Modeling large Ethernet networks for the ATLAS high level trigger system using parameterized models of switches and nodes

Large local area Ethernet networks are strong candidates to connect data sources and processing nodes in high energy physics experiments. In the high level trigger system of the ATLAS LHC experiment several Gbytes/s of data, distributed over 1700 buffers, have to be delivered to around a thousand processing nodes. Due to the network size, its performance and scalability can only be assessed by modeling. To avoid lengthy simulation runs, and concentrate only on characteristics important for network transfers, the components of the system need to be parameterized. The network performance depends on traffic patterns generated by processing nodes and switching capabilities of the network, we therefore evaluated and modeled both processing nodes and switches. We have developed a parameterized model of a class of switches, where a limited set of parameters, collected from measurements on real devices, is used to model switching characteristics. Another set of simple measurements is used to collect values for parameters used to model processing nodes running the Linux operating system and the TCP/IP communications protocol suite. In this paper we present the set of parameters used in the models together with measuring procedures used to calibrate our models. Calibrated models are used to model small test-bed setups with random traffic to validate our approach.