Subject: Re: connection bonding?
To: None <tech-net@NetBSD.org>
From: der Mouse <mouse@Rodents.Montreal.QC.CA>
Date: 12/08/2005 15:23:45
> der Mouse's $DAYJOB wants to spread a (single) TCP connection between
> a *single* pair of hosts ("devices"), over multiple links, as if
> those multiple links were a [single] fatter pipe to the [i.e.,
> single] peer device.
I'm not entirely sure about wanting to spread a single TCP connection.
The spec I got was not fully precise; I've started the process of
pushing queries back to the customer to find out more details.
I'm also not sure whether the other end is a host or a router or what.
Also possibly significant is that I find the links potentially being
aggregated are actually wireless.
> (OTOH, I do know of switches which can be persuaded to do round-robin
> scheduling across links in a link-aggregation group, but that's a
> definite case of caveat emptor.)
I'm not sure I'd want round-robin; it works well only if your traffic
is evenly distributed. If, for example, you have an even number of
interfaces and your traffic displays odd-even imbalances, you'll have
problems with RR. That's why I was thinking of shortest-queue. But
they all have reordering potential, unless you slap another layer above
it. Which appears to be close to what 802.3ad does, though it's an
almost stateless layer. (I really should follow that link and see if I
can't find a copy to read over to see what it really says. Gah, PDFs.)
/~\ The ASCII der Mouse
\ / Ribbon Campaign
X Against HTML email@example.com
/ \ Email! 7D C8 61 52 5D E7 2D 39 4E F1 31 3E E8 B3 27 4B