No abstract

Prior studies have shown that XCP [l], while showing high potential for effective congestion control, may have significant performance problems if it misestimates the bottleneck link capacity [3]. To explore the magnitude of this implication and ultimately to contain it, we have conducted a control theoretic analysis of the XCP protocol and have studied its properties in the presence of capacity estimation errors. Under such conditions, by using a revised fluid model, we have discovered that XCP will not settle at zero steady-state error. However, we found that the steadystate errur is bounded by the estimation error, and this bound can be exploited in XCP router queue size planning. This preliminary analysis explains what we have observed experimentally in an implementation study ofXCE I. INTRODUCTION Recedy, a new congestion control protocol called XCP [ 11 has attracted much attention in the Internet community because it has shown promise to potentially obtain high utilization in high bandwidthdelay product networks, while maintaining IOW standing queues in routers, and without keeping per-flow state in routers. Both theoretical and simulation studies have shown that XCP controllers are stable, efficient, and fair. However in a previous exyerimental study [3], we have discovered that XCP can have significant performance problemsif it operates over a contention-based link or over a hybrid network with non-XCP queues, We have further shown that the core of the problems is that XCP routers often mis-estimate the link capacity under such conditions. Th!s discovery has serious practical implications because shared. contention-based links are rather common in the Internet: Ethernet, 802.11 wireless LAN, cable modem system, etc. And so are non-XCP queues because it will be impossible to instantly tum every link to be XCP-aware throughout the Internet. We must therefore deal with such capacity mis-estimatiom in XCP,In this research, we analyze the control system behind XCP

No abstract

No abstract