Currently the team is engaged in developing a new version of the system, dubbed SWEB++, which extends the scheduling scheme to include client resources. We model the computation as a sequence of subtasks which are chained together to produce the desired response. These subtasks are then scheduled on an appropriate server node and possibly on the client. This is made possible by the advent of platform-independent code (typically Java), which allows us to transfer computation as well as content. We are also aided by increasing bandwidths between our project and the outside world.
Current client/server solutions typically have a fixed point at which the functionality is divided. In digital libraries, there are circumstances which dictate dynamically scheduling workload due to widely varying client resources and network bandwidths. For example, some users may have a workstation-class machine such as a Pentium Pro connected via a fast network connection to the Internet. In such cases, large amounts of data can be sent to the client for processing. On the other hand, the user may have a Personal Digital Assistant (PDA) utilizing a wireless radio link. Such links typically have a throughput of only 9.6Kbps, forcing every response to use the minimum bandwidth possible. We have analytically modeled this situation and determined that for a "typical" digital library both scenarios should be accounted for.
We have studied the application of this adaptive techniques in wavelet browsing operations, in which the reconstruction computation can be done on either the client or the server. Initial results indicate that by utilizing client resources the server can fulfill over six times more requests per second. We see enormously reduced demands on the server virtual memory system and reduced I/O. We are investigating dynamic document browsing and expect similar results.
During the past year we have tested the system on the ADL machines with good results, which involved porting the code from Solaris to DEC OSF/1, as well as making a number of other minor changes to suit the new configuration. We contributed portions of our Java code to the ADL demonstration suite. Various demonstrations of our software have been presented to our partners, including the Navy (NRAD) and the Alexandria advisory board. At the NRAD visit we were able to demonstrate a reduction of approximately 40% in OS file system overhead through a system reconfiguration utilizing multiple copies of the SWEB software running in parallel.
We are also engaged in rebuilding the system using the Java language and CORBA as a communications substrate. CORBA has been defined as the official access method to CIO collections, and is also being used by the Stanford DLI project, the "InfoBus". We hope to integrate the system with the ADL experimental system over the course of the year.