Achieving truly large-scale, real-time distributed simulation has remained a highly elusive goal primarily due to limitations in current generation network technology. While the progression from early, broadcast-based DIS approaches to the current multicast-based publish/subscribe approaches (Van Hook and Calvin, 1994) embodied in the HLA (Defense Modeling and Simulation Office, 1998) has significantly extended the number of simulated hosts that can be supported, limitations in the networking infrastructure still force compromises that historically have constrained simulation exercises to only a few tens of hosts. This constraint is overwhelmingly the result of the networking infrastructure's inability to adequately control the delivery of unneeded and unnecessary traffic to the simulation hosts - a problem so severe that the overhead of reading and discarding unneeded state update messages can critically impair a host's ability to perform its primary simulation tasks. This unintended distributed denial of service problem also results in an enormous waste of network and human resources. In particular, the approach taken for United Endeavor of over-provisioning WAN links with multiple T3 circuits coupled with multiple man-weeks of trial-and-error to find a "workable" multicast grouping is extremely expensive and time-consuming.
Under the Defense Advanced Research Projects Agency (DARPA)-sponsored Specialized Active Networking technologies for Distributed Simulation (SANDS) project, we are developing Active Networks-based capabilities to significantly improve multicast-based distributed simulation performance1. In particular, we have created a capability for dynamically and automatically configuring and reconfiguring application-specific content management filters directly within intermediate network routers where they are most effective. With this approach, we have been able to achieve the ultimate goal of being able to eliminate all irrelevant network traffic at the earliest opportunity within the network, providing optimal use of both network and host resources. Moreover, through the use of filter processing acceleration techniques, such as tree-based filtering mechanisms, information theoretic-based filter complexity reduction techniques, and cooperative, distributed content filtering strategies, we have been able to craft an extremely low-latency (<100 usec) interest filtering capability that provably scales to support millions of simulated entities. This capability is intentionally designed to mesh seamlessly with the High Level Architecture (HLA) Declaration Management (DM) and Data Distribution Management (DDM) services, and a working version has been demonstrated using ModSAF and the Georgia Tech HLA-compatible Run Time Infrastructure (RTI).
This paper reviews our active-networks-based interest filtering architecture, the interrelationships with HLA DM and DDM services, and provides technical details and performance measurements of the various components that clearly demonstrate the performance and scalability claims for our approach.