The Simple SARA AppLeS demonstrates that application-level scheduling is a useful strategy for data retrieval. We are currently extending the Simple SARA AppLeS to provide performance for an ``end-to-end'' SARA tool. Figure 2 illustrates the application architecture being used to develop the SARA AppLeS agent. In particular, our ``end-to-end'' SARA AppLeS focuses on applications which require multiple files per image, and may process the image at a variety of different sites (at the data server, at an intermediate compute server, at the web site, etc.). The Simple SARA resource selection strategy will provide a basic building block for an ``end-to-end'' SARA AppLeS.
Note that the performance model will be more complex for the ``end-to-end'' SARA. In particular, the interactions of the client and the compute server(s), the compute servers and data servers, as well as data servers of various sorts of media will all have to be modeled. Scheduling the ``end-to-end'' application will involve decisions about which compute server(s) to use, whether to move the computation to the data, the data to the compute server, or to compute at an intermediate site, and which of the different data servers to target. Optimizing each of these activities individually may not necessarily optimize the execution performance of the ``end-to-end'' application, so the application must be scheduled as whole.
We also note that SARA is representative of a larger class of distributed data applications. High energy physics applications from the Cleo/Nile project [10], Digital Sky [7], San Diego Supercomputer Center's Storage Resource Broker (SRB) [12], and other applications have a similar computational structure to SARA and similar computational goals, although the size of the data files, location of client, servers and visualization sites, computation to be performed on the data, and other details may differ. We plan to use the ``end-to-end'' SARA AppLeS as a model for the development of a distributed data AppLeS template which can be used for scheduling other distributed data applications on the Grid.