[e2e] Re: RFC 2414

[Mark Allman]

> The IETF Transport Area Directors have received a request to move
> RFC 2414, 
>      Increasing TCP's Initial Window. M. Allman, S. Floyd, C.
>      Partridge. September 1998. (Status: EXPERIMENTAL)
> to standards track.
> 
> We'd like to ask the authors to re-publish here some supporting comments 
> that they gave to the IESG, and we'd appreciate hearing from
> the broader end2end-interest community if the change to standards
> track is now appropriate.
    
The following is the summary of a conversation between Sally, Craig
and myself.  (Although, I'll note that neither of my co-authors
reviewed this text specifically.  I hope to have captured our
conversation, but don't want to put words in their mouths.)

Two data points of larger initial windows since the publication of
RFC 2414:

  * Sally and Jitu's TBIT results:

    Results from the TBIT tool with Jitu, at
    "http://www.aciri.org/tbit/initwin.512.txt.gz", show that, with
    512-byte packets, of the web servers tested:
    
    214 used an initial window of 1 segments, 
    456 used an initial window of 2 segments,
    407 used an initial window of 3 segments,
     30 used an initial window of 4 segments,
     12 used an initial window of 5 segments,
     95 used an initial window of 6 segments,
      2 used an initial window of 8 segments,
    and
      8 used an initial window of >8 segments
    
    So clearly initial windows of two and three are common at web
    servers. 

    The four web servers with an initial window of 17 segments have
    been tracked down to something old and obsolete, as I recall...

  * There are some fairly recent results on using a larger initial
    cwnd in our web server in:

    Mark Allman. A Web Server's View of the Transport Layer. ACM
    Computer Communication Review, 30(5), October 2000. 
    http://roland.grc.nasa.gov/~mallman/papers/webobs-ccr.ps

    (The results are basically in agreement with the arguments made
    in 2414).

And, a problem with the document:

  * A point has been raised that is not addressed in RFC 2414.
    That is, using a larger initial cwnd may potentially hose the
    RTO on slow links.  If we measure the RTT of the (small) 3WHS,
    it does not predict the RTT caused by 4KB of data sent into the
    network back-to-back.  This is not necessarily a problem caused
    by using more than one segment in the initial cwnd.  We see the
    same problem if we use large packets.  However, using a larger
    initial cwnd has the potential to aggrivate this problem.  We
    should probably address this with some sort of workaround (don't
    time the 3WHS when using larger initial windows?).

It seems our consensus is that we'd like to make a pass at the
document to address the RTO problem.  In addition, we can add some
pointers to more recent work in this area (not as important as the
RTO problem).  And then we'd love to see larger initial cwnds put on
standards track.

allman


---
Mark Allman -- BBN/NASA GRC -- http://roland.grc.nasa.gov/~mallman/