Name: Rose P. Tsang Title: Senior Member of Technical Staff Affiliation: Sandia National Laboratories Postal address: PO BOX 969, MailStop 9011 Livermore, CA 94550 Email address: rtsang@ca.sandia.gov Phone: 510-294-3200 Fax: 510-294-1225 Signalling Mechanisms for Supporting Dynamic QoS for Large Heterogenous IP Networks Rose P. Tsang rtsang@ca.sandia.gov Sandia National Laboratories Livermore, California BACKGROUND. Distributed multimedia applications share several important features which represent a major departure from conventional computer data applications. The first is that they usually require real-time transmission of continuous media information such as digital video and audio. Digital video, of any substantial quality, is particularly demanding due to its dynamic high bit rate variability as well as real-time requirements [TSANG96,TSANGA97]. In addition to supporting real-time traffic, the network must also support the integration of traditional computer data traffic while maintaining the integrity, or Quality of Service (QoS), of each traffic type. Many distributed multimedia applications are also distribution oriented. In such applications, a sender may desire to multicast data streams to a large dispersed group of users. An example of such an application would be video- on-demand or remote videoconferencing. The group of users may be dynamic; users may leave and enter the group on demand. The users or receivers in such environments may also have varying QoS requirements. This may be based upon their system capabilities as well as their personal preferences. For instance, some users, with low-end equipment may only be capable of receiving a lower-resolution NTSC MPEG-2 signal. A network which sent an entire signal, containing high and low resolution components, would be wasting network resources by sending to such a low- end receiver [TSANGA97]. Thus another important feature the network must provide is multicasting and broadcasting capabilities in a heterogenous QoS environment. CURRENT STATE. Today's IP networks offer a best effort datagram service [CHEN97]. IP networks based upon IP version 6 will provide two fields, a priority field and a flow label field, where a user can request a real-time level of service. However, the actual mechanism which will allow the desired real-time capabilites to be actualized must be found in the underlying network. Real-time service support for continuous media traffic entails the notion of resource reservation based upon some form of connection management [TSANGA96]. The Resource ReSerVation Protocol (RSVP) was intended to provide IP networks with real-time support capabilities. RSVP allows heterogenous reservations within a multicast session, receiver initiated reservations, "soft-state" in routers, dynamic QoS requests, and multiple reservation styles. The IETF and ATM Forum have both defined signalling protocols independently of one another and are in the process of merging signalling methodologies in order to provide IP-level real time services support over ATM networks. ATM provides resource reservation through a connection setup process whereby the sender initiates the connection and all parties involved in the connection must agree upon the level of resource allocation. The ability of this type of signalling to efficiently handle RSVP-style reservations is not obvious. The following are previously proposed mappings: (i) single ATM Virtual Circuit (VC) per RSVP flow, (ii) multiple RSVP flows per ATM VC, and (iii) multiple ATM VCs per RSVP flow. The first case would support the RSVP notion of dynamically changing QoS by tearing down an old VC and setting up a new VC; ATM does not allow the renegotiation of the QoS of a VC. The drawback to this approach is the excessive signalling overhead; in ATM, tearing down and re-establishing VCs requires a non-trivial amount of processor time as well as network delays. In the second case, large VCs would be established between IP routers in an ATM network. The major disadvantage to this approach is its static nature; accurately provisioning a large VC may be difficult, and it eliminates the possibility doing QoS dependent routing. The third case would allow the possibility of greater heterogeneity since potentially a separate VC could be created for each distinct QoS for a multicast session. The obvious drawback is the excessive amount of duplicate traffic. PROPOSED APPROACH The traffic characteristics of digital VBR video, such as JPEG, MPEG-1 and MPEG-2, are a function of many varying facets of the video such as video content (e.g., video-teleconferencing versus high-action movies), frequency of scene changes, and encoding techniques (e.g., DCT-based inter/intra frame coding) [TSANG97,TSANGA97]. Thus in ATM, specifying the QoS for VBR video traffic is an inherently difficult task. If the QoS is specified such that the VC is under-provisioned, cells belonging to large bursts, which usually correspond to the reference frames (I frames) in MPEG, will have an increased probability of being lost. Highly compressed video such as MPEG is quite sensitive to cell loss. Lost cells from reference frames may trigger errors in resolving the dependencies in subsequent frames. Aware of these issues, the ATM Forum has proposed that peak resources be allocated for real-time VBR traffic while, for the sake of network utilization, Available Bit Rate (ABR) traffic is allowed to consume the remaining network resources. The goal is high network utilization. However, there exist scenarios where the network is primarily used to transport real-time VBR traffic (such as video traffic in a video server environment) and there is very little ABR traffic. In such cases, allocating peak resources would be enormously wasteful. Allocating less than peak implies higher network utilization. However then the possibility for unpredictable statistical fluctuations, which may result in unacceptable cell loss and delay, arises. Our approach is to dynamically renegotiate the QoS, as necessary, throughout the lifetime of a VC. This method raises several issues such as the latency incurred by the renegotiation process, the frequency of renegotiation and the manner in which renegotiation is triggered. A fundamental issue in this approach is the tradeoff between the penalty of latency, incurred by the renegotiation process, and the penalty incurred by the network, in terms of under-utilization, or the VC, in terms of excessive loss and delay due to the lack of resources. In our method, the renegotiation process will be a function of the particular type of traffic and the network feedback mechanism. The type of video may be stored video, in which the optimal renegotiation points may be computed in advance, or live video, which could contain a number of unpredictable variables such as scene complexity, scene change frequency, etc. The network feedback mechanism is based upon a predictive approach where the amount of network resources consumed by the VC is monitored. For instance, a VC's buffer occupancy at a switch port may be tracked. If the buffer occupancy increases at a high enough rate, the network feedback mechanism can be set to trigger a renegotiation for a higher QoS level. Our deliverables will be duofold. Firstly, we will provide analytical and simulation analysis of various renegotiation methods. We will perform this in the context of providing a framework of metrics for evaluating the efficiency and performance of these and other signalling processes. Providing the functionality offered by RSVP through ATM, as described in the preceding section, may be described as awkward at best. We plan to use our approaches and results to extend the existing standards to support dynamic QoS mechanisms for large IP networks. RSVP already contains a number of features which are suitable for a heterogenous environment such as the proposed NGI. Clearly extensions to the currently evolving ATM signalling standard Q.2931, must be proposed and studied. BIBLIOGRAPHY [CHEN97] H. Chen, R. Tsang, J. Brandt, J. Hutchins, ``A Survey of IP- over-ATM Architectures'', to appear in ConneXions. [TSANG96] R. Tsang, A. Pavan, D. Du, ``Digital Video Transmission over ATM Networks'', ACM/Springer-Verlag Journal on Multimedia Systems. [TSANGA96] R. Tsang, P. Keat, T. Chang, J. Hsieh and D. Du, "Fast Packet Switching Algorithms for Dynamic Resource Control over ATM Networks", to appear in the Journal of Computer Communications, Special Issue on Enabling ATM Networks. A short version of this paper also appears in the IEEE Proceedings of Global Communications, London, November 1996. [TSANG97] R. Tsang, J. Hsieh, D. Du, "An Experimental Study of VBR Video over Various ATM Switch Architectures", Proceedings of IFIP High Performance Networking, White Plains, New York, May 1997. Submitted to IEEE/ACM Transactions on Netowrking. [TSANGA97] R. Tsang, H. Chen, J. Brandt, J. Hutchins, ``Digital Video Technologies and Their Network Requirements'' paper in preparation.