NOTE: The following is a discussion of the history of netdb implementation and is not current information. See the main netdb page for current documentation.

History

The netDb is distributed with a simple technique called "floodfill". Long ago, the netDb also used the Kademlia DHT as a fallback algorithm. However, it did not work well in our application, and it was completely disabled in release 0.6.1.20.

(Adapted from a post by jrandom in the old Syndie, Nov. 26, 2005)
The floodfill netDb is really just a simple and perhaps temporary measure, using the simplest possible algorithm - send the data to a peer in the floodfill netDb, wait 10 seconds, pick a random peer in the netDb and ask them for the entry to be sent, verifying its proper insertion / distribution. If the verification peer doesn't reply, or they don't have the entry, the sender repeats the process. When the peer in the floodfill netDb receives a netDb store from a peer not in the floodfill netDb, they send it to all of the peers in the floodfill netDb.

At one point, the Kademlia search/store functionality was still in place. The peers considered the floodfill peers as always being 'closer' to every key than any peer not participating in the netDb. We fell back on the Kademlia netDb if the floodfill peers fail for some reason or another. However, Kademlia was then disabled completely (see below).

More recently, Kademlia was partially reintroduced in late 2009, as a way to limit the size of the netdb each floodfill router must store.

The Introduction of the Floodfill Algorithm

Floodfill was introduced in release 0.6.0.4, keeping Kademlia as a backup algorithm.

(Adapted from posts by jrandom in the old Syndie, Nov. 26, 2005)
As I've often said, I'm not particularly bound to any specific technology - what matters to me is what will get results. While I've been working through various netDb ideas over the last few years, the issues we've faced in the last few weeks have brought some of them to a head. On the live net, with the netDb redundancy factor set to 4 peers (meaning we keep sending an entry to new peers until 4 of them confirm that they've got it) and the per-peer timeout set to 4 times that peer's average reply time, we're still getting an average of 40-60 peers sent to before 4 ACK the store. That means sending 36-56 times as many messages as should go out, each using tunnels and thereby crossing 2-4 links. Even further, that value is heavily skewed, as the average number of peers sent to in a 'failed' store (meaning less than 4 people ACKed the message after 60 seconds of sending messages out) was in the 130-160 peers range.

This is insane, especially for a network with only perhaps 250 peers on it.

The simplest answer is to say "well, duh jrandom, it's broken. fix it", but that doesn't quite get to the core of the issue. In line with another current effort, it's likely that we have a substantial number of network issues due to restricted routes - peers who cannot talk with some other peers, often due to NAT or firewall issues. If, say, the K peers closest to a particular netDb entry are behind a 'restricted route' such that the netDb store message could reach them but some other peer's netDb lookup message could not, that entry would be essentially unreachable. Following down those lines a bit further and taking into consideration the fact that some restricted routes will be created with hostile intent, its clear that we're going to have to look closer into a long term netDb solution.

There are a few alternatives, but two worth mentioning in particular. The first is to simply run the netDb as a Kademlia DHT using a subset of the full network, where all of those peers are externally reachable. Peers who are not participating in the netDb still query those peers but they don't receive unsolicited netDb store or lookup messages. Participation in the netDb would be both self-selecting and user-eliminating - routers would choose whether to publish a flag in their routerInfo stating whether they want to participate while each router chooses which peers it wants to treat as part of the netDb (peers who publish that flag but who never give any useful data would be ignored, essentially eliminating them from the netDb).

Another alternative is a blast from the past, going back to the DTSTTCPW (Do The Simplest Thing That Could Possibly Work) mentality - a floodfill netDb, but like the alternative above, using only a subset of the full network. When a user wants to publish an entry into the floodfill netDb, they simply send it to one of the participating routers, wait for an ACK, and then 30 seconds later, query another random participant in the floodfill netDb to verify that it was properly distributed. If it was, great, and if it wasn't, just repeat the process. When a floodfill router receives a netDb store, they ACK immediately and queue off the netDb store to all of its known netDb peers. When a floodfill router receives a netDb lookup, if they have the data, they reply with it, but if they don't, they reply with the hashes for, say, 20 other peers in the floodfill netDb.

