I have categorized the papers by their problem areas. Each list is preceded by a small description of the problem.
Next-generation internet terminals (like PDAs, network
computers, communicators) are expected to support intensive
continuous-media applications like audio and video. These terminals are
inherently leaner in terms of processing and memory than powerful PCs or
workstations. Traditionally audio and video applications have been
deployed on powerful Internet hosts. This way, they could avoid clogging
the network by using adaptive hooks provided by RTP/RTCP protocols. Such
protocols allow them to tradeoff network bandwidth for
CPU processing on end-hosts. For example, an audio server can select
ADPCM over PCM to reduce output bit rate. However, this switch incurs
a much higher processing delay on hosts.
We are trying to answer the following questions in this context:
Current flow-based reservation mechanisms (like RSVP or ATM signaling) are geared toward guaranteeing per flow media QoS. Intensive efforts will ensure that these guarantees are in place sooner or later. I claim that many of the sophisticated applications also require other high-level QoS measures. To illustrate a specific example of ``application QoS'', consider a multiclient application that involves clients which add or leave the application during its lifetime. If new clients cannot add (because of insufficient network bandwidth), flow-based reservation rejects such clients. This impacts not only the clients that wish to add but also the existing clients in the application. The resulting service degradation is at a layer higher than media degradation like packet loss or delay, and I refer to it as ``client-level degradation''. An application that demands a guarantee of no more than a certain level of client-level degradation cannot be supported using current flow-based reservation schemes natively.
In these papers, we propose the framework of application reservation to capture and guarantee such sophisticated demands. It turns out that flow reservation is a special case of this framework. A related concept called advance reservations are also a special case of this framework. Application reservation can be implemented using existing flow reservation protocols. Thus, no changes are required to be made to installed RSVP or ATM code.
Embedded systems are increasingly used to support intensive applications in network hosts and servers. Such systems can implement protocols (like RTP, RTCP) that permit adaptive real-time processing of multimedia streams based on the state of the network. This imposes non-deterministic processing requirements on the embedded system architecture. Hence, traditional compile-time tools cannot be used to design or evaluate such systems. Typically simulation is the preferred choice, but it slows down the design process because it is very slow and intensive.
To speed up the design process, we propose a fast analytical tool. We assume that applications that adapt the processing of their protocol stacks or switch them based on the state of the network can be simplistically modeled by probabilistic execution paths. Given a set of such concurrent applications, the tool uses an analytical framework to derive several performance measures of the system architecture. This includes measures like the average nodal wait, the average resource wait, and the resource utilization. These measures can be used to design and to verify systems much faster than by simulation. A software prototype of the tool has been implemented.
The fundamental problem is this: If a network carries applications that are implemented using only two types of connections, simple point-to-point connections that connect a source to a receiver and complex multipoint connections that connect a source to a time-varying number of receivers, how should the network resources be allocated to these connections? Simple connections and complex connections have different service objectives. Simple connections may not want their setup request to be blocked more often than a certain value. Complex connections may require a cap on the blocking of the initial participants as well as that of the new additions. In these papers, we present a simple threshold-based scheme for partitioning the resources between the two connection types on each link. We show how this scheme can achieve optimality (throughput) and fairness (equal degradation to both connection types) on an isolated link. An iterative algorithm is designed to approximately design the threshold on all the links of the network.
Last Modified March 27, 1996