Re: [Aironet] Re : Firmware/packet loss/802.11 (long)

[Jean-Pierre E. <eb...@ee...>]

Hi Jean, hi all,

> > Caceres, Iftode: The effects of Mobility on Reliable Transport
> > Protocols, in Proc. of the 14th
> > Intl. Conf. on Distributed Comp. Sys., pp. 14-20, Jul. 1994
> >
> > or
> >
> > A. DeSimone, Chuag, Yue, Throughput Performance of Transport-Layer
> > Protocols over Wireless LANs, in Proc. of Globecom'93,  Dec.
> 
>         The *BIG* problem of those papers are outdated with regards to
> the hardware we are talking about, because they don't consider MAC
> level retransmissions.
>         Without MAC level retransmission, you would reach their
> conclusions (even though I much prefer SNOOP to Split-TCP). With MAC
> level retransmissions, the conclusion might be different.
>         By the way, SNOOP is a form of MAC level retransmission, and
> their paper shows that it out-perform other solutions. Of course, with
> 802.11 MAC acs, we shouldn't need SNOOP any more.

Wait wait, first the paper are not that outdated as you think. Just
realize the general well expressed idea: If one have retransmissions
below TCP (whatever level it will be (LLC/MAC)) than one has to ensure
that the retransmit timers on TCP level and and the one on the lower
levels should be harmonized. Just a small example: Think of a packet
which got lost on the radio link. The MAC retransmit this packet several
times until it was succesfully delivered (this may take several seconds
with aironet4800 cards!!!). In between the retransmit timer of TCP times
out (retransmit timer of TCP are dynamic from 200ms - several seconds)
and TCP retransmit the packet. The point is, although the MAC has
successfully delivered the packet, TCP will retransmit it! So the same
packet might be transmit several times. This is what the performance
degradation makes.

Three conclusions could be drawn:
- Leave retranmission to higher level protocols (not on MAC). The
problem with this method is that retransmission on MAC level is the most
efficient way to recover from channel errors. So let see what the next
two points bring
- Harmonize MAC level and TCP Timeout Timers. I never read a paper about
that this is definitely a topic for research and might improve TCP
performance significantly. I have some rather hazy ideas how to do that.
Whoever is interested in that topic, can contact me.
- Tell TCP about transmission state of a packet (a not very handy
solution).

Snoop (also the indirect TCP approach has it's advantages) is a great
thing. It does already a this harmonization, whereas the focus goes a
bit in another direction but not on the MAC level(!). The most
efficient  approach to recover from radio link errors still are MAC
level retranmission with all its problems.
See also a paper from my colleague related to that topic: 

M. Schläger, B.Rathke, S. Bodenstein, A. Wolisz; 
                   " Advocating a Remote Socket Architecture for
Internet Access using Wireless LANs"  
                   "accepted for publication in Mobile Networks and
Applications (The Journal of special Issues on
                   mobility of systems, users, data and Computing),
Special Issue on Wireless Internet and Intranet
                   Access, to appear Published by: Balzer Science
Publishers/ACM 


> 
> > You will also find some papers on WLANs, MAC and performance topics at
> > my Web site and websites of my colleaques.
> 
>         URL please ?
> 

see here
http://www-tkn.ee.tu-berlin.de/publications/pub.html
and 
http://www-tkn.ee.tu-berlin.de/~ebert/publication.html

> 
> P.S. : By the way, are you french ? I am...

No, my father is from a former french colony (Vietnam). From there is my
name, the rest of mine is german all the way.