|
Hi Julianne & others,
On Mon, 18 Apr 2005, J. Sansa wrote:
> i have looked at the tsunami protocol, aparent results are sure
> aggressive, am afraid too aggressive and unfair to anyother traffic on
> the link. would this do for evlbi, isnt there other traffic using the
> same link???
Well, you have to think things through.
The TCP congestion control algorithm is trying to be fair to oother users.
The critical fault there is that the algorithm tries to measure network
congestion by measuring packet loss. This was a valid approach twenty
years ago, when the links were slow and congested.
However, in modern networks there rarely is congestion or packet loss.
Almost all the packet loss is caused by the slowness of the receiving
computer. The TCP congestion algorithm wrongly interpretes this to be
congestion and slows dowm. This is the main reason why most of the eVLBI
tests have been spectacular failures.
If we use the best available microcomputers, we see almost no packet loss
below 400 Mbit/s speed, about 2.6% packet loss at 512 Mbit/s and 7% at 640
Mbit/s speed. Packet loss in backbone networks (Nordunet, GEANT, SURFnet,
Abilene) has always been below 0.05%.
For nice load graphs see
http://kurp.hut.fi/pub/temp/tsunami/evlbischiphol2005.pdf
Cheers,
Jouko
|
