Re: [Roll] reliability metrics for digital radios

[JP Vasseur <jvasseur at cisco.com>]

On Oct 14, 2009, at 9:48 PM, Jonathan Hui wrote:

> On Oct 14, 2009, at 6:13 PM, Philip Levis wrote:
>
> > Right -- the assumption is that the real link estimation is with  
> > packets. Receiving a packet with a very high chip correlation index  
> > means that the link *at that time* had a high SNR, meaning it is  
> > likely a good enough link to try (white bit set). If the white bit  
> > isn't set, then a link has to go through a little bit of packet- 
> > based estimation before CTP Noe will try using it.
>
> My experience with chip correlation is that it can be a leading  
> indicator of packet loss, but it's ability to discern 'good' links  
> from 'okay' links is limited.  In the end, I think a 'link  
> classifier' should use whatever information it can get - that leads  
> me to want a metric that doesn't really have a 'unit' but instead  
> identifies a 'class'.

That is the flag approach that I was referring to. Classification  
should then be left to implementation.

> > I agree that it's reasonable for an implementation to heavily favor  
> > "good" links over "okay" links (tradeoff potential efficiency for  
> > stability), or discard the use of okay links altogether. However,  
> > user/application preference of behavior in the presence of only OK  
> > links should be left out of the protocol specification. Some  
> > applications would prefer "don't work unreliably" and others would  
> > prefer "do your best."
>
> Of course.  I just wanted to say that there are times when you want  
> to ignore and other times you don't want to ignore 'okay' links.

True. Should be dictated by circumstances and OF.

Cheers.

JP.

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