth. The concept of ""workflow skeleton"" is enhanced by letting single tasks decide about starting a migration based on a predefined set of policies.