1 Architectures of Multimedia Communication Subsystems Erich Rütsche Institute Eurecom, Sophia-Antipolis ruetsche@eurecom.fr Matthias Kaiserswerth IBM Research Division, Zurich Research Laboratory kai@zurich.ibm.com Abstract This paper analyzes the particular requirements that multimedia communication imposes on the network adapter and the I/O subsystem of a workstation. We show the drawbacks of current (parallel) communication subsystems and develop new architectural concepts applicable to multimedia communication subsystems. The key ideas are the separation of isochronous and asynchronous traffic, parallelism between sender and receiver, and the execution of per-byte operations in a hardware pipeline. We adapt these concepts to the design of a Gb/s adapter and to the design of a light-weight, slower speed adapter and present some scenarios for their application. 1. Introduction New types of networks such as FDDI, ATM, and FCS provide bandwidth in the range of 100 Mb/s to 1 Gb/s to a single workstation. Multimedia communication requires broadband communication of data, still images at high resolution, and moving pictures (video, animated graphics). The aggregation of these data generates not only high-volume continuous data streams but also dynamic bursts of data. An example of this is cooperative teleworking with video connections between multiple parties who also exchange high-resolution images, e.g., doctors holding a teleconference discussing X-ray images. The communication of these data imposes a new type of load on the network, on the network adapter, and on the I/O subsystem of the workstation. Processing requirements of multimedia protocols are very different from the requirements of traditional transport protocols. Isochronous multimedia traffic requires the transmission of a large amount of data with low delay and bounded delay jitter but may accept relatively high bit-error or packet loss rates [Hehmann 89]. In a video, for example, bit errors or even lost frames are hardly visible, whereas a delay jitter of more than 10 ms is perceived as disturbing. Asynchronous traffic, for example file transfer or a remote procedure call, requires comparatively less throughput but tolerates no errors. In the conference example above, the video connection between the doctors could tolerate errors, whereas the X-ray images must be displayed error-free. Conventional protocol processing on a workstation is limited by the available processing power and the I/O bottleneck [Ramakrishnan 93]. Communication subsystems based on single processors such as Nectar [Steenkiste 92a] or based on multiprocessors, such as the Parallel Protocol Engine (PPE) [Giarrizzo 89, Kaiserswerth 93] and similar architectures proposed by [Zitterbart 89], [Braun 92], and [Ulrich 93] were built to off-load protocol processing from the workstation. These systems were successful for transport protocol stacks at a network speed of up to 150 Mb/s. However, these approaches do not fulfill the quality of service requirements of real-time multimedia data streams transported in higher speed networks. A successful multiprocessor subsystem for continuous multimedia at moderate speed was demonstrated in [Blair 93]. In this paper we analyze the new requirements that the transmission of multimedia data impose