UDP vs. TCP for data transfer in lossy networks

[Stefan Winter <stefan.winter@xxxxxxxxxx>]

Hello,

there was an interesting talk at the 72nd IETF about using reliable vs 
unreliable transports when transmitting data, and how to achieve maximum 
throughput.

I found this quite interesting, not only in regards to RadSec. It 
provided quite a few insights. 
(http://www3.ietf.org/proceedings/08jul/slides/opsarea-2.pdf)

First off, he figured that you can tweak the timeouts in UDP so that you 
achieve a better throughput with UDP(!). Unfortunately, we can not tweak 
it in this way for RADIUS, because the tweak was about reducing the 
waiting time just a bit over the expected round-trip time. In his case, 
setting it to one second wating time made UDP underperform TCP, but 
everything below .2 seconds beat TCP (slide 26). Unfortunately we are 
stuck with 5 seconds of RADIUS timeout when using UDP, which in turn 
means we are much better off with a TCP-based transport.

Another interesting result is that TCP is only performing better up to a 
certain threshold of loss in the network (he terms it as "mildly lossy 
networks"). Around 40% loss, TCP completely breaks down, and 
transmitting with UDP gets better seults (slides 20-22). That was 
somewhat unexpected to me. Luckily, in our roaming backbone, we will 
hopefully never have to cope with such loss probabilities! (And so, TCP 
is again fine)

There is more interesting stuff towards the end, about TCP behaving 
chaotically when the loss gets too high, I'd generally recommend a brief 
look at the presentation.

Greetings,

Stefan Winter

--
Stefan WINTER
Ingenieur de Recherche
Fondation RESTENA - Réseau Téléinformatique de l'Education Nationale et de la Recherche
6, rue Richard Coudenhove-Kalergi
L-1359 Luxembourg

Tel: +352 424409 1
Fax: +352 422473