Parallel Hybrid Network Traffic Models

Fluid-based network traffic models are attractive due to their execution efficiency. They run much faster than the corresponding discrete-event packet-oriented simulation, especially when we study the aggregate traffic behavior of large-scale network scenarios. The efficiency, however, comes at a cost—fluid modeling does not include packet-level details. The ability to accurately capture the interaction between the packets and the network routers and hosts visited by the packets is essential for real-time network simulations, where the simulator must be able to interact with real applications in real time. In particular, the virtual network must be able to carry real packets subject to proper delay and loss and be able to react to these real packets (such as traceroute). Previously we presented a hybrid network traffic model that combines a continuous-time fluid model and the discrete-event packet-oriented simulation. In this article, we examine a parallel processing method for simulations of large-scale networks using the hybrid model. Our method benefits from the observation that the time it takes to propagate fluid characteristics along the path taken by the traffic flows has a lower bound equal to the minimum link delay as manifested by the governing ordinary differential equations (ODEs). A better lookahead can thus be used to allow parallel simulation of the hybrid model to run without more synchronization overhead than the corresponding discrete-event packet-oriented model. We derive an analytical model comparing the fluid model and the packet-oriented model both for sequential and parallel simulations. We demonstrate the benefit of the parallel hybrid model through a series of simulation experiments of a large-scale network consisting of over 170,000 hosts and 1.6 million traffic flows on a small parallel cluster.