I think that there are actually two distinct though related issues going on here, which are downtime and polling.
Downtime means that the data you want is in the network, but you can't get to it right now because one of the nodes that the request needs to go through isn't up right now (e.g. a gateway out of a darknet). Polling means that the data you want isn't in the network yet, but you'd like to get it sometime in the future when it is inserted. DOWNTIME The downtime problem is related to delay-tolerant networking - http://en.wikipedia.org/wiki/Delay_Tolerant_Networking http://www.dtnrg.org/wiki One way to think about it is to think about routing in time as well as space. Suppose you want to send a message along the route A-B-C, but A-B is only up on Monday and B-C is only up on Tuesday. You can complete the route by holding the message at B for a day - effectively 'routing' between B_on_Monday and B_on_Tuesday. The practical implication is that if your first-choice destination is down, sometimes it may be better to hold the request for a while and wait for it to come back up, rather than immediately rerouting to your second-choice destination. Or, you route to the second-choice destination, but if it fails, you hold onto it. When the first-choice destination comes back up, it triggers a retry - a sort of "active" passive request. (A terrible term - perhaps "persistent request" or "delay-tolerant request"?) POLLING Polling, on the other hand, is related to publish-subscribe: http://en.wikipedia.org/wiki/Publish/subscribe You set up a subscription that registers your interest in some key(s), together with a route showing where to find you later on. When the data is later inserted, it gets pushed to you. Here, the trigger event is data arriving, rather than a node coming up. There is no further forward routing, only sending data back along an already-existing route. If the route to the subscriber no longer exists, that's another problem, but in this case I think we should just drop it rather than try to find the subscriber. If s/he really wants it, s/he can just request it again. Thoughts? theo 2008/7/14 Matthew Toseland <toad at amphibian.dyndns.org>: > 1. Many people have proposed over the years that we have a "bulk" flag which > can be set when the timing of a request is less important (e.g. for splitfile > fetches), or a priority class for a request which is visible at the network > layer. I have always opposed this mostly because it makes traffic profiling > slightly easier and any sort of priority scheme would need careful regulation > to prevent race-to-the-top. > > 2. Long-term, and in particularly nasty places, Freenet will have to be mostly > darknet, because it is much easier to attack opennet nodes, or to block them > in bulk. One of the biggest practical problems with a pure darknet is the > 24/7 issue: more people have laptops than have real PCs nowadays, and this > trend is likely to continue and accelerate, but even if people have a desktop > PC, many users won't run it 24x7 for various reasons: power consumption, > noise, security (with encrypted disk, do you want to leave it unattended?), > etc etc. Fanless home server appliances might be able to run 24x7, but that > means additional expenditure to buy them. > > 3. FMS, even more than Frost, makes heavy use of SSK polling, and this is > likely to expand as the network grows and FMS becomes more newbie friendly. > Also various innovative applications require fast propagation of data once > inserted (although there are frequently security issues with this). And > widely-wanted data which is hard to find can be effectively polled by much of > the network, causing excessive load. > > 4. The solution to SSK polling etc is some form of passive requests. In 0.7, > we have ultra-lightweight passive requests, which are a very limited and > unreliable mechanism but nonetheless should help significantly. The basic > principle of ULPRs is that once a request completes, each node on the network > remembers who wants the data and who it has asked for it, for a short time, > without making any effort to reroute if connections are lost; if the data is > found it is propagated quickly to everyone who wants it. > > 5. True passive requests (0.9) would be a mechanism whereby a node could send > out a request, which once it failed would be remembered permanently, subject > to a (long) timeout and/or periodic renewal from the originator. It would be > automatically rerouted if the network topology changes. Passive requests > would introduce a number of new technical challenges such as load management > for persistent requests, evaluating a peer's competence in performing them, > and so on, but they could greatly reduce the cost of SSK polling, > rerequesting common but absent data, and enable such things as medium > bandwidth high latency publish/subscribe for for example audio streams. > Passive requests would probably have to have a priority level setting. It's a > big job, but a big prize... > > 6. Passive requests would go a long way to solving the uptime problem. Say you > have a small darknet, say 5 nodes. Its nodes are only online during evenings > local time. Its only connection to the outside world is through one node > which is connected to two of the small darknet, which is only online on > Thursdays. Right now, except on Thursdays, the network would be essentially a > leaf network: our real-time routing assumes that the network is fully > connected. Most data will be very difficult to obtain. Real-time routing > requires real-time load balancing, which means that all the nodes would > request whatever it is they want constantly, generating load to no good > purpose, except on Thursdays when the requests would get through, but > severely limited by load management, and