Composing Distributed Services for Selection and Retrieval of Event Data in the ATLAS Experiment

TAGs are event-level metadata allowing a quick search for interesting events for further analysis, based on selection criteria defined by the user. They are stored in a file-based format as well as in relational databases. The overall TAG system encompasses a range of web services providing functionality for the required use cases. The data as well as the services are replicated to several ATLAS sites, i.e. inside each service group there exist several concrete deployments, differing only in site-related non-functional attributes. In order to satisfy a user’s request, the above mentioned atomic data sources and web services have to be composed on demand to provide the full functionality. As several instances of each service exist, one service has to be selected out of each group. The overall goal is to maximize the system’s throughput, in order to give to as many users as possible efficient access to the TAGs, while meeting end-to-end quality of service (QoS) requirements. Many approaches can be found to the service selection problem; however most of them are based on a fixed objective function as well as a fixed set of QoS constraints. We argue that in a long lasting experiment like ATLAS, where usage access patterns can change significantly and technologies can evolve, the exact optimization objective can change as well. In this paper, we present a dynamic service selection approach for heterogeneous service environments such as the TAG system, as well as our implementation for monitoring services, deployments and data distribution.