On Mon, Sep 26, 2011 at 7:46 AM, Sebastian Kiesel <[email protected]> wrote: > On Sat, Sep 24, 2011 at 11:03:00AM -0700, Richard Alimi wrote: >> On Tue, Sep 13, 2011 at 1:49 PM, Sebastian Kiesel <[email protected]> >> wrote: >> > On Tue, Sep 13, 2011 at 10:35:13AM -0400, Bill Roome wrote: >> >> How many PIDs do we expect an ALTO server to handle? >> >> >> >> I ask because I've discovered that code that works fine with a few hundred >> >> pids dies when I go to 5,000 pids. >> >> >> >> The biggest problem is the full cost map response message. With 5,000 >> >> pids, that has 25 million cost entries. >> > >> > I'm wondering whether retrieving the full N x N cost map is an >> > important (or even a realistic?) use case. >> >> Yes I do believe it is important and realistic. If someone is going >> to be doing post-processing on the map to optimize for a a network as >> a whole, then this would be the desired interface to use. > > I have no doubts that there would be interesting things that could > be done with the full map (in particular if we think about a tracker). > > The question for me is more: will a single organization be able to > gather the data? That depends on the metric / rating criterion: > latencies or AS paths from the point of view of the ALTO server can > easily be found. For other criteria, such as (normalized) monetary > costs for transmitting traffic from one PID to another, ISPs would have > to cooperate and disclose the data, I'm not sure this will ever happen.
Okay - without detracting from the original comment, it sounds like we both agree that the NxN matrix can be a reasonable request for an ALTO Client to make. I agree that the quality of information (and whether the ALTO Server provides certain cost types for addresses outside of its administrative domain) is going to vary depending on many factors. > >> > Consider an ALTO client embedded into a (P2P) application running on a >> > PC connected to ISP A, say in PID 1. It makes sense to ask ISP A's ALTO >> > server how expensive it is to get from PID 1 to candidate peers located >> > in PIDs 2 (other access network of ISP A), 3 (ISP B) and 4 (ISP C), >> > respectively. That is, it would need to know an 1 x N vector. >> > But why should the application be interested in knowing the cost from >> > PID 3 to 4? And even if it was interested, how would ISP A know the cost >> > between ISP B and C's networks? >> >> Yes - some of the feedback from IETF81 was to give an ALTO Server a >> way to say "I don't know" (one of the ways to do this might be to just >> omit such entries from the cost map, but that's another thread so I >> won't comment on it more here). > > So we need an efficient encoding for a sparse matrix ... Yep - saving that for the other thread :) > >> > If, on the other hand, we consider a big resource directory (e.g. P2P >> > tracker) it would make sense to know the N x N matrix. >> >> Agreed. >> >> > But I doubt >> > that there would be a single ALTO server that can deliver this matrix >> > with a high level of detail. >> >> Why not? I think I would disagree with this, depending on what you >> mean by "a high level of detail". An ALTO Server does not need to >> reply based on information from within other ISPs. The semantics of >> ALTO maps and information in general is that it is from the ALTO >> service provider's perspective. > > Agreed. But not sure how useful is information from the point of view > of the ALTO server. FWIW, I fully expect there to be ALTO Clients out there who will crawl all publicly-available ALTO Servers and do data mining on that aggregated set. iPlane is an example of gathering data on a global scale without ALTO, and I would imagine that ALTO would only serve as another helpful input. But that is a separate question from the size of a map of a single ALTO Server. Rich > > -- S. > _______________________________________________ alto mailing list [email protected] https://www.ietf.org/mailman/listinfo/alto