by the fact that more than one of > the small darknet may be asking for the same file. So on Thursdays, some > progress would be made, but often not very much. > > Now, with true passive requests, things can be very different. From the user's > point of view the semantics are essentially the same: they click a link, it > gets a DNF (fairly quickly), and they click the button to queue it to the > global queue; some time later, they get a notification that the content is > available. But performance could be much higher. If a node requests a block > while the network is "offline", the request will propagate to all 5 nodes, > and then sit there waiting for something to happen. When we connect to the > wider network, the request is immediately rerouted to the node that just > connected (either because it's a better route, or because there are spare > hops). It propagates, fails, and is stored as a passive request on the wider > network, hopefully reaching somewhere near the optimal node for the key. When > the link is lost, both sides remember the other, so when/if the data is found > on the wider network, it is propagated back to the originator. Furthermore, > the load management would be optimised for passive requests: when the small > network connects, it can immediately send a large number of passive requests > for different blocks of the same file or for different files. These are not > real-time requests, because they have already failed and turned into passive > requests; so they can be trickled out at whatever rate the recipient sees > fit. Also, they are not subject to the anti-polling measures we have > introduced: Polling a key in 0.7 means requesting it 3 times, sleeping for > half an hour, and repeating ad infinitum. Further similar measures may need > to be introduced at the node level to try to deal with increasing load caused > by FMS, but because we reroute on getting a connection, we can immediately > route the requests. When we reconnect, hopefully our peer will have found > most of the data we requested and can transfer it at link speed (or whatever > limit may be imposed for security reasons). The transfer might take longer > than the intersection, but I expect the whole system will be significantly > faster than it would be now. It's even better if you have more than two > network fragments: on a large darknet you might have subnetworks coming > online and going offline constantly, so that you never actually have a fully > online network. Passive requests would happily search out every relevant nook > and cranny of the network. > > Note that much of this is only feasible on darknet, because of the trust > connection: on opennet, passive requests probably will have to last only as > long as the connection is open, and bulk transfer of passive requests is > certainly not feasible on opennet. > > With regards to security, it may be possible to determine whether an FMS > poster (for example) is on the local network, if you know when his posts come > in. This is of course feasible now on such a topology, but on the other hand > if nobody uses it because it's unusable, there's no threat. Passive requests > would probably make it a little easier. Some form of tunneling, preferably > with long client controlled delays for inserts, might help to solve it, but > we would have to have a way of determining that the network is too small to > provide useful anonymity. > > 7. Even longer term, many ISPs and countries may deploy traffic flow analysis > hardware to identify and block all (unlicensed?) peer to peer networking. The > only way to beat traffic flow analysis is to not send data continually over > the internet. The obvious ways to do this are: > 1. Parasitic transports: Steal the video stream in a VoIP call to a friend. > Note that if VoIP calls are rare this won't work well, and if they are > artificially common to speed downloads up that will probably be detectable. > 2. Fake timing: Make the transports look like e.g. a private gaming server, > and fake the timing based on statistical models. This is classic stego. It's > a race between you and your opponent for whoever has the better model. Given > that your opponent may store traffic data indefinitely and not act > immediately, this is very dangerous... > 3. Wifi etc, non-internet constant data transfer. > 4. Sneakernet and physical-rendezvous-based protocols (the latter working on > the Freenet threat model assumptions, so still a form of darknet, rather than > Haggle's free-for-all-networking system which is also interesting but IMHO > dangerous in the long term, and certainly isn't Freenet). > > Passive requests are again the right tool to deal with this IMHO. They lend > themselves to efficient stream subscriptions, and also enable long-term > downloads without assuming a fully connected network at all times. Of course > for the very high latency options (sneakernet), there are other challenges, > such as how to assign locations without being able to swap continually. But > for medium latency caused by sparse darknets not often having many nodes > online simultaneously, and for transports which have the same effect despite > nodes' host computers actually being switched on, it should work well. > > Comments? Am I spouting nonsense? :) Apologies for the length of this mail, > it's a somewhat complex subject! > > Given the enormous implications, maybe we should postpone passive requests to > post 1.0 ... but I'm worried that FMS, and polling in general, may force our > hand. It's likely that there will be different parts of it implemented over > different versions of Freenet in any case... > > _______________________________________________ > Tech mailing list > Tech at freenetproject.org > http://emu.freenetproject.org/cgi-bin/mailman/listinfo/tech >
