Faculty Recruiting Make a Gift

RAPID: RTT-Fair Congestion-Control for Terabit Networks

09 Dec
Tuesday, 12/09/2008 5:30am to 7:00am
Seminar

Jasleen Kaur
University of North Carolina

Computer Science Building, Room 151

TCP congestion-control is fairly inefficient in utilizing bandwidth in high-speed and dynamic environments. The main culprit is the slow bandwidth-search process used by TCP,which may take up to several thousands of round-trip times (RTTs) in searching for and acquiring the end-to-end spare bandwidth. Unfortunately, while several alternate congestion-control protocols have been proposed to speed up the search process, most of these struggle to remain non-intrusive to cross-traffic while achieving high speed --- consequently, even these "high-speed" protocols may still take hundreds to thousands of RTTs in searching for available bandwidth in 1-10Gbps networks. Experimental evaluations show that such protocols can utilize no more than 60% of the spare bandwidth on 10G links.

We argue that the legacy design framework of window-based transmission and control, that fundamentally operates at an RTT timescale, is responsible for the poor scalability of existing protocols. We design a new rate-based framework for congestion-control that allows TCP connections to boldly search for, and adapt to, the available bandwidth within a handful of RTTs. Our key insight is to rethink the timescale at which congestion-control probes for available bandwidth---we show that by shrinking this timescale, it is possible to design a protocol that achieves a high bandwidth-search speed without significantly overloading the network. Our resultant approach relies on carefully orchestrated inter-packet gaps at the sender---that help quickly probe for several different rates using only a few packets---and estimates the available bandwidth based on gap increases at the receiver.

We use this framework to design a new protocol, referred to as RAPID, using mechanisms that promote efficiency, queue-friendliness, and fairness. Our experimental simulations with 1-10Gbps networks indicate that RAPID: (i) converges to an updated value of bandwidth within 1-4 RTTs; (ii) helps maintain fairly small queues; (iii) has negligible impact on co-existing regular TCP traffic aggregates; and (iv) exhibits excellent fairness among co-existing RAPID transfers. The rate-based design allows RAPID to be truly RTT-fair.