Wormhole routing is a cost-effective way to provide very-high speed, very-low latency communication for emerging distributed real-time applications with high-bandwidth demands. Unfortunately, the lack of traffic buffering in the switches makes it difficult to give real-time guarantees to traffic streams that contend for communication links.
In this paper we propose a simple blocking model for traffic on wormhole-routed networks and derive a fully distributed admission control scheme for periodic traffic streams. The scheme is based on a three-pass message exchange among the switches on the connection route. We also developed a CAC protocol to implement the proposed algorithm. MyRTP is real-time connection management protocol which provides the framework for the implementation. Experimental comparisons of a straight-forward implementation of our scheme with a fully centralized algorithm are encouraging. By providing tighter bounds in the calculation of our calculated worst-case blocking times and appropriately modifying the protocol, we can improve the performance of the distributed algorithm as compared to the centralized scheme.
Rate control at the sender host [6] is known to significantly improve the real-time performance of wormhole networks. We are currently studying how to use our current distributed algorithm as a framework to include and make use of information about rate control at the sender.
This distributed scheme has a number of non-obvious benefits. First, it can be seamlessly integrated into existing admission control schemes for store-and-forward networks (e.g. [8, 11]), and thus allow a well-integrated real-time service over hybrid (that is, store-and-forward and wormhole-routed) networks. Second, guaranteeing timing requirements subsumes deadlock avoidance. Our scheme can therefore support a distributed, adaptive routing scheme, which in turn rely on our scheme to guarantee deadlock freedom. Since this study is strictly about admission control, we are currently not taking advantage of this, and establish the connection with a a priori route. This is not required by our protocol, however, and we are currently investigating adaptive, load-sensitive routing schemes.
During connection establishment, resources can be allocated without involving nodes that are not on the connection route. We can also incorporate resource deallocation done by network management. This periodically discovers allocated, but unused resources ( due to connections torn down ), and makes them available to new connections.
The only assumption we make about the switches in this work is that they serve waiting packets in a round-robin manner. There are current efforts (e.g. [10]) to study the benefits of special support for real-time traffic in wormhole-routed switches. We are currently investigating how such additional functionality (e.g. priorities for packets) affects the performance of our scheme.
Substantial work is currently being done on host and switch supported multicasting over wormhole networks (e.g., [9]). The problem of how to guarantee real-time constraints in such a setting still needs to be addressed.