Looking at it from a network economics perspective, the floodfill netDb is quite similar to the original broadcast netDb, except the cost for publishing an entry is borne mostly by peers in the netDb, rather than by the publisher. Fleshing this out a bit further and treating the netDb like a blackbox, we can see the total bandwidth required by the netDb to be:

  recvKBps = N * (L + 1) * (1 + F) * (1 + R) * S / T
where
  N = number of routers in the entire network
  L = average number of client destinations on each router 
      (+1 for the routerInfo)
  F = tunnel failure percentage 
  R = tunnel rebuild period, as a fraction of the tunnel lifetime
  S = average netDb entry size
  T = tunnel lifetime
Plugging in a few values:
  recvKBps = 1000 * (5 + 1) * (1 + 0.05) * (1 + 0.2) * 2KB / 10m
           = 25.2KBps
That, in turn, scales linearly with N (at 100,000 peers, the netDb must be able to handle netDb store messages totaling 2.5MBps, or, at 300 peers, 7.6KBps).

While the floodfill netDb would have each netDb participant receiving only a small fraction of the client generated netDb stores directly, they would all receive all entries eventually, so all of their links should be capable of handling the full recvKBps. In turn, they'll all need to send (recvKBps/sizeof(netDb)) * (sizeof(netDb)-1) to keep the other peers in sync.

A floodfill netDb would not require either tunnel routing for netDb operation or any special selection as to which entries it can answer 'safely', as the basic assumption is that they are all storing everything. Oh, and with regards to the netDb disk usage required, its still fairly trivial for any modern machine, requiring around 11MB for every 1000 peers (N * (L + 1) * S).

The Kademlia netDb would cut down on these numbers, ideally bringing them to K over M times their value, with K = the redundancy factor and M being the number of routers in the netDb (e.g. 5/100, giving a recvKBps of 126KBps and 536MB at 100,000 routers). The downside of the Kademlia netDb though is the increased complexity of safe operation in a hostile environment.

What I'm thinking about now is to simply implement and deploy a floodfill netDb in our existing live network, letting peers who want to use it pick out other peers who are flagged as members and query them instead of querying the traditional Kademlia netDb peers. The bandwidth and disk requirements at this stage are trivial enough (7.6KBps and 3MB disk space) and it will remove the netDb entirely from the debugging plan - issues that remain to be addressed will be caused by something unrelated to the netDb.

How would peers be chosen to publish that flag saying they are a part of the floodfill netDb? At the beginning, it could be done manually as an advanced config option (ignored if the router is not able to verify its external reachability). If too many peers set that flag, how do the netDb participants pick which ones to eject? Again, at the beginning it could be done manually as an advanced config option (after dropping peers which are unreachable). How do we avoid netDb partitioning? By having the routers verify that the netDb is doing the flood fill properly by querying K random netDb peers. How do routers not participating in the netDb discover new routers to tunnel through? Perhaps this could be done by sending a particular netDb lookup so that the netDb router would respond not with peers in the netDb, but with random peers outside the netDb.

I2P's netDb is very different from traditional load bearing DHTs - it only carries network metadata, not any actual payload, which is why even a netDb using a floodfill algorithm will be able to sustain an arbitrary amount of I2P Site/IRC/bt/mail/syndie/etc data. We can even do some optimizations as I2P grows to distribute that load a bit further (perhaps passing bloom filters between the netDb participants to see what they need to share), but it seems we can get by with a much simpler solution for now.

One fact may be worth digging into - not all leaseSets need to be published in the netDb! In fact, most don't need to be - only those for destinations which will be receiving unsolicited messages (aka servers). This is because the garlic wrapped messages sent from one destination to another already bundles the sender's leaseSet so that any subsequent send/recv between those two destinations (within a short period of time) work without any netDb activity.

So, back at those equations, we can change L from 5 to something like 0.1 (assuming only 1 out of every 50 destinations is a server). The previous equations also brushed over the network load required to answer queries from clients, but while that is highly variable (based on the user activity), it's also very likely to be quite insignificant as compared to the publishing frequency.

Anyway, still no magic, but a nice reduction of nearly 1/5th the bandwidth/disk space required (perhaps more later, depending upon whether the routerInfo distribution goes directly as part of the peer establishment or only through the netDb).

The Disabling of the Kademlia Algorithm

Kademlia was completely disabled in release 0.6.1.20.

