Figure 1. Global Common Time Reference.
Moreover, the deployment of new high bandwidth multimedia applications will boost network traffic and consequently the demand for very high capacity transmission technologies, such as Wavelength Division Multiplexing (WDM). Networks will suffer (i) electronic switching bottlenecks among high-speed links and (ii) communications link bottlenecks between high capacity core technologies and low speed access technologies.
This paper addresses the design of interactive systems for applications such as toll quality telephony, videotelephony and videoconferencing, highlighting the benefits brought by the availability of global common time reference derived from GPS (Global Positioning System). Common time reference is essential to keep the user perceived delay within the 100 ms bound while avoiding the two above mentioned bottlenecks. The proposed solution can be applied to both IP and ATM networks, does not require changes to any of the existing protocols, and enables traffic aggregation in the core of the network–thus not requiring nodes to keep state information on microflows–while providing a guaranteed quality service to individual applications.
Many interactive applications, such as, telephony, videotelephony and videoconferencing require at the receiver continuous playing of samples captured at the sender. Continuous playing implies a constant delay service to be provided at the application layer, i.e., where samples are acquired and played. Since some of the end-to-end delay components can vary during a session, specific action is required at the receiver to keep constant the end-to-end delay between the application layers, thus enabling continuous playing. Before samples are played, delay variations should be "smoothed out" by buffering the samples that have experienced (in the network and in the decoder) a delay shorter than the maximum. This introduces a resynchronization delay component that is typically the time spent in a replay buffer [4]. On exiting the replay buffer, all the samples have experienced the same delay since the time they were acquired at the sender side. Such an overall delay is equal to or larger than the delay bound the system can guarantee.
This paper shows that the end-to-end delay bound can be minimized by a system with a global common time reference, independently of the underlying packet technology (e.g., ATM or IP) and the session rate. Resynchronization components can be kept small, e.g., 125 µs, as well as the queuing delay in each switch, e.g., 250 µs. Global time is used in two ways:
The next section describes how synchronous switches use the common time reference to implement Time-Driven Priority (TDP) forwarding and how such synchronous switches can be deployed in any network–one significant specific case being the Internet–together with traditional asynchronous switches. Moreover, the creation of Synchronous Virtual Pipes (SVPs) enables traffic aggregation while providing service guarantees to individual packet flows. The following section discusses the benefits stemming from the application of time-driven priority forwarding to voice traffic. The deployment of the same technology for the transmission of constant and variable bit rate video is addressed. Then, the advantages of using the global time in the video capture card and the video display card are described. Scalability properties of synchronous packet switches are highlighted in the following section. Finally, conclusions are drawn.
Thus, all switches around the globe have an identical time structure that is collectively called a Common Time Reference (CTR) and can be used to coordinate the acquisition of samples at the sender with the playing of them at the receiver, as discussed the subsequent sections. Moreover, the CTR enables the implementation of Time-Driven Priority (TDP) [1][2] for periodically forwarding real-time packets, for example inside IP and ATM networks. Real-time packets are periodically granted the highest priority, while "best effort" traffic is sent with lower priority. In order to guarantee that a packet can be forwarded during a predefined TF, the right to transmit the corresponding amount of bits on links during the TFs is reserved beforehand by applications. Periodic forwarding indicates that the reservation, and hence the forwarding pattern, repeats itself in every time cycle and in every super cycle. TDP guarantees that the end-to-end delay jitter is less than one TF and that reserved real-time traffic is transferred from the sender to one or more receivers with no loss due to congestion.
TF reservations for each micro-flow should be made in each node in order to guarantee constant delay and no loss due to congestion. Even though no per-micro-flow information is needed to forward packets, the processing load introduced by signaling can affect scalability in the network core. Maximum scalability can be obtained by reserving TFs for semi-permanent Synchronous Virtual Pipes (SVPs). Packets traveling through the core are carried within an SVP. In this scenario, when an application attempts to reserve network resources for its packet micro-flow, its signaling message is not forwarded to core switches. Rather, as shown in Figure 2, it is processed by the Access Bandwidth Broker (ABB) at the ingress of the SVP the micro-flow is to traverse. If the available TFs–among those reserved for the SVP–are compatible with the application’s request (i.e., an end-to-end schedule is feasible), the ABB assigns the corresponding TFs to the new micro-flow and forwards the signaling request to the ABB at the other end of the SVP. Even though intermediate switches traversed by the SVP are not aware of the TFs reserved to each micro-flow, they can forward packets according to the TDP forwarding principles since a reservation had been placed for the SVP. As a result, packets experience the same guaranteed quality service as if resources had been reserved to their micro-flow, but only the ABB is aware of the TF assigned to the micro-flow.
In case a micro-flow gets to the device at the ingress of the SVP through an asynchronous network, as in the configuration shown in Figure 3, the border device buffers each packet until the proper TF–among those reserved to the SVP–assigned to the micro-flow. Notice that SVPs can be set up over multiple synchronous subnetworks interconnected by asynchronous ones. Only the border device at the boundary with an asynchronous network, possibly co-located with the ABB, needs to be aware or the TF assigned to the micro-flow.
