Modeling and Analysis to Estimate the End-System Performance Bottleneck Rate for High-Speed Data Transfer

The bandwidth available in a high-speed backbone network today is much greater than the capacity at the edge devices. As an example, when the transmission line rate is 10 Gbps, the end computing machine (end-system) may not be able to cope with the arriving data rate and is hence often a bottleneck for the data transfer. This performance bottleneck depends on the end-system workload. A key deficiency in most transport protocols is the lack of an accurate estimate of this bottleneck rate at which the edge device, such as a receiving end-system, can handle network I/O. In this work, we propose and investigate an analytical model of an end-system to estimate the optimal rate for the best network performance. By using a Stochastic Reward Net (SRN) queuing network model, we determine the service-time distribution for processing network I/O. We represent the Network Interface Card (NIC) as a M/M[N]/1/K bulk-service finite-capacity queue; and by applying the above service-time distribution, we calculate the packet-loss probability as a function of the packet-arrival rate which corresponds to the sending data rate. The transfer time for a given file on a path with known propagation delay at a fixed sending rate is determined from the loss probability at that rate. We define the end-system effective bottleneck rate as the sending rate which yields the minimum transfer time for the specified path and file. We then validate the effective bottleneck rate from this model using an experimental testbed where two machines are connected by a 1-Gbps Ethernet connection. We demonstrate that the experimental sending rate which yields the minimum file transfer time is very close to the effective bottleneck rate estimated from the analytical model.