Scalability and Accuracy in a Large-Scale Network Emulator

Experimental distributed services are diÆcult to evaluate due to their scale and the complexity of their host environment|the Internet. Unfortu-nately, it is diÆcult to deploy development software at more than a handful of sites or to evaluate it under controlled conditions. Simulation is a useful alterna-tive, but it abstracts key details of real implementa-tions. Lack of an accepted evaluation methodology that can predict real behavior is a key obstacle to progress in this otherwise thriving research area. ModelNet is an Internet emulation environment de-signed to remove this obstacle. ModelNet emulates a wide-area network on a scalable gigabit LAN cluster, enabling researchers to deploy unmodied software prototypes in a congurable Internet-like environ-ment and subject them to faults and varying net-work conditions. Edge nodes running user-specied OS and application software are congured to route their packets through a set of ModelNet core nodes, which cooperate to subject the traÆc to the band-width, latency, and loss prole of a target network topology. This paper describes and evaluates the ModelNet ar-chitecture and its implementation, including novel techniques to balance emulation accuracy against scalability. Experiments with several large-scale dis-tributed services demonstrate the generality and ef-fectiveness of the infrastructure