Operating System Techniques for Distributed Multimedia

In designing operating system support for distributed multimedia, we target three areas for improvement: reduced copying, reduced reliance on explicit kernel-user interactions, and provision of rate-based flow control. Towards these goals, we propose an ar

Operating System Techniques for Distributed MultimediaDavid K.Y. Yau and Simon S. Lam Department of Computer Sciences The University of Texas at Austin Austin, Texas 78712-1188 TR-95-36 July 17, 1995 January 2, 1996 (revised)Abstract

In designing operating system support for distributed multimedia, we target three areas for improvement: reduced copying, reduced reliance on explicit kernel-user interactions, and provision of rate-based ow control. Towards these goals, we propose an architecture that includes the concept of I/O e cient bu ers for reduced copying, the concept of fast system calls for low latency network access, and the concept of kernel threads for ow control. Also included is a concept called direct media streaming which is suitable for applications that require limited user processing of media data. These concepts have been implemented as an extension to SunOS 5.3 (the operating system component of Solaris 2.3). We report some experimental results on the performance of our current system.

Research supported in part by National Science Foundation under grant no. NCR-9506048 and by a grant from AT&T Foundation. An earlier version of this paper appeared in Proceedings IS&T/SPIE Multimedia Computing and Networking (MMCN '96), January 1996.

In designing operating system support for distributed multimedia, we target three areas for improvement: reduced copying, reduced reliance on explicit kernel-user interactions, and provision of rate-based flow control. Towards these goals, we propose an ar

Data buffers

In designing operating system support for distributed multimedia, we target three areas for improvement: reduced copying, reduced reliance on explicit kernel-user interactions, and provision of rate-based flow control. Towards these goals, we propose an ar

Memory-to-memory copying incurs a per-byte overhead. Such overhead is substantial for distributed multimedia applications which require bulk data transfer. For example, an MPEG-2 source, targeted at 4{9 Mbps for broadcast TV quality video, would provide about 1 Mbyte of data per second. And as network bandwidth increases, it is approaching the same order of magnitude as memory bandwidth. Therefore, memory bandwidth may become the communication bottleneck if memory copies are not carefully minimized 5]. Second, data ows generated by multimedia applications have characteristics that are more predictable than conventional data tra c. In a server for video playback, for example, pictures are made available for sending once every frame period. To transfer control between kernel and user for such isochronous tra c, it would be unnecessary to use explicit system calls. There are two kinds of system calls in Unix: synchronous and asynchronous. A synchronous call blocks until its operation can be performed. For example, a read that nds an empty connection blocks until the connection has packets. An asynchronous call that does not succeed immediately returns with an indication that the user process should try again later. The retry may be facilitated by a signal delivered to the user process when the blocking condition is lifted. In either case, an I/O operation succeeds only when the system call successfully returns. To better understand the characteristics of multimedia applications, we studied a working video-conferencing tool called nv (for network video), which is widely used in mbone multicasts. We found that nv sends each picture as a sequence of smaller bu ers of about 1040 bytes. Each bu er corresponds to a packet of RTP, which is an application level protocol for multimedia transport over the Internet. Figure 2 pro les nv's use of the send() function call for sending to the network. This pro le shows that nv makes fairly frequent use of system calls for data transfer in a periodic manner. Third, the delivery of media data generally requires real-time QoS (such as delay and throughput) guarantees to ensure smooth playback. We therefore expect that future multimedia applications will negotiate with networks for reserved-rate connections|such as those considered in 2] and 8]. To obtain QoS guarantees, the packet arrivals to a reserved-rate connection are required to satisfy certain ow speci cations. Operating system support for implementing rate-based ow control is essential. The balance of this paper is organized as follows. In Section 2, we present an overview of operating system techniques in our proposed architecture designed to support distributed multimedia applications. In Section 3, we elaborate upon I/O e cient bu ers for reducing the need for memory-to-memory data copy. In Section 4, we introduce a fast system call for low latency network access. In S

ection 5, we describe kernel threads intended for reducing the use of system calls and for rate-based ow control. In Section 6, direct media streaming is introduced for applications that are mainly concerned with transport|rather than processing|of media data. In Section 7, we present performance measurements of primitive operations in the architecture. In Section 8, we illustrate how to apply the primitive operations for reducing the use of system calls and for implementing rate-based ow control in a distributed multmedia system.

In designing operating system support for distributed multimedia, we target three areas for improvement: reduced copying, reduced reliance on explicit kernel-user interactions, and provision of rate-based flow control. Towards these goals, we propose an ar

1200

1000

800

600

400

200

0Size (bytes)0200400600800100012001400

Time (ms)

In designing o …… 此处隐藏：20179字，全部文档内容请下载后查看。喜欢就下载吧 ……