On Feb 1, 2012, at 5:50 PM, Sebastian Kiesel wrote: > Bill, all, > > On Wed, Feb 01, 2012 at 10:07:58AM -0500, Bill Roome wrote: >> I guess I'm also late to this discussion. But I always assumed: >> >> 1. The ALTO costs are end-to-end. Always. > > ack. > >> 2. An ALTO server is expected to give the costs between *every* pair of >> PIDs. It can omit a cost only if it cannot determine that cost at all. > > My assumption was different:
the differences in assumptions tend to demonstrate we would need to be more precise on the semantic. > I'd say that the protocol specification > must allow the ALTO server to return a full NxN matrix indicating all > end-to-end costs between all N PIDs (maybe even different costs in the > two directions between two given PIDs). > > But I also assume that in a real deployment we will rather have a very > sparse matrix, with the default value "I don't know / I am not allowed > to tell". I could even imagine that my ISP would operate an ALTO server, > which gives my hosts only an 1xN vector, which tells me how "expensive" > it is for me to reach all possible destinations in the Internet. > >> 3. You cannot infer a topology from the 'holes' in the cost map. > > ack. > >> 4. If it were possible to calculate the missing costs by using SPF (or >> whatever) on the other costs, then the ALTO server should do that! The >> ALTO server shouldn't push that task onto the clients. > > ack. > > > Btw: which application needs to know the topology? E.g.: a CDN may want to operate autonomously with maps rather than using ECS. Not that I'm advocating this but it all depends on where do you want to place the complexity of topology hint: inside ALTO server (ECS like of approach) or inside the application (Maps that allows you to compute paths/trees). > If I remember > correctly once upon a time our primary goal was assisting peer selection > (i.e., I know that I can get the desired resource form IP address A, B, > or C, so please ALTO service tell me which address I should give a > connect()-try first). To that end, the 1xN vector would be enough. The CDN is not substantially different in its logic. s. > > -- 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
