[e2e] was Re: A message to authors - nsdi

[Jim Gettys]

On Thu, 2008-01-10 at 14:42 +0000, Jon Crowcroft wrote:
> so google do application layer retries in a
> lot of things
> 
> the problem with that approach
> is that automatic systems above
> the network layer have no meaningful
> equivalent to the RTT estimation process that
> drives the retransmit timer process, and 
> packet conservation mechanisms in transport
> that is a part of the congestion control
> loop, so we could end up either clobbering the
> net and the end system, or using incredily
> conservative timers for the retries...
> 

And since the kernel interfaces are lacking to RTT and/or other
information such as the delay/bandwidth product, I've seen multiple
application developers have to kludge their own estimates....

One might think it would be wise to expose some of the information
already available to TCP to application programmers; not all of them are
clueless, contrary to popular opinions I've sometime seen echoed by some
on this list.
                                - Jim

-- 
Jim Gettys
One Laptop Per Child