High-performance simulation of low-resolution network flows

摘要

Simulation of large-scale networks demands that we model some flows at coarser time scales than others, simply to keep the execution cost manageable. This article studies a method for periodically computing traffic at a time scale larger than that typically used for detailed packet simulations. Applications of this technique include computation of background flows (against which detailed foreground flows are simulated) and simulation of worm propagation in the Internet. The approach considers aggregated traffic between Internet points of presence (POPs) and computes the throughput of each POP-to-POP flow through each router on its path. This problem formulation leads to a nonlinear system of equations. The authors develop means of reducing this system to a smaller set of equations, which are solved using fixed-point iteration. They study the convergence behavior, as a function of traffic load, on topologies based on Internet backbone networks. They find that the problem reduction method is very effective and that convergence is achieved rapidly. The authors also examine the comparative speedup of the method relative to using pure packet simulation for background flows and observe speedups exceeding 3000 using an ordinary PC. They also simulate foreground flows interacting with background flows and compare the foreground behavior using their solution with that of pure packet flows. They find that these flows behave accurately enough in their approach to justify use of the technique in their motivating application. The authors parallelize the algorithm on a distributed-memory multiprocessor. They exploit the flexibility offered by noncommittal barrier synchronization that permits a processor to handle computation messages even after it invokes a barrier primitive. They also take advantage of application-specific knowledge to minimize synchronization cost, study the performance of their parallel algorithm with both fixed and scaled problem sizes, and observe excellent scalability on a multiprocessor supercomputer.