To: end2end-interest@ISI.EDU
cc:
Bcc:
Fcc: +sent
From: Sally Floyd <floyd@aciri.org>
Subject: checking for Reno TCP in web servers, FYI

Here is a procedure that I think would work for checking whether
a particular web server is using Reno, Tahoe, NewReno, or SACK TCP.
It requires the ability to configure the receiver's advertised
window at the web client, and a mechanism (such as Dummynet) to
drop specific arriving data packets:

* Set the receiver's advertised window so that the maximum window is
  5 packets.

* Request some data that will require at least 25 packets. 

* After the congestion window has reached five packets, drop
  arriving data packets n and n+3.  For example, drop the 
  16-th and 19-th arriving data packets (assuming that the data
  packets have been arriving in-order and without drops).

Given a SACK-capable client, it is easy to check for SACK - just see 
if the SYN and SYN-ACK packets both advertised "SACK-permitted".
So assume that we have a non-SACK connection.

(1) If the server waits for a retransmit timer to expire before
retransmitting packet n+3, then the server is using Reno.
This can be detected, with tcpdump, by the long time interval
before the second transmission of packet n+3.

(2) If the server unnecessarily retransmits packet n+4, the
the server is using Tahoe.  (The Tahoe server will not
have to wait for a retransmit timer to expire, however.)

(3) If the server does not do (1) or (2), then it is
some variant of NewReno or SACK.

(I have not tried to do this myself, however.)

The validation test "./test-all-testReno" in NS illustrates
these four scenarios (and will be available in the nightly
snapshot posted tonight).

- Sally