* News & Events


ROSS.Net to be Launched in Summer '04

April 15, 2004

rossnet poster There is a deliberate need for large-scale simulation of various networking protocols in order to understand their dynamics. For example, there are several issues in routing that needs to be understood, such as cascading failures, inter/intra-domain routing stability, and interactions of policy-based routing with BGP features. One needs to perform large-scale simulations of inter-domain routing protocols along with various traffic engineering extensions, in order to see their dynamics cause or effect various performance problems in the current Internet.

Chris Carothers and his team are addressing this need using three techniques. First, they leverage an optimistic synchronization protocol to enable efficient execution on a hyper-threaded, multiprocessor system. Here, simulation objects, such as a host or router, are allowed to process events unsynchronized without regard for the underlying topology or timestamp distribution. If an out-of-order event computation is detected, the simulation object is rolled back and re-execute in the correct timestamp order. Unlike previous optimistic protocols, such as Time Warp, the rollback mechanism is realized using "reverse computation". Here, events are literally allowed to execute backward to undo the computation. This approach greatly reduces the amount of state required to support optimistic event processing as well as increases the performance.

Next, they devise an extremely light-weight model implementation framework called ROSSNet that is specifically designed for large-scale network simulation. The framework poses the question: "what do you really need to model in order to answer a particular protocol dynamics question in a large-scale scenario?" For example, are all layers in a protocol stack really necessary? Can a host just be a TCP sender or just a TCP receiver? Does the simulated host really need to be both? By asking these kinds of questions, the framework enables a single TCP connection state to be realized in just 320 bytes total (both sender and receiver) and 64 bytes per each packet-event.

Third, on issue of design-of-experiments (DOT), they develop a large-scale experiment design component that is part of ROSS.Net. This component allows us to characterize and optimize protocol response. In general the protocol response is a function of a large vector of parameters, i.e. is a response surface in a large-dimensional parameter space. The result of this work includes a unified search and optimization framework with demonstrated ability to pose meaningful large-scale design questions and provide "good" characterizations rapidly.

* Return to main News page