The queuing delay introduced by TDP is independent of the packet size, session rate, and the amount of resources reserved to it. This is unlike other asynchronous scheduling algorithms, such as, weighted fair queuing, in which the queuing delay is proportional to the packet size and inversely proportional to the rate. Such asynchronous algorithms possibly require overallocation of resources to meet the end-to-end delay requirement [3]; in some cases where propagation delay is large, voice compression may lead to a resource allocation larger than without compression. Paradoxically, it may be the case that where voice compression would be most beneficial (e.g., trans-continental phone calls) it becomes useless.
On one hand, since TDP does not require overallocation, the bandwidth reduction stemming from compression can provide full benefit. Moreover, the low queuing delay bound in the network allows the sender to:
The elimination of such a long transmission delay is achieved by transmitting the captured video frame as a short burst. Packet switching allows burst transmission of video frames in packets, i.e., as shown in Figure 5, a video frame is captured, put into a packet, and then transmitted as a burst into the network. Therefore, the only way to transmit video frames for interactive applications with minimum delay is over packet-switched network.
Since video frames are captured periodically, in order to minimize the delay bound, periodic resource allocation with periodic transmission synchronized with their capture are required. TDP is the only known way to satisfy those requirements, while guaranteeing absence of loss with minimum delay bound.
It is worth noting that even though all the video frames are not encoded with exactly the same amount of bits, the capacity reserved on the links is not wasted since it used to forward "best effort" (i.e., non-reserved) traffic. Avoidance of loss due to congestion is guaranteed to all the video frames, provided that the amount of bits encoding them does not exceed the reservation.
It may be inefficient to transfer such a compressed video stream over a constant bit rate channel, e.g., the one provided by a circuit switched network (see [4] for a detailed discussion). If the encoder is operated in such a way that it produces a constant bit rate flow, it can introduce a delay up to the time between two successive I-frames.
Figure 6. Complex Periodicity Scheduling of MPEG Video Stream.
Complex periodicity scheduling allows MPEG video frames to be transmitted as soon as they are encoded, analogously to what is described in Section 4 for fixed size video frames. TDP together with global time facilitates the realization of complex periodicity scheduling, which provides deterministic quality of service guarantees to variable bit rate traffic. In complex periodicity scheduling the amount of transmission capacity reserved on the links traversed by a session varies in a repetitive manner. This is what is needed to transmit an MPEG encoded video stream as shown in Figure 6. A related study [5] showed that an MPEG encoder can be successfully implemented in a way that encoded video frame size never violates the resource allocation.
TDP enables to control traffic patterns across each switch, i.e., the maximum number of packets that during each TF are to be moved from every input to the same output. This can be leveraged of in the switch design: optimal input-output switching is obtained with low (x2 or lower) speed up in the switching fabric, or even without speed up at all. Instead, asynchronous switches require high speed up in order to achieve high throughput.
A non-blocking fabric allows any possible input-output connection at any time; a simpler (blocking) fabric allows only a limited number of simultaneous input-output connections. With TDP, packet arrival can be controlled in a way that incompatible input-output connections are not required during the same TF, thus avoiding unfeasible switching configurations. Thus, thanks to the increased flexibility introduced by the time dimension, it is possible to design synchronous packet switches based on blocking switching fabrics that achieve a throughput comparable to that of asynchronous switches with non-blocking fabrics. Or, in other words, given the state of the art aggregate switching fabric capacity, synchronous switches can accommodate more ports, each at a higher capacity, than asynchronous switches.
When a TDP network is deployed to carry voice calls, compression can be fully benefited to reduce the amount of link capacity used by each call. This is not the case with other asynchronous queuing schemes that possibly require overallocation to satisfy end-to-end delay requirements.
When dealing with videotelephony, encoded video frames are transmitted in bursts of packets with controlled delay and no loss. The delay perceived by users is lower than the one obtained by carrying video calls over a circuit switching network which requires delay to be introduced for smoothing out the burstiness of the video source.
Global time and TDP forwarding offer the only solution for the transmission of video frames also when they are encoded with a highly variable amount of bits, such as with MPEG. Each video frame can be transmitted in a burst of packets as soon as it is encoded with no shaping delay and no loss due to congestion. Through complex periodicity scheduling, resource reservation is fitted to the size of encoded video frames thus leading to efficient resource utilization.
The paper also shows that global time can be beneficial when its deployment is not ubiquitous, as it will be the case especially in the initial transition phase when synchronous switches will start to be installed on backbones and will coexist with legacy asynchronous ones. Moreover, the concept of synchronous virtual pipe enables flow aggregation, which is particularly useful in the core of global networks in order to increase scalability.
Finally, global time is beneficial to increase the scalability of packet switches. Switch scalability is essential to satisfy the increasing demand for electronic switching capacity that will be boosted by new bandwidth-consuming multimedia applications. Moreover, when very high capacity backbones will be built to satisfy the above demand, the congestion free operation guaranteed by TDP will avoid potential bottlenecks at the boundary with lower speed access networks.
