[e2e] Are we doing sliding window in the Internet?
lars.eggert at nokia.com
Sun Jan 6 01:22:32 PST 2008
On 2008-1-6, at 7:17, ext Joe Touch wrote:
> Injong Rhee wrote:
>> I agree that the process could have been more diplomatic and it could
>> have been improved. However, what's done is done. Wide-scale tests in
>> some sense have been done for three to four years, as you said,
>> unknowingly. If you don't trust that as a test, I am puzzled on what
>> kind of wide-scale tests (with volunteers) we can do without causing
>> disruption on the production networks.
> For all you know there have been problems - either failure to provide
> benefit (which is otherwise harmless), or real problems - which you
> haven't bothered to detect or collect statistics on. Closing your eyes
> when you run an experiment doesn't declare it a success.
It's important to remember the two reasons for congestion control from
Sally's RFC2914: preventing congestion collapse and establishing some
degree of fairness.
CUBIC has been out there and the Internet hasn't collapsed. That's
good. It tells at that at least with the current mixture of traffic in
the Internet, CUBIC in concert with whatever other congestion control
is out there prevents collapse. (We don't know, however, if this
remains true under all possible conditions, though. But it's very
likely - preventing collapse is the relatively easy part of congestion
What the current unmonitored deployment of CUBIC isn't giving us is
data about how it affects fairness. If fairness-related issues (low
performance, stalls, etc.) happen to the CUBIC users, there is at
least a chance that they'll notice this and draw the right conclusions
- if they know they're using CUBIC. But if these issues happen to non-
CUBIC users that happen to share a path with CUBIC users, how are they
ever going to figure out that the problems they see are due to
somebody else using CUBIC? That's the sort of issues I worry about.
So while the current deployment of CUBIC gives us some useful data, it
doesn't give us all the data we'd need to evaluate safety.
(And I don't want to single out CUBIC here - the same is true for any
deployment of a congestion control variant.)
More information about the end2end-interest