Readings for Lecture Sep 18, 2008

Fundamental Design Issues for the Future Internet

The paper discusses the fundamental architectural design issues of “future“ Internet. The Internet at the paper’s time offers only single class of “best-effort” service and is neutral; there is not admission control and the network does not guarantee the about when or if packets will be deliver.

The emerging real-time applications like video and voice have different characteristics and requirement from data application could significantly alter the Internet architecture since real-time applications do not perform adequately when running over the current Internet due to the variations in transmission delay. In addition, real-time applications cannot back off during congestion and they continue contend for bandwidth with data application resulting in the lack of bandwidth for data applications.

The above issue can be resolved by using fair queueing and modifying the implementation of applications to adapt to delay variations. However, without admission control, the network can run into the situation where fair bandwidth is too small. Thus, it is necessary to extend current internet service model from one best effort class to include a variety of service classes.

The goal of the network design then is to maximize the performance of the resident applications, called efficacy V≡ΣiUi(si). V can be improved without the necessity to supply more bandwidth by provides a wider variety of services rather than just a single class of best effort service.

To determine the service for a flow, there are two possible ways. First, network can implicitly supply; this means that the application does not tell the network about its service requirement. However, this requires embedding application information into the network, which violates the Internet tenet. Second, applications can explicitly request their desired service class. This also has some disadvantages like the incentive for users and the lack of flexibility of service model.
The paper does not address the problem of delay bound and also assumes that the bandwidth will be extremely inexpensive.

Supporting Real-time Applications in an Integrated Services Packet Network Architecture and Mechanism

The paper proposes an Integrated Services Packet Network (ISPN) architecture that supports two distinct kinds of real-time services: guaranteed service which involves pre-computed worst case delay bounds and predicted service which uses the measured performance of the network in computing delay bound.

The characteristics of a class of real-time application dubbed playback applications are discussed to indentify requirements for the design proposal. The playback class has four important points: 1) it prefers low delay, 2) The applications need to know about the absolute or statistical bound of the delay, 3) Data just need to arrive before the play-back point, 4) The play-back applications can tolerate the loss of a certain fraction of packets.

The guaranteed service is if the network hardware is functioning and the client is conforming to its traffic characterization, then the service commitment will be met (does not require the conformant of other network clients). The predicted service is that the network commits if the past is the guide to the future, then the network will mot its service characterizations.

The Weighted Fair Queueing (WFQ) algorithm is suitable for guaranteed service. This scheduling does not favor the bursty sources. The FIFO algorithm is better for packet scheduling in predicted service, which provide better the utilization of the network, however, does not provide any isolation. The unified algorithm of WFQ and FIFO can handle both types of services in that WFQ scheme is used as a framework into which other scheduling algorithms can be fitted. All predicted services is assigned a pseudo WFQ flow (this flow has a constant data rate).

The paper assumes that most of future real-time application will fit the playback applications. Other types of real-time applications are not discussed in this paper like real-time control systems, which do not tolerate the loss of data.