(this is adapted from an IRC conversation with jrandom 11/07)
Kademlia requires a minimum level of service that the baseline could not offer (bandwidth, cpu), even after adding in tiers (pure kad is absurd on that point). Kademlia just wouldn't work. It was a nice idea, but not for a hostile and fluid environment.

Current Status

The netDb plays a very specific role in the I2P network, and the algorithms have been tuned towards our needs. This also means that it hasn't been tuned to address the needs we have yet to run into. I2P is currently fairly small (a few hundred routers). There were some calculations that 3-5 floodfill routers should be able to handle 10,000 nodes in the network. The netDb implementation more than adequately meets our needs at the moment, but there will likely be further tuning and bugfixing as the network grows.

Update of Calculations 03-2008

Current numbers:

  recvKBps = N * (L + 1) * (1 + F) * (1 + R) * S / T
where
  N = number of routers in the entire network
  L = average number of client destinations on each router 
      (+1 for the routerInfo)
  F = tunnel failure percentage 
  R = tunnel rebuild period, as a fraction of the tunnel lifetime
  S = average netDb entry size
  T = tunnel lifetime
Changes in assumptions:
  • L is now about .5, compared to .1 above, due to the popularity of i2psnark and other apps.
  • F is about .33, but bugs in tunnel testing are fixed in 0.6.1.33, so it will get much better.
  • Since netDb is about 2/3 5K routerInfos and 1/3 2K leaseSets, S = 4K. RouterInfo size is shrinking in 0.6.1.32 and 0.6.1.33 as we remove unnecessary stats.
  • R = tunnel build period: 0.2 was a very low - it was maybe 0.7 - but build algorithm improvements in 0.6.1.32 should bring it down to about 0.2 as the network upgrades. Call it 0.5 now with half the network at .30 or earlier.
  recvKBps = 700 * (0.5 + 1) * (1 + 0.33) * (1 + 0.5) * 4KB / 10m
           ~= 28KBps
This just accounts for the stores - what about the queries?

The Return of the Kademlia Algorithm?

(this is adapted from the I2P meeting Jan. 2, 2007)
The Kademlia netDb just wasn't working properly. Is it dead forever or will it be coming back? If it comes back, the peers in the Kademlia netDb would be a very limited subset of the routers in the network (basically an expanded number of floodfill peers, if/when the floodfill peers cannot handle the load). But until the floodfill peers cannot handle the load (and other peers cannot be added that can), it's unnecessary.

The Future of Floodfill

(this is adapted from an IRC conversation with jrandom 11/07)
Here's a proposal: Capacity class O is automatically floodfill. Hmm. Unless we're sure, we might end up with a fancy way of DDoS'ing all O class routers. This is quite the case: we want to make sure the number of floodfill is as small as possible while providing sufficient reachability. If/when netDb requests fail, then we need to increase the number of floodfill peers, but atm, I'm not aware of a netDb fetch problem. There are 33 "O" class peers according to my records. 33 is a /lot/ to floodfill to.

So floodfill works best when the number of peers in that pool is firmly limited? And the size of the floodfill pool shouldn't grow much, even if the network itself gradually would? 3-5 floodfill peers can handle 10K routers iirc (I posted a bunch of numbers on that explaining the details in the old syndie). Sounds like a difficult requirement to fill with automatic opt-in, especially if nodes opting in cannot trust data from others. e.g. "let's see if I'm among the top 5", and can only trust data about themselves (e.g. "I am definitely O class, and moving 150 KB/s, and up for 123 days"). And top 5 is hostile as well. Basically, it's the same as the tor directory servers - chosen by trusted people (aka devs). Yeah, right now it could be exploited by opt-in, but that'd be trivial to detect and deal with. Seems like in the end, we might need something more useful than Kademlia, and have only reasonably capable peers join that scheme. N class and above should be a big enough quantity to suppress risk of an adversary causing denial of service, I'd hope. But it would have to be different from floodfill then, in the sense that it wouldn't cause humongous traffic. Large quantity? For a DHT based netDb? Not necessarily DHT-based.

Floodfill TODO List

NOTE: The following is not current information. See the main netdb page for the current status and a list of future work.

The network was down to only one floodfill for a couple of hours on March 13, 2008 (approx. 18:00 - 20:00 UTC), and it caused a lot of trouble.

