Re: [DNA] Definition of "Link Up" and "Link Down" events?

[Bernard Aboba <aboba@internaut.com>]

> > I agree in theory.  In practice I may implement to bring the link-down, if
> > it is not my only link-access, as far as IP stack is concerned.  Reason is
> > node/device may have multiple link-access mediums and this one is not
> > performing.  I think we all understand where I am going. My reason for
> > stating the obvious is does this context affect what we do within DNA?
>
> I think what we are doing is coming up with definitions of the concepts
> of link up/down that are useful for DNA, and does not prevent things
> like "link quality" to be taken into account, especially for hosts with
> multiple interfaces.

"Link-aware" routing metrics date back to the original ARPAnet packet
radio experiments, and more recent work has established the importance of
taking link quality into account.  For example, see:

http://research.microsoft.com/mesh/papers/metrics.pdf
http://pdos.csail.mit.edu/~rtm/papers/etx.pdf
http://pdos.csail.mit.edu/papers/grid:losstr02/paper.pdf
http://www.ee.ucla.edu/~kulkarni/papers/Routing.pdf

> It is probably the case that a host with multiple interfaces will work
> best if it takes into account a number of factors (from packet loss
> rates and bandwidth, to monetary cost for using a NIC and application
> requirements) when choosing which NIC to use (and for what application
> traffic).

All this can be taken into account by adjusting routing metrics.  In
particular, routing metrics can automatically make a decision about
whether to route packets over another interface once the link quality
begins to deteriorate.

> I also suspect that such strategies doesn't require that DNA
> stop considering an interface. For instance, if the stack+applications
> decide to use the GPRS interface for VoIP traffic, it is probably still
> useful to have DNA monitor the 802.11 interface to detect when the host
> gets back in range of the AP.

Routing metrics only become relevant when an interface is "up" and
configured for communicating IP packets.  DNA helps speed up that process,
but it isn't involved in making the decision whether to enable an
interface or not.  Similarly, bringing an interface down is a separate
decision from deciding whether to use it to route a particular packet.

However, I do think that decisions on enabling/disabling interfaces do
need to take routing metrics into account, or else inconsistencies can
result.  For example, if you bring an interface up that has a higher
routing metric than another interface, it won't be used except perhaps if
it can reach prefixes that the other interface can't.  If "idle timeout"
is in effect, then the unused interface will only be brought down again.

Similarly, if "dial on demand" is in effect, then if you bring an
interface down that had a lower routing metric than an existing interface,
then it will only be brought up again.

So my take is that routing metrics are actually fairly fundamental to
multi-homing behavior.  For example, MIT LCS has an interesting paper
on rate negotiation algorithms:
http://pdos.csail.mit.edu/~rtm/srcrr-draft.pdf

> But that clearly shouldn't prevent implementations that just use simpler
> strategies such as "haven't heard from the AP in 30 seconds" to just
> abandon the use of the 802.11 interface.

That would be a particularly poor algorithm to use since the detection
time is larger than RTOmin.  In general, a NIC should probably start
scanning after only 3 retransmissions (unlikely to occur merely due to
congestion) and should consider moving to another AP before RTOmin.  For
example, even poor rate negotiation implementations typically can move
from the highest bit rate to the lowest (typically 1 Mbps) in 200 ms.

> I don't see the definitions that Bernard updates as restricting the
> implementation in any way; could be as simple or as complicated as it
> wants when handling multiple NICs.

I don't see them restricting an intelligent implementation either.  But
I've seen enough bad ones that I'd prefer to rule out the most egregious
behavior.