[Mldonkey-users] Re: mld 2.02-16 with MTU 1500 | was: FYI:CVS2.02-16

[Brett Dikeman]

At 11:09 AM +0100 2/21/03, Sven Hartge wrote:

> Why does mldonkey need to know the MTU anyway? All other programs using
> the IP-Stack don't need this value and work perfectly.

I got the impression from the emails traded back+forth that the value 
is used for bandwidth estimation.

If I'm wrong and it's actually used for assembling packets, that's -quite- bad.

Here's an explanation for those who don't know- MTU is a function of 
layer 2 on the ISO model(see below).  MLdonkey (being an application) 
is level 7.  Or, if you prefer the tcp/ip 4-layer model, it's layer 1 
vs 4.

In both cases, the whole point of a layered approach is to isolate 
each layer from the peculiarities/specifics of the level(s) below 
it(or above it, to boot).

The application does not need to know how to interact with a wireless 
card vs. a modem versus an ethernet connection; the network 
drivers(layer 2) handle that sort of thing.  Similarly, my ethernet 
switch(layer 1) really doesn't give a hoot what kind of TCP/IP 
traffic I pass through it.

This sums it up nicely:

http://www.uwsg.iu.edu/usail/network/nfs/network_layers.html

Fun stuff :-)

Oh, also- the Maximum Transmission Unit is a "maximum" for a reason. 
It is not what you should use all the time for your packets- it has 
to be dynamic, like window size.  If the link between two users 
happens to get more congested, packet size and window size are 
supposed to drop until things start flowing again.

It sounds counterintuitive from an efficiency standpoint, especially 
if the link is congested...until you look at the cost of loosing an 
entire packet's worth of data(and hence the window!), and the 
handshake and retransmission involved as a result.

Dropping a packet from 1500 to 1000 bytes only results in a loss of 
about 1% in efficiency, but having to retransmit just one packet 
results in a much more dramatic loss of efficiency- at least 50%.

B
--