On Saturday 10 May 2008 15:57, Michael Rogers wrote: > Matthew Toseland wrote: > > We could implement darknet sneakernet connections by exchanging USB sticks. > > E.g. if you meet somebody every day (e.g. a coworker), you could exchange > > (cheap) 8G sticks, plug them in overnight, and then do the same again the > > next day. This would produce approx 100K/sec (1Mbps) each way for each person > > you did it with. > > The speed is fine, but what about the latency? If a request travelled > for five hops the round-trip time would be ten days. We'd have to > reconsider things like redirects and splitfiles to minimise the number > of round trips.
Hence request priorities, so that the requests for the top blocks go over the UDP connections. > > > The main technical prerequisite is > > token passing load management, unless we implement a completely different > > load management system for it. > > Load management is hardly the biggest stumbling block. For example, with > a round-trip time of ten days and a bandwidth of 8 GB per peer per day > you're going to have a massive number of outstanding requests - far more > than you can hold in RAM. Hmm, I was assuming we already have long term requests. But strictly, we don't. We have ULPRs, which could be easily adapted to persist longer if the timeout is different. But even those are in-RAM-only. So the routing code could be very similar to the current code, but we would presumably have to store the stats in a database of some kind. Full passive requests will likely have some on-disk form anyway, because there are likely to be a lot of them, and they don't timeout (at least, they'd be renewable, not time out without asking the originator first). So they'd be very similar, but not identical. Even full passive requests don't have a routing stage - but that's not such a difficult adaptation. So we're talking about post-full-passive-requests most likely, i.e. probably Freenet 0.9. > > > I am of the view that these networks are mutually > > complementary and therefore should talk to each other: Darknet over UDP isn't > > safe in hostile environments, and off-grid darknets a) work much better if > > parts of them are online (certainly we could expect some covert wireless > > links in places, but being able to link to a functional on-grid darknet would > > surely be a benefit; long links are going to be rare on a pure off-grid > > darknet), and b) would be much easier to bootstrap from a working on-grid > > darknet. > > It might be a good idea for the sneakernet to support internet links, > but I'm not sure it's a good idea for sneakernet requests to take > opportunistic shortcuts across Freenet. They can't just be transformed > into Freenet requests when they cross into Freenet and transformed back > when they cross back: if there are any sneakernet links on the path then > the round-trip time will be on the order of days rather than seconds, so > the whole path needs to use appropriate timeouts and store the pending > request on disk rather than in RAM. That means the Freenet nodes on the > path need a sneakernet implementation even if they don't have any > sneakernet links. Again, request priorities. Maybe two completely different kinds of requests, fast (darknet only, no priority) and slow (can go over sneakernet or real-time darknet links, different sub-priorities, high latency). I did think about tiered routing a la NetworkManager, but it's not quite the same: The sneakernet network may be close to us, but have unacceptable latency (for some requests). The darknet network may be close to us, but unavailable because it's occupied by higher priority requests. So requests need to indicate their maximum latency somehow. > > Cheers, > Michael -------------- next part -------------- A non-text attachment was scrubbed... Name: not available Type: application/pgp-signature Size: 189 bytes Desc: not available URL: <https://emu.freenetproject.org/pipermail/devl/attachments/20080512/b72b6d06/attachment.pgp>
