%A Derek L. Eager %A Michael C. Ferris %A Mary K. Vernon %T OPTIMIZING THE DELIVERY OF RADIATION TO CANCER PATIENTS %D August 1998 %R 1385 %I COMPUTER SCIENCES DEPARTMENT, UNIVERSITY OF WISCONSIN %C MADISON, WI %X Systems for on-demand delivery of large, widely-shared data can be expected to have a hierarchical architecture containing one or more (remote) shared central servers and multiple (local) regional servers. This paper develops two significant advances in the state-of-the-art design of such systems. The first advance is an extension to the {\em dynamic skyscraper} delivery technique, recently proposed for single-server video-on-demand systems and shown to have considerable cost/performance advantages over conventional delivery, so that it can be applied in hierarchical systems. The extension allows objects to be partially stored at the regional servers and makes use of a new construct that can also be employed to improve performance in single-server systems. The second advance is an analytic model for determining optimal allocations of objects, within the context of the proposed hierarchical delivery scheme. Results of the model show it is often more cost-effective to store the initial segments of many data objects, rather than the complete data for fewer objects, at the local servers. The results also show how the basic structure of suitable object/segment allocations depends on key system and workload parameters, including the relative constraints of disk bandwidth and storage capacity at the local servers, and the relative cost of local and global channel resources.