As the initial stage of the development of an ``end-to-end'' SARA AppLeS, we implemented an AppLeS scheduling agent for the data retrieval phase of the SARA application only (which we term ``Simple SARA''). Figure 2 illustrates the basic structure of the SARA application, with the component to which this work applies shown in bold. In particular, we focused on the transfer of raw data for some region of a single track from the storage site to a processing node. Since some of the data may be stored in multiple archives, we concentrated on the more difficult case where data can be retrieved from multiple servers. (If data can be accessed only from a single server, the SARA tool would of course access it from there). In the multiple data server case, wide performance variations on the networks between the data servers and the processing node have considerable impact on data transfer rates. The selection of which data server to use for the fastest transfer of remote data is the focus of the Simple SARA AppLeS.
 |
For the initial prototype, we assumed that the files to be retrieved
were on disk rather than tape, so access time to the data is uniform
for each of the potential data servers. This is consistent with the
SARA application as it is currently used. The performance model used
for resource selection by the Simple SARA AppLeS is
straightforward:
