.addition, when the system is overloaded with continuous media applications, a way of identifying applications of lesser or greater importance to the users can allow the system to automatically perform service trade-offs rather than forcing it to degrade all applications equally at best, or randomly at worst. Armed with such information, the system can manage its resources in such a way as to maximize the total value delivered to the end user. Towards this end, we are creating a new scheduling framework, based on Time-Driven Resource Management [6,8, 9], that provides the flexible control and delivered performance required for multimedia applications.Finally, note that the existence of the strict-priority realtime scheduling class in standard SVR4 in no way allows a user to effectively deal with these types of problems. In addition, it opens the very real possibility of runaway applications that consume all CPU resources and effectively prevent a user or system administrator from regaining control without rebooting the system. AcknowledgmentsMonica Lam provided many insightful suggestions, especially during the formative stages of this work. This research was supported in part by an NSF Young Investigator Award and Sun Microsystems Laboratories, Inc.‡ UNIX is a trademark of UNIX System Laboratories. References1 New Timesharing ClassA new timesharing scheduling class was developed in order to correct the problems demonstrated in these experimental runs. In particular, the modified version removes the anomalies of identifying batch jobs as interactive, and vice versa. In addition, it attempts to ensure that each process that can run is given the opportunity to make steady progress in its computation, while retaining a bias in favor of interactive processes. Finally, it reduces the feedback interval over which CPU behavior is monitored and penalties and rewards given. The timesharing scheduling class contained in Sun's Solaris 2.3 operating system is based on this work.The results of the default use of this class for all applications and the window server process are given by Figure 2i. As can be seen, this delivers significantly better results for the continuous media and interactive applications than any combination of the standard SVR4 scheduling classes. It should also be noted that this scheduling policy achieves this level of performance without significantly starving the batch application, which still receives approximately 30% of the available CPU time.Additional tests were performed by adjusting user priorities and by combining this new scheduling class with the SVR4 RT class (as was done with SVR4 TS class). However, with the exception of the cases where there was sufficient load in the RT class to consume all CPU cycles and starve the new scheduling class, this resulted in no pathologies and showed a direct and predictable relationship between user priorities and application performance. Conclusions and Future WorkThrough trial and error, it may be possible to find a particular combination of priorities...

The first workshop session was dedicated to the issue of real-time support for multimedia applications. Audio and video are commonly referred to as time-dependent continuous media. Their timing dependencies need proper support by the computer system. This calls for the use of classical techniques from real-time computing. Yet, audio and video have slightly different requirements than traditional real-time applications. During the first talk, Jim Hanko from SUN Microsystems presented a paper co: authored with Eugene Kuerner, Duane Northcutt, and Gerald Wall on "Workstation Support for Time-Critical Applications," introducing the topic. The initial thesis of the talk was that while today's workstations have a great deal of computational power, this power cannot be effectively delivered to support multimedia applications because system resources are not organized and managed in the necessary manner. Existing workstations provide abundant CPU capacity, but poor I/O support. While in traditional computing the lack of I/O power can be masked by caching, continuous-media applications that do not reuse data cannot benefit from it. Workstations do not dedicate resources to I/O processing as mainframes do. The single CPU has to help out to perform I/O, constituting a performance bottleneck. That the CPU is scheduled based on "fairness" rather than urgency worsens the problem. The authors point out that existing real-time systems cannot be used to overcome this problem because they rely on deterministic behavior and known load, assumptions not well suited for a workstation environment. Their statement that existing real-time scheduling techniques such as rate-monotonic scheduling are not appropriate for multimedia has, however, generated opposition from the audience. A new resource management technique, Time-Driven Resource Management (TDRM), is proposed. The technique bases its decisions on the requester's deadline, importance, and expected resource requirements. The technique does not provide sharp guarantees for a certain quality of service, it rather encourages "graceful degradation." The matter of how a resource management technique should affect system behavior was subject to major discussion. Many reservation schemes proposed (e.g., those from Anderson or Ferrari of the University of California at Berkeley) provide a fixed quality of service and reject new service requests when the provision of the service would endanger previously given guarantees. To a user, this may mean that his request for retrieving a video is turned down. Such a system can be compared to the telephone system: If a busy signal is received, the user will have to wait and dial again. Adaptive policies provide more flexibility, but cannot guarantee f'bxed quality levels. It was mainly agreed that an ideal resource management technique should be a combination of both cases. Some fictitious "costs" can be chosen to decide which ap

Session IX continued the presentation and discussion of multimedia research projects from Session VI. It showed a very diverse set of approaches, but integration of new media in the existing workstation environment was a common theme to all of them. Only some fraction of the session can be captured in this summary because three of the four speakers made use of video presentations (Andy Hopper did not even have a paper) or even plugged their PC into the projector (Klaus Meissner).

No abstract

No abstract