Optimizing TCP Start-up Performance

Abstract

The performance of many networking protocols is dependent on a handful of tuning parameters. However, it is not obvious how to set or adapt these parameters to optimize performance. We believe that this optimization task can benefit from passive monitoring of current network performance. In this paper, we apply this methodology for initializing TCP parameters, such as initial congestion window size and slow start threshold for short connections. We analytically derive the optimal initial parameters, and use simulations to study its effectiveness. Our innovations include: (i) derivation of optimal TCP initial parameters as a function of link characteristics; (ii) abstract a communication path as a virtual link and model cross traffic as perturbation on it; (iii) an efficient architecture for network performance discovery; (iv) a new pacing algorithm that combines leaky bucket flow control with traditional window-based flow control. Our results show this approach leads to significant performance improvement.