Large-Scale Network Parameter Configuration Using An On-line Simulation Framework

As the Internet infrastructure grows to support a variety of services (eg: VPNs), its legacy protocols (eg: OSPF, BGP) are being overloaded with new functions such as traffic engineering. Today, operators engineer such capabilities through clever, but manual parameter tuning. In this paper, we propose a back-end support tool for large-scale parameter configuration that is based on efficient parameter state space search techniques and online simulation. The framework is useful when the network protocol performance is sensitive to its parameter settings, and its performance can be reasonably modeled in simulation. In particular, our system imports the network topology, relevant protocol models and latest monitored traffic patterns into a simulation that runs online in a network operations center (NOC). Each simulation evaluates the network performance for a particular setting of protocol parameters. A recursive random search (RRS) technique is proposed to explore the large-dimensional parameter state space, where each sample point results in a single simulation. In other words, the overall parameter configuration problem is modeled as a "black-box" optimization problem. The goal of RRS is efficiency, i.e., to find "good" network protocol configurations for the current traffic conditions quickly. An important feature of the framework is its flexibility: it allows arbitrary choices in terms of the simulation engines used (eg: ns-2, SSFnet, future scalable simulators etc), network protocols to be simulated (eg: OSPF, BGP, RED, MPLS etc), and in the specification of the optimization objectives. We demonstrate the flexibility and relevance of this framework in three scenarios: joint tuning of the RED buffer management parameters at multiple bottlenecks, traffic engineering usi..