Hi Wendy, this is very true! But the question is, from where does the tracker get the 5000x5000 matrix, if there is no single ALTO server that provides it?
One answer may be that the tracker will have to query multiple ALTO servers and assemble the map on its own. (That's why I am working on and lobbying for cross-domain discovery alias 3pdisc for quite a long time...) But in this case the maps pulled from the individual ALTO servers would be smaller, and the tracker could control the update intervals individually. Sebastian On Fri, Jul 11, 2014 at 01:44:54PM -0400, Wendy Roome wrote: > I don¹t know whether any ALTO server could maintain a full 5000x5000 cost > matrix, but I can give you a very reasonable example of a client who like > one. > > Consider a large P2P tracker. It has thousands of peers for each file, > scattered over the planet. It gets lots of requests from peers, who are > also scattered all over the planet. When a peer in New York asks for > peers, the tracker needs the costs from New York to every other peer. And > the tracker needs detailed local costs for the New York area: for a New > York peer, Boston is better than Chicago, and Chicago is better than San > Francisco. But to a New York peer, Tokyo & Kyoto are the same. > > But when the tracker gets a request from a Tokyo peer, the tracker needs > detailed local cost data for Tokyo & Japan. For that peer, Tokyo is better > than Kyoto and Boston and New York are the same. > > So a large P2P tracker love to have a fully populated 5k x 5k cost matrix. > > - Wendy > > > On 07/11/2014, 05:30, "Sebastian Kiesel" <[email protected]> wrote: > > >On Fri, Jul 11, 2014 at 09:24:21AM +0000, Scharf, Michael (Michael) wrote: > >> > On Wed, Jul 09, 2014 at 03:14:56PM -0400, Wendy Roome wrote: > >>> > What do I mean by "large"? A Network Map with 5,000 PIDs, 250,000 > >> > > prefixes, and up to 25,000,000 cost points. > >> > > >> > A full 5000x5000 cost map would be in the order of 130 MB gzipped json > >> > (see "The size of the cost map" thread on the ALTO list Fri, 22 Mar > >> > 2013). > >> > > >> > do we have any idea whether 5000 PIDs is a realistic assumption for > >> > foreseeable deployments? > >> > >> Indeed, 5000 seems like an extreme case. Yet, I think one could really > >> end up there, depending on how ALTO is used. > >> > >> For instance, let's assume an ALTO use case like > >> draft-scharf-alto-vpn-service for a large enterprise with many > >> branches. As argued in the draft, only could possibly use one PID for > >> each VPN endpoint. > >> > >> In that case, 5000 is not entirely unrealistic. As a random data > >> point, Bank of America is reported to have 5,377 branches in the US. > >> Other large organizations (e.g., retailers) may have a similar order > >> of magnitude of branches. A L2VPN or L3VPN could be used to > >> interconnect such branches. > >> > >> Obviously, there are certain ways to reduce the network/cost size in > >> such a scenario, e.g., by introducing a hierarchy, or by topology > >> encoding (for cost map). Also, I doubt that an application would > >> really require an accurate cost map entry for any given combination of > >> PIDs. Thus, 5000 looks like a kind of worst-case scenario. > > > >In this use case, would you really have a full(!) 5000x5000 matrix? > >Or would it be more like a sparse matrix, defining cost only for > >5000 branch offices X 100 data centers, > >while most (all?) other costs (i.e., branch office to branch office) are > >"default"? > > > >Thanks, > >Sebastian > > _______________________________________________ alto mailing list [email protected] https://www.ietf.org/mailman/listinfo/alto