Two changes implemented in 0.6.1.33 should reduce the disruption caused by floodfill peer removal or churn:

  1. Randomize the floodfill peers used for search each time. This will get you past the failing ones eventually. This change also fixed a nasty bug that would sometimes drive the ff search code insane.
  2. Prefer the floodfill peers that are up. The code now avoids peers that are shitlisted, failing, or not heard from in half an hour, if possible.

One benefit is faster first contact to an I2P Site (i.e. when you had to fetch the leaseset first). The lookup timeout is 10s, so if you don't start out by asking a peer that is down, you can save 10s.

There may be anonymity implications in these changes. For example, in the floodfill store code, there are comments that shitlisted peers are not avoided, since a peer could be "shitty" and then see what happens. Searches are much less vulnerable than stores - they're much less frequent, and give less away. So maybe we don't think we need to worry about it? But if we want to tweak the changes, it would be easy to send to a peer listed as "down" or shitlisted anyway, just not count it as part of the 2 we are sending to (since we don't really expect a reply).

There are several places where a floodfill peer is selected - this fix addresses only one - who a regular peer searches from [2 at a time]. Other places where better floodfill selection should be implemented:

  1. Who a regular peer stores to [1 at a time] (random - need to add qualification, because timeouts are long)
  2. Who a regular peer searches to verify a store [1 at a time] (random - need to add qualification, because timeouts are long)
  3. Who a ff peer sends in reply to a failed search (3 closest to the search)
  4. Who a ff peer floods to (all other ff peers)
  5. The list of ff peers sent in the NTCP every-6-hour "whisper" (although this may not longer be necessary due to other ff improvements)

Lots more that could and should be done -

  • Use the "dbHistory" stats to better rate a floodfill peer's integration
  • Use the "dbHistory" stats to immediately react to floodfill peers that don't respond
  • Be smarter on retries - retries are handled by an upper layer, not in FloodOnlySearchJob, so it does another random sort and tries again, rather than purposefully skipping the ff peers we just tried.
  • Improve integration stats more
  • Actually use integration stats rather than just floodfill indication in netDb
  • Use latency stats too?
  • More improvement on recognizing failing floodfill peers

Recently completed -

  • [In Release 0.6.3] Implement automatic opt-in to floodfill for some percentage of class O peers, based on analysis of the network.
  • [In Release 0.6.3] Continue to reduce netDb entry size to reduce floodfill traffic - we are now at the minimum number of stats required to monitor the network.
  • [In Release 0.6.3] Manual list of floodfill peers to exclude? (blocklists by router ident)
  • [In Release 0.6.3] Better floodfill peer selection for stores: Avoid peers whose netDb is old, or have a recent failed store, or are forever-shitlisted.
  • [In Release 0.6.4] Prefer already-connected floodfill peers for RouterInfo stores, to reduce number of direct connections to floodfill peers.
  • [In Release 0.6.5] Peers who are no longer floodfill send their routerInfo in response to a query, so that the router doing the query will know he is no longer floodfill.
  • [In Release 0.6.5] Further tuning of the requirements to automatically become floodfill
  • [In Release 0.6.5] Fix response time profiling in preparation for favoring fast floodfills
  • [In Release 0.6.5] Improve blocklisting
  • [In Release 0.7] Fix netDb exploration
  • [In Release 0.7] Turn blocklisting on by default, block the known troublemakers
  • [Several improvements in recent releases, a continuing effort] Reduce the resource demands on high-bandwidth and floodfill routers

That's a long list but it will take that much work to have a network that's resistant to DOS from lots of peers turning the floodfill switch on and off. Or pretending to be a floodfill router. None of this was a problem when we had only two ff routers, and they were both up 24/7. Again, jrandom's absence has pointed us to places that need improvement.

To assist in this effort, additional profile data for floodfill peers are now (as of release 0.6.1.33) displayed on the "Profiles" page in the router console. We will use this to analyze which data are appropriate for rating floodfill peers.

The network is currently quite resilient, however we will continue to enhance our algorithms for measuring and reacting to the performance and reliability of floodfill peers. While we are not, at the moment, fully hardened to the potential threats of malicious floodfills or a floodfill DDOS, most of the infrastructure is in place, and we are well-positioned to react quickly should the need arise.