Efficient End-Host Resource Management with Kernel Optimizations for Multimedia Applications

Abstract

Multimedia applications have timing requirements that cannot be satisfied using time-sharing scheduling algorithms of general operating systems. Our approach is to provide for a resource reservation mechanism to cater to the real-time resource requirement[5] of multimedia applications. We propose the design of a Resource Manager which allocates and manages the end-host resources among the processes[7]. We identify three important resources at the end-host namely, the processor, memory and system bus cycles. A process reserves these resources by negotiating with the Resource manager.

The goals that we seek to achieve are: (a) real-time resource management using kernel supported reservation mechanisms, (b) optimal utilization of the various resources of the end-system and (c) kernel optimizations for reducing end-host communication overheads in distributed multimedia applications. We use a two-pronged approach to accomplish our goals. First, we adopt a reservation strategy coupled with priority process scheduling13,14 to achieve real-time resource management. The reservation mechanism includes a processor and memory reserve abstraction which controls the allocation of processor cycles and memory space to the processes. The reservation scheme can allow applications to dynamically adapt in real-time based on system load and application requirements. Device requirements of a certain multimedia application is abstracted out as a kernel process for system bus reservation and device activation. Secondly, we adopt kernel optimizations to minimize end-host communication overhead in real-time multimedia applications. To improve the end-host performance in distributed multimedia applications, we unveil a new connectionless protocol, Reliable-UDP, in the kernel.We also present aggressive caching mechanism as a scheme for improving end-host performance. The performance of the Resource Manger was tested out with the generation of processes at random times and the results match the expected theoretical results. The connectionless protocol was tested out in a local distributed system and the results are also presented.