Hi Wendy and Sebastian,

Please see inline,
Sabine

>>-----Message d'origine-----
>>De : alto [mailto:[email protected]] De la part de Sebastian Kiesel
>>Envoyé : jeudi 22 septembre 2016 18:55
>>À : Roome, Wendy (Nokia - US)
>>Cc : IETF ALTO
>>Objet : Re: [alto] cost metrics vs cost modes
>>
>>On Mon, Aug 01, 2016 at 11:11:42AM -0400, Wendy Roome wrote:
>>> In the early days of ALTO, we decided that cost-metrics and cost-
>>modes
>>> were orthogonal: any metric could be presented in either mode. We
>>were
>>> able to get away with that when routingcost was the only metric,
>>> because routingcost values are totally arbitrary, with no intrinsic
>>meaning.
>>>
>>> But now we are considering new metrics, like delay and bandwidth.
>>> These have intrinsic meaning. If a client wants a server with a
>>> latency of less than (say) 50 ms or an available bandwidth of more
>>> than 1 mb/sec, that client *must* have a numerical-mode delay. An
>>> ordinal-mode delay is useless.
>>
>>Couldn't there be a case where a client says "I want to talk to the top
>>three of the list of candidate peers, ordered by latency,
>>irrespectively what the absolute latency is"?
>>
>>>
>>> Contrast that with routingcost: most clients just pick the server
>>with
>>> the lowest cost, so ordinal-mode and numerical-mode are
>>interchangeable.
>>>
>>> So here is my suggestion: rather than being orthogonal to the metric,
>>> the mode is really *part* of the metric. That is, a metric is defined
>>> as either numericsl or ordinal. Or absolute or relative, if you
>>prefer.
>>>
>>> Thus we would have the metrics:
>>>
>>> routingcost: A numerical cost. Scalable, proportional, comparable
>>> between queries.
>>>
>>> ord-routingcost: A ranked cost. The values can be compared to others
>>> in this response, but are not scalable or proportional, and cannot be
>>> compared to costs from previous queries.
>>>
>>> hopcount: The actual hop count.
>>>
>>> ord-hopcount: A ranking that tracks the actual hop count for
>>(src,dst)
>>> pairs in this query. Useful for finding the pair with the lowest
>>> hopcount, but not for estimating the actual number of hops.
>>>
>>> delay: The actual delay, in seconds (or millisec, or whatever)
>>>
>>> ord-delay: A ranking that tracks the delay, without revealing the
>>> actual delay. If, indeed, this metric is useful at all.
>>
>>Isn't that just concatenating the two fields?
>>
>>Is there really a metric where the ord-xxx does not make sense at all?
>>And is that such a big problem that we should change the spec (maybe we
>>should have done that differently, but is it worth changing now?).
>>
[SR     ] Agree. Many metrics provided in numerical modes are likely to be 
requested in ordinal mode. For instance a client may ask : "as long as their 
path bandwidth > xx mb/sec, rank those k servers". This may apply to most of 
the ALTO performance metrics. 
Other potential modes such as "Boolean", "string", "path-vector", are often 
uniquely associated to a metric. In those cases indeed, requests and responses 
with just "cost-name" : "mode-metric" would be lighter. 

However how do we ensure backwards compatibility with RFC7285 ? 


>>
>>Sebastian
>>
>>_______________________________________________
>>alto mailing list
>>[email protected]
>>https://www.ietf.org/mailman/listinfo/alto

_______________________________________________
alto mailing list
[email protected]
https://www.ietf.org/mailman/listinfo/alto

Reply via email to