Re: State of TCP Fairness Question

From: Alhussein Abouzeid ([email protected])
Date: Mon Mar 12 2001 - 17:48:23 EST

  • Next message: Andreas Pitsillides: "INFOCOM 2001, Anchorage Alaska, 22-26 April 2001 --> Call for participation"

    An additional point: In the case of congestion, TCP fairness is improved
    by a factor of sqrt if you use RED instead of Tail Drop. Thus, if the
    ratio of the lowest to the highest flow (in steady state) through a
    congested Tail Drop gateway is 0.2, this ratio is improved to 0.45 if you
    use an equivalent RED.

    For details/derivation, please see the paper
    Alhussein A. Abouzeid, Sumit Roy, "Analytic Understanding of RED Gateways
    with Multiple Competing TCP Flows", in Proceedings of GLOBECOM'00, also
    available electronically at
    http://students.washington.edu/abouzeid/globecom.pdf

    -Hussein.

    On Mon, 12 Mar 2001, Chris Metz wrote:

    > Hi-
    > Seeking to understand that status of the TCP fairness issue. Here are my
    > data points:
    >
    > - TCP thoughput is a function of packet size, RTT, loss probability and RTO
    > as shown in the "Modeling TCP Throughput ..." paper by Padhye, et al from
    > Sigcomm '98.
    >
    > - TCP may be biased against connections with long RTTs because the additive
    > increase time is roughly one segment per RTT. Connections with longer RTTs
    > take longer to ramp up and while this is happening those with smaller RTTs
    > can acquire available bandwidth. TCP constant rate algorithm as suggested
    > by Floyd and later Henderson, et al paper could overcome this limitation by
    > "equalizing" the rate by which senders with different RTTs increase their
    > sending rates during congestion avoidance. Not implemented.
    >
    > - New techniques (larger initial windows, limited transmit) can help the
    > performance of small flows.
    >
    > Anything else to look at? What about TCP and Diffserv? It was suggested by
    > recent Yeom and Reddy paper that it may be difficult to achieve
    > "contracted" TCP throughputs for higher rates and that those with smaller
    > rates would fare better.
    >
    > My apologies if this is not relevant for the mailing list but I was
    > thinking that sat links can encompass the entire spectrum of
    > characteristics that impact TCP throughput.
    >
    > Thanks ...
    >
    >
    >
    >
    >
    >
    > Chris Metz
    > Lead IP Architect
    > Solutions Integration
    > Service Provider Line of Business
    > Cisco Systems
    > email: [email protected]
    > offic phone: 408-525-3275
    > home office: 914-241-0423
    > pager: 800-365-4578
    > Internal URL: http://wwwin-people.cisco.com/chmetz/chmetz.htm
    >
    >



    This archive was generated by hypermail 2b29 : Mon Mar 12 2001 - 18:35:42 EST