Re: [patch] TCP/IP delacks disabled with MPI

George Bonser (grep@shorelink.com)
Tue, 1 Jun 1999 13:35:40 -0700 (PDT)


On Tue, 1 Jun 1999 kuznet@ms2.inr.ac.ru wrote:

> Hello!
>
> > Channel bonding (or Trunking or EtherChannel) support is something I would
> > like to see too. At least the addition of the IFF_RUNNING stuff that was
> > talked about before so it will fail over peroperly.
>
> Please, remind me what was talked about IFF_RUNNING?
>
> BTW you may try teql to trunk ethernets. It is a bit more
> expensive than beowulf bonding, but more clever yet.
>
> Alexey
>

The problem is that the channel bonding will notice that a link has gone
bad only if its xmit buffer fills. Many cards will never do this because
they will simply throw away any data they get if the link is down. This
results in the channel bonding round-robin to continue using this
interface resulting in 1/N packet loss where N is the number of
interface cards bonded. If you have two cards, you get 50% packet loss
forever. THe goal was to have the round-robin check for IFF_RUNNING
instead of the xmit buffer but the card drivers have to be taught how to
set it. Or at least that was my understanding of the problem.

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.rutgers.edu
Please read the FAQ at http://www.tux.org/lkml/