Re: [alto] Adopting draft-kiesel-alto-xdom-disc as WG item?

["Mingming Chen" <[e-mail]>]

Hi all,
First, I agree with Kai :
"
>> I think whether to use the source/destination or to split up queries
>> depends on the application.  If it is selecting a destination for a
>> given source, it certainly can choose XDOM-DISC(SRC).  If the selection
>> is about sources, probably XDOM-DISC(DST) is more reasonable. 
"
Then, Xin's idea is reasonable for me. Suppose XDOM-DISC(X) just can get cost(X,*), an ALTO Client want to ask a reasonable cost(X,*) with a given X, it can achieve through XDOM-DISC(X), but when it want to ask a reasonable cost(*,Y) with a give Y, it should sent to all servers(except Y) to get the info, so I agree with Xin's here. XDOM-DISC(X) should register cost( X, *), cost(*,X), so it will save the query processor.


Besides that, I think we can reduce some extra cost information of XDOM-DISC(X) by training in a specific situation(It depends on different applications and scene). When some cost information has not been called for a long time, we can consider to remove them to simplify cost data store.


Then, a typo in 2.3, paragraph 2, line 3: and it is rooted in the in the special domain "IN-ADDR.ARPA."








At 2016-07-15 23:14:44, "Sebastian Kiesel" <[e-mail]> wrote:
>Hi Kai and all,
>
>On Fri, Jul 15, 2016 at 10:27:14AM +0800, Gao Kai wrote:
>> On 13/07/16 05:09, Sebastian Kiesel wrote:
>> > On Tue, Jul 12, 2016 at 02:55:52PM -0400, Wendy Roome wrote:
>> >> My comments on draft-kiesel-alto-xdom-disc-02:
>> >> Section 3.2:
>> >>
>> >> If a client wants the cost from X => Y, why just do cross-domain discovery
>> >> on address X? Why not do it for Y instead? Or do it for both X & Y?
>> > This is again to cope with the not so well-defined "routingcost" metric.
>> >
>> > Our assumtion is that XDOM-DISC(X) will discover one ALTO server,
>> > which will be able to express costs (X,Y) and costs (X,Z) using a
>> > (locally) consistent routingcost metric, i.e., with the meaning that
>> > a lower value indicates a higher preference.
>> >
>> > If we used XDOM-DISC(Y) and XDOM-DISC(Z) we would probably discover
>> > two servers, which could use completely incompatible definitions.
>> >
>> >
>> > In other words, the example query given in sec. 11.5.1.7. of RFC 7285
>> > should be sent to the server yielded from XDOM-DISC(192.0.2.2).
>> >
>> > If the "srcs" list contained more than one IP address, the query
>> > should be split up in several queries, each containing only one
>> > IP address in "srcs" (but keep all in "dsts"), and each of these
>> > "sub-queries" would need its own call to XDOM-DISC.
>>
>> I had the same question as Wendy because for me, sources and
>> destinations are somewhat symmetric 
>
>if src/dst and the costs between are really symmetric, it doesn't
>make a difference.
>
>> so it is tempting to think there
>> should be something like XDOM-DISC(X, Y), which might be overcomplicated
>> though.
>
>That would be too complicated indeed.
>
>
>> I think whether to use the source/destination or to split up queries
>> depends on the application.  If it is selecting a destination for a
>> given source, it certainly can choose XDOM-DISC(SRC).  If the selection
>> is about sources, probably XDOM-DISC(DST) is more reasonable. 
>
>I agree.  So the idea is, that we call XDOM-DISC(X), in oder to find an
>ALTO server that can answer both ECS queries, whether
>
>routingcost(src=X, dst=Y) > routingcost(src=X, dst=Z).
>
>and whether
>
>routingcost(src=Y, dst=X) > routingcost(src=Z, dst=X).
>
>
>In other words, we use the common address X that appears in all the
>tuples we want to compare as parameter for the XDOM-DISC.
>
>For the Endpoint Cost Service ECS, queries may have to be split in
>a way that either the "srcs" or the "dsts" list in the query contains
>only one IP address - and this address is used for XDOM-DISC.
>What that means for the other ALTO services we still need to define.
>
>> If the
>> application wants to pick the best source-destination pair, maybe the
>> queries should never be split up.
>
>Not splitting the queries would be the best thing to do from an
>optimization perspective.  It would be great if there were
>a bunch of ALTO servers all with the same knowledge, so we could
>just discvover any one of them (e.g. using RFC7286) and ask it
>arbitrary questions or download a comprehensive NxN cost map.
>But for various reasons I believe that this is not going to happen
>on an Internet scale (and after all, we are the Internet Engineering
>Task Force ;) ).  So we need to find a solution that works if knowledge
>is partitioned.
>
>Thanks
>Sebastian
>
>_______________________________________________
>alto mailing list
>[e-mail]
>https://www.ietf.org/mailman/listinfo/alto