The increased development and use of networked real-time applications in distributed multimedia, remote laboratoring, and distributed virtual reality has generated a large amount of interest in the development of real-time communication protocols to support such applications. Such protocols provide real-time guarantees for message deliveries to the network clients, typically in terms of bounded delay and jitter. This in turn enables the applications to meet their end-to-end timing requirements. In order to provide such guarantees on message delivery, resources in the network must be appropriately allocated. The admission control component of the protocol must ensure that enough resources are available in the system so that a new connection can in fact be guaranteed the required performance. Such admission control protocols have been extensively studied for packet-switched networks [4, 8, 11, 13].
Due to the increasing bandwidth requirements in a large number of applications (such as incorporation of uncompressed video streams in control loops for remote laboratoring), a number of very-high speed networking technologies are currently under investigation for their applicability to support hard-real-time communication, such as ATM [2], FDDI [1], HIPPI [3]. We are investigating the applicability of wormhole-routed network technologies for this class of applications.
As exemplified by recently developed technology, wormhole-routed
networks are a promising approach for high-bandwidth, low-latency
communication for small and medium-sized networks. For example,
Myrinet switches [5, 7] have a link bandwidth of
1.28Gb/s with a worst-case routing latency of 550ns for a 
switch. In the following discussion, we will be using Myrinet as a
reference for the system model.
A well known problem with wormhole-routed packet networks is the potentially large amount of blocking that packets can experience because of contention for links. Because of the very limited amount of buffering in such networks, blocked packets ``remain'' in the network and keep using network resources. Thus, they may in turn cause other packets to be blocked. This may affect a large number of packets over a large portion of the network. Proper connection management strategies and appropriate protocols must be devised that ensure that blocking of packets due to link contention is bounded. Zhao et. al [6] have developed a transmission-control scheme that regulates the rate of connections at the entrance to the network. A worst-case achievable utilization of 50% could be proved, allowing for a simple admission control scheme: new connections are accepted as long as the utilization of the network does not exceed the worst-case achievable utilization.
In this paper we propose a distributed admission control scheme with the following two objectives:
The paper is organized as follows. In Section 2 we describe our model of packet transmission over wormhole-routed networks and derive simple worst-case bounds for the time packets can be blocked during transmission. We also describe the basic traffic model used in the protocol and how the admission of a new connection affects existing connections. From this we derive a general admission control scheme. In Section 3, we describe our distributed admission control scheme. In Section 4, we describe the protocol for the implementation of the distributed algorithm. In Section 5 we describe the MyRT protocol which is the real-time connection management protocol based on the proposed distributed admission control scheme. {Section 6 describes simulation results which illustrate that our scheme compares favorably against centralized schemes. It also contains measurements caried out to time the establishment process with the MyRT protocol over different underlying transport layers. We will conclude and give an outlook on future research in Section 7.