Re: [Roll] reliability metrics for digital radios

[JP Vasseur <jvasseur at cisco.com>]

Hi Richard,

On Oct 12, 2009, at 4:26 PM, Richard Kelsey wrote:

> JP,
>
>   From: JP Vasseur <jvasseur at cisco.com>
>   Date: Mon, 12 Oct 2009 14:46:18 +0200
>
> > Below is a discourse on the reliability characteristics of
> > 802.15.4 and similar digital radios and what this means for
> > link metrics.  Even if the goal is to maximize something
> > other than reliability, you still need the links to be
> > reasonably reliable in order to reduce churn and because no
> > matter what the metric, broken routes have little utility.
>
>   Yes, just to make sure that we are in line, there are two
>   orthogonal aspects here:
>   1) Link UP/DOWN: if indeed, the link quality crosses some threshold
>   it can be put in DOWN state locally by the node
>   2) If the Link stays in DOWN state for some period of time, the
>   node may decide to update the metrics in its RA-DIO.
>
> My point is that the key metric for reliability is not the
> RTX for a message right now, or any similar metric, but an
> estimate of the stability of the link itself.  We need to
> preferentially use stable links.
>
> This is independent of what a node does when a link actually
> does stop working.

Completely agreeing with you.

> > Non-radio links, or other types of radios, will have very
> > different characteristics.  This may make it difficult to
> > have a good relibility metric that is independent of the
> > link layer.
>
>   Yes but ... RPL is L3 thus we want path to be calculated
>   according the link reliability (and I think that we all
>   agree on this requirement), this is what we need to
>   do. And I agree, this is not easy, the idea is to have a
>   good enough way to abstract the link reliability using a
>   single metric. The way to compute that metric is
>   implementation specific and L2 dependent, which is fine.
>
> There may not be a single metric that we can use.  In
> particular, for 802.15.4 radios, both stability and
> latency/power must be taken into account.  We have to
> strongly prefer links that are stable (to avoid churn) and
> links between always-on devices (to minimize battery use and
> latency).  These may be specific to 802.15.4, but they are
> the overriding concerns for those networks.  I do not see
> how they can be collapsed into a single value.  A single
> metric is possible, but it will likely have to contain
> distinct stability and power/latency values.

Again agreeing ... and the metric ID is specifying a set of metrics.  
In any case, applications
will require different metric according to the OF and an OF may  
require multiple metrics.

>   Should we try to use the ETX and couple it with filtering
>   to get the result of distinguishing OK links with Good
>   links ?
>
> I do not see how ETX can capture the desired information.
> ETX averaged over the short term can't measure long term
> stability.  ETX averaged over the long term can't tell
> you which links are currently usable.  I am not sure what
> you mean by 'filtering' here.

I meant two things:
1) There are many ways to compute the ETX (implementation specific)
2) Short transient issues may not have to be captured by the ETX: some  
transient
events may be filtered out by a node.

What is your preference in term of reliability metric ?

Thanks.

JP.

>                              -Richard Kelsey
> _______________________________________________
> Roll mailing list
> Roll at ietf.org
> https://www.ietf.org/mailman/listinfo/roll