On Wed, Aug 12, 2015 at 3:46 PM, Juliusz Chroboczek
<j...@pps.univ-paris-diderot.fr> wrote:
>> ECMP could be one component in this going forward. Why are you so
>> opposed to [t]his
>
> I'm not opposed to ECMP, quite the opposite, I think it would be a fun
> thing to implement. I'm trying to understand the use cases, since I have
> the policy of not implementing anything until it is requested by my users
> (fewer TLVs = better protocol).
>
> Alia's use case is different from yours, and the two may require different
> heuristics (Alia's: balance over any two non-interfering links,
> yours: balance over wired links only). So finding out what the realistic
> use cases are is necessary in order to get something done.
I am extremely busy with make-wifi-fast, but have two cents to add in
here, related to congestion and utilization, and
what I would call staged routing updates (please tell me a better
name), giving similar results to ecmp. instead of converging on one
best route, multiple "best" routes are chosen.
Consider a network with the following topology, all gigE.
A B C D
------------ (switch)
| |
E F
| |
---------- (switch)
G H I K
E and F are equal cost. In ecmp packets are basically FQ'd across E
and F, resulting in massive packet loss for all flows if either E or F
fails - or weird delays as one becomes overutilized.
If instead, A and B chose E, and C and D chose F, then we get flows
utilizing the full bandwidth available, with no possibility of
re-ordering (except on a route change), but no fq (so we can't get
twice the bandwidth for any given pair of hosts), which seems to me to
be more of the "right thing".
By "staged routes" I kind of mean - make a small route change, then
measure the result, then make more, until the links are as balanced as
possible. There is no single "best route" here, and indeed, the "best
route" for all stations is that they all choose a different E or F.
> -- Juliusz
>
> _______________________________________________
> homenet mailing list
> homenet@ietf.org
> https://www.ietf.org/mailman/listinfo/homenet
--
Dave Täht
worldwide bufferbloat report:
http://www.dslreports.com/speedtest/results/bufferbloat
And:
What will it take to vastly improve wifi for everyone?
https://plus.google.com/u/0/explore/makewififast
_______________________________________________
homenet mailing list
homenet@ietf.org
https://www.ietf.org/mailman/listinfo/homenet
