[e2e] Timeout or dupacks?

[Jennifer Rexford]

Tamas,

> Does anybody know about studies that investigate the typical packet
> loss probabilities and proportion of timeouts and dupacks in the
> today's Internet?

The loss probabilities probably differ dramatically depending on where
you measure (and which links are traversed by a particular flow) and
what type of traffic you are looking at (e.g., bulk transfer vs. Web).
Some papers that might be of interest include:

  Vern Paxson, "End-to-End Internet Packet Dynamics," IEEE/ACM
  Transactions on Networking, 7(3):277--292, June 1999.
    ftp://ftp.ee.lbl.gov/papers/vp-pkt-dyn-ton99.ps.gz

  H. Balakrishnan, V. Padmanabhan, S. Seshan, M. Stemm, and R.
  Katz, "TCP Behavior of a Busy Internet Server: Analysis and
  Improvements," Proc. IEEE INFOCOM, March 1998.
    http://www.ieee-infocom.org/1998/papers/02d_3.pdf

> Recent observations have found that, in many cases, retransmission
> due to timeout is a more frequent event - or at least as frequent -
> during a TCP transfer than retransmission due to dupacks.

The large number of short Web transfers tends to contribute to this
effect.  See Section 8.1 of the book "Web Protocols and Practice"
(http://www.awlonline.com/product/0,2627,0201710889,00.html) for some
discussion of this issue.  Also see the Balakrishnan et al paper above.

-- Jen