On Real-time Scheduling, Utilities and Frameworks

The premise of Shenker’s paper is that while the Internet of the mid-90’s have scaled to support a growing number of users, a new set of applications with radically different traffic requirements were beginning to appear at that time. Shenker claims that with such applications in the horizon, the design and architecture of the Internet needs to be reexamined in the context of possibly introducing a new service model, its associated invocation and the use of admission policies similar to what is used in circuit switched networks. There were a number of  future design considerations presented but I will concentrate on the utility function and how Shenker described the abstract framework around this economic model. It is important to point out that while mechanisms developed at the time can be  fair and have nice network properties such as efficiency and stability, they  are unable to differentiate one type of traffic from another. In short the existing mechanisms  lack a user-centric criteria to determine which users require more resources and more importantly, whether the user’s expectations were met. Shenker proposes the use of utility functions to capture this expectation. With users specifying their expectations through this function, the goal of the network can therefore be simply to maximize overall utilities (of course there could be some variation on this optimization). With the utility function, Shenker was able to motivate the use of more than one service class using a simple example. He then begins to introduce the notion of incentives and how one who requires less resources could be compensated by paying less. This allows the network to potentially be both efficient and economically viable while meeting the resource demands of different classes of users. In addition, it is proposed that some form of admission control be in place to avoid overloading. Since this paper was written in 1995, we have the benefit at looking at what has happened over the last 10 years. While the idea of using utility functions sound reasonable, it seems to have not taken off. Personally, when using Voip or streaming video applications, I do not specify my utility functions but I do give feedback in some cases (Skype, YM). From what I understand, these functions seem to be subjective and differ from one user to another. So I will assume that either my preferences were guessed by the network correctly or the Internet is still the same except that we have better links and higher capacities available.

The next reading by Clark and et. al. (interestingly Shenker is also in here and I will go back to this note in a moment) describes an architecture that defines two types of service commitments for real-time applications and the mechanisms to implement these commitments. The first type is the guaranteed service class which may be considered as having hard requirements — that is the delay estimation is based on worst-case static bound. The second type is the predictive class or the tolerant class. From an initial estimate of the source behavior, this class introduces the idea of measurement-based admission control. This allowed for the adaptation or flexibility to admit new connections and therefore improve the utilization. They then begin to describe the mechanisms to implement these service classes and detail the results of their experiments wherein they can offer differentiated services to a number of defined user classes.

From the point of view of the authors, it is both important to put in place admission control mechanisms to limit load and to introduce pricing to allow for incentives. Going back to my note, I believe some of the experiences of one of the authors (Shenker) allowed him to generalize the ideas in this paper and therefore developed the framework in the first paper (2nd paper if we are strict with the chronology) which incorporated the service classes, admission control, the user criteria and the incentives.