Abstract

Container shipping for interdomain transfers involves no physical copying. Its cost is proportional to the amount of data actually accessed, often a fraction of the data transferred. ew I/O devices with data rates ranging from 10 to 100 Mbytes per second are becoming available for personal computers and workstations. These include human-interaction devices for video capture and display (and audio record and playback), high-capacity storage devices, and high-speed net-work communication devices. Along with continual improvements in processor, memory, and bus technology, these devices have enabled I/O-intensive applica-tions for desktop computing that require input, processing, and output of very large amounts of data. In this article, we focus on an important aspect of operating system support for these applications: efficient transfer of large data objects be-tween the protection domains in which processes and devices reside. A rapidly growing class of IIO-intensive applications is multimedia computing, in particular, applications that acquire or present video (or image) and audio streams, possibly transforming them in novel ways under programmer control. These applications are often distributed and interactive, imposing real-time con-straints for the delivery of large volumes of data transported over potentially long distances. Examples include video teleconferencing with shared work spaces, re-mote scientific visualization and sonification, and distributed virtual reality. Multimedia computing applications often require the manipulation of images whose sizes typically range from 1 to 10 Mbytes, but they can be larger. For exam-ple, an uncompressed high-definition television frame requires 6 Mbytes, and a high-quality computer-generated video frame from a motion picture such as Jurus-sic Park requires 36 Mbytes. While compression can sometimes reduce these sizes by one to two orders of magnitude, usually processing must be done on uncom-pressed data. Scientific computing and visualization are even more demanding. These applications often operate on data in the form of sequences of very large

Citations