Michele C. Weigle 
clark@cs.unc.edu
http://www.cs.unc.edu/~mcweigle


Investigating the Use of Synchronized Clocks in TCP Congestion Control

  In TCP Reno, the most common implementation of TCP, packet loss is the
sole indicator of network congestion.  Even so, TCP hosts are not
explicitly notified of lost packets, but must rely on timeouts and
duplicate acknowledgements (ACKs) to indicate loss. Current efforts to
provide explicit congestion notifications (ECN) to end systems require
modifications to both end systems and routers. We investigate methods in
which end systems can monitor their performance and detect network
congestion without explicit router support.

  We believe that information about one-way transit times (OTTs) can be
used to detect congestion. (OTT is the actual time between the sending of
a packet and its receipt by the receiver.) Queues in routers build up
before they overflow, resulting in increased OTTs. We have developed a
variant of TCP, called Sync-TCP, which uses synchronized timestamps to
gather a connection's OTT data. Sync-TCP includes timestamps and OTT
information in each packet such that upon receiving an ACK, a sender can
determine the OTT on the data path, the OTT on the ACK path, the time the
data packet was received, and the amount of time the ACK was delayed at
the receiver.

  In preliminary experiments, we have shown an efficiency improvement
(i.e., fewer retransmissions) with Sync-TCP over TCP Reno.  In these
experiments, we reduced the sending rate by 1/2 when the virtual queue
became over half full.  The queue size threshold and sending rate
adjustment were picked arbitrarily as a first test of Sync-TCP.  We want
to improve Sync-TCP's performance by using a virtual Random Early
Detection (RED) queue at the sender as an indicator of congestion in the
network and experimenting with various methods of adjusting the sending
rate after congestion detection. Challenges include congestion feedback
time (the amount of time between the congestion event and the detection of
the congestion by the sender) and detecting congestion over multiple
bottleneck links.