[e2e] TCP, retransmissions, and ISPs with byte caps

[Jon Crowcroft]

a la recherche de temps perdue


a single TCP
connection in possessiion of a small fortune in source
data
must be in search of
a 10Gbps fiber transcontintental link
with an especially unusual loss characteristic

as jane austin once said

is not per packet processing


its simple feature of the control loop

if you dont get it, go back to engineering grade school - its not
retransmssion....

sure there's lots (most) of lie
when it doesnt happen

but when it does, its your highest paying custimer, so listen up.

In message <20020617165354.GA24946 at philou.ch>, Philippe Strauss typed:

 >>On Tue, Jun 18, 2002 at 12:10:41AM +1000, grenville armitage wrote:
 >>>=20
 >>> On a different tangent, has anyone got pointers to
 >>> studies of TCP retransmission rates in the context of
 >>> 'typical' consumer broadband access? (And, e.g.,
 >>> the (dis?)incentive for a dodgy ISP to reduce packet loss
 >>> rates if they're also counting retransmitted packets against
 >>> a customer's weekly/monthly usage cap.)
 >>
 >>At TCP level, I dont'know.
 >>But i have a "smokeping" running on my home PC connected to
 >>the net thru CATV (terayon teracomm, not DOCSIS), you can take
 >>a look:
 >>
 >>http://philou.ch/cgi-bin/smokeping.cgi
 >>
 >>It show packet loss rate on long term average (several hours).
 >>
 >>But, as an ex-CATV-based-ISP network admin, I can say that
 >>there is no incentive to let packet loss rate grow, except
 >>if your are struggled by the internet bandwidth fee toward your
 >>upstream, which used to be the case in '96 - '98 for the emerging
 >>broadband business but not anymore.
 >>
 >>If you want to maximize customer traffic, reduce packet drop.
 >>Retransmit are nothing compared to a connection that works well.
 >>
 >>--=20
 >>Philippe Strauss
 >>http://philou.ch/
 >>
 >>L'indiff=E9rence est le plus grand risque de notre temps,
 >>la forme civilis=E9e de la cruaut=E9.  -- Zenta Maurina
 >>--
 >>

 cheers

   jon