[e2e] TCP flow count estimation

[Christian Huitema]

Jon,

This algorithm breaks badly if a box on the way implements some form of
leaky bucket admission control, and if the leaky bucket can allow more
than 2 packets. The test packets will detect the underlying link
capacity, while a stream of real packets would be constrained by the
token arrival rate.

-- Christian Huitema

> -----Original Message-----
> From: Jon Crowcroft [mailto:Jon.Crowcroft at cl.cam.ac.uk]
> Sent: Thursday, November 29, 2001 7:02 AM
> To: David P. Reed
> Cc: Jon Crowcroft; end2end-interest; Jon.Crowcroft at cl.cam.ac.uk
> Subject: Re: [e2e] TCP flow count estimation
> 
> 
> In message <5.1.0.14.2.20011129094501.0391f488 at mail.reed.com>, "David
> P. Reed" typed:
> 
>  >>Jon - A cute estimator, but besides doubling the header overhead
> and
>  >>routing overhead network wide ...
> 
> well, most apps that do long transfers are using MSS and many MSSs are
> 1500 bytes, so
> its not so bad:-)
> 
>  >>  Has the IP layer's function been redefined so that it guarantees
> that all
>  >>datagrams must be sent along the same physical path and through
>  >>order-preserving queues?
> 
>  >>We know that IP does not have order preserving queues in general -
> fast
>  >>path processing and service-types violate order.  We also know that
> in some
>  >>cases, path stability is violated even in the short term.
> 
> well its easy to try look at this with traceroute - time for some
> experiments
> 
> but the various pathchar programs rely on this hack, and the tiems
> when i have seen them
> break are always when there's a level two hop with some type of
> bandwidth on demand
> (e.g. ATM VC, or frame relay hop with committed access rate policing
> via a window/token
> bucket scheme).....
> 
> so i dont think there's too many cases in the real net out there that
> would confuse this
> besides, like MTU discovery, you only need to do it at the start of
> every route's
> lifetime (icmp "route changed" message considered useful ? :-)
> 
> 
> 
> 
>  >>At 06:24 AM 11/29/2001 +0000, Jon Crowcroft wrote:
>  >>
>  >>
>  >>>here's a daft idea for xmas:
>  >>>
>  >>>if tcp is sending instead of MSS, alternate packets of size
>  >>>MSS-1, and size 1
>  >>>(well payload size)
>  >>>and every packet elicits an ack,
>  >>>it can use path char style estimation of the bottleneck line rate
>  >>>
>  >>>it can use equation basedestimation of the bottlenck avaialble
>  >>>capacity
>  >>>
>  >>>between these two, it can deriv an estimate of the number of other
>  >>>tcps (wel module notknowing their RTTs)
>  >>>
>  >>>this might be useful for somethig....
>  >>>
>  >>>j.
>  >>>oh, if every other packet elicits an ack, then you need to
> intersperse
>  >>>the small/large alternation over  3 packets - simple tho
>  >>>
>  >>>think of it like running 2 seperate tcp flows tho, one with MSS 1
>  >>>bute, one with mtu-41 and estimators for both
>  >>
> 
>  cheers
> 
>    jon