> True P2P networks don't scale, because every node has to store an (accurate if partial) representation of the whole network
Former LimeWire dev here... which P2P networks use a fully meshed topology? LimeWire and other Gnutella clients just have a random mesh with a fixed number of (ultra)peers. If the network gets too large, then your constrained broadcast queries hit their hop count before reaching the edge of the network, but that seems fine.
Last I checked, Freenet used a variation on a random mesh.
Kademlia's routing tables take O(log(N)) space and traffic per-peer to maintain (so O(N log(N)) for global total network space and traffic). Same for Chord (though, twice as much traffic due to not using a symmetric distance metric like XOR).
There are plenty of "True" (non-centralized) P2P networks that aren't fully meshed.
Creator of Freenet here. Freenet[1] relies on peers self-organizing into a small world network[2].
Small world networks have the advantage of being able to find data in log N time where N is the network size, they're also completely decentralized, self-healing, and distribute load evenly across peers. The principle is similar to DHTs like Kademlia but more flexible and intuitive IMO, while having similar scaling characteristics.
It's surprisingly common for people to confuse small world networks with "scale free networks", but scale free networks rely on a subset of highly connected peers which do a disproportionate amount of the work - which isn't truly decentralized.
The new Freenet design incorporates adaptive learning into the routing algorithm. When a peer is deciding where to route a message, it predicts the response probability and time for each neighboring peer based on past performance and chooses the best. With conventional "greedy routing", peers choose the neighbor with a location closest to the data being retrieved. The new approach is adaptive to actual network performance.
[1] Both the original Freenet from 1999 and the new sequel we're currently building - see https://freenet.org/ for more. We hope to launch the network in the next few weeks.
As far a second-generation Freenet, I heard i2p started out as a proposed re-factoring and generalization of Freenet's encrypted transport layer. Are there any plans on using i2p to carry Freenet traffic?
I2P was created by someone who was previously involved with Freenet, but its design is a lot closer to Tor than to Freenet. Both I2P and Tor are anonymizing proxies, they allow services to be hidden, but they're still centralized.
While they are quite different, there is enough overlap that running Freenet over I2P (or Tor) would be wildly inefficient and slow, so I wouldn't recommend it. Freenet is designed to run over UDP directly.
The new Freenet is designed to allow the creation of completely decentralized services. Briefly, it's a global key-value store in which keys are webassembly code that specify what values are permitted under that key, and the conditions under which those values can be modified. This key-value store is observable, so anyone can subscribe to a key and be notified immediately if the value changes.
This is just scratching the surface, for anyone interested in a much more comprehensive explanation of the new Freenet please see this talk I gave a few months ago: [1] You can also find a FAQ here: [2]
I think ive been following the dev chat long enough to answer that the new freenet is a new, seperate network to the original (now called hyphanet I think) that handles transport by itself, and end to end encryption is not in scope of the project but can be built on top
> the new freenet is a new, seperate network to the original
This is correct - while old and new Freenet both rely on a small-world network, they are very different and not compatible. Borrowing from our FAQ[1], the main differences are:
Functionality: The previous version of Freenet (now called Hyphanet) was analogous to a decentralized hard drive, while the current version is analogous to a full decentralized computer.
Real-time Interaction: The current version allows users to subscribe to data and be notified immediately if it changes. This is essential for systems like instant messaging or group chat.
Programming Language: Unlike the previous version, which was developed in Java, the current Freenet is implemented in Rust. This allows for better efficiency and integration into a wide variety of platforms (Windows, Mac, Android, MacOS, etc).
Transparency: The current version is a drop-in replacement for the world wide web and is just as easy to use.
Anonymity: While the previous version was designed with a focus on anonymity, the current version does not offer built-in anonymity but allows for a choice of anonymizing systems to be layered on top.
Doesn't Java have the widest variety of hardware running it, thanks to its Virtual Machine ?
I can even remember my Motorola Razr being arguably (almost) a smartphone because, while a far cry from Symbian, it could already run Java applications ! (Notably, IIRC, Opera mini ?)
P.S.: Also, I tried Freenet about around that time too ! I'm a bit confused about this being a "new" project... why not naming it "Freenet 2" then ? Why did Freenet "1" had to change its name ??
> Doesn't Java have the widest variety of hardware running it, thanks to its Virtual Machine ?
Java has the advantage that you can run it on a wide variety of hardware platforms without recompilation, but it has largely failed to attain broad usage/support for desktop apps and so it's a bad choice for something like Freenet in 2024.
A systems programming language like Rust gives us a lot more control over things like memory allocation, allowing apps to be a lot more efficient. This is important with Freenet because we need it to run in the background without slowing down the user's computer.
Rust can also be compiled to run on all major platforms, Windows, Mac, Linux, Android, iOS, etc.
> P.S.: Also, I tried Freenet about around that time too ! I'm a bit confused about this being a "new" project... why not naming it "Freenet 2" then ? Why did Freenet "1" had to change its name ??
Using the name for the new software was a difficult decision and not without risk.
The "Freenet" name was never intended to belong to a specific codebase. From the start we viewed the original Java implementation as a prototype, which is one reason we never actually released version 1.0 (even 7 years after the project started we were still on version 0.7). At the time I had no idea that it would be over 20 years before I had a design I thought would be suitable, but here we are.
This new Freenet is the original concept but designed, not as a prototype, but as software that can gain broad adoption. In that sense it is the fulfilment of my original vision.
I did consider calling it Freenet 2, but these days not that many people have heard of the original, so on balance I believe it would have been confusing for the (hopefully) much bigger userbase we hope to reach.
I've always taken interest into p2p, but never heard of Freenet so thanks for being here!
Question: how good is the latency once connections are already established, say for a real time video call over Freenet, or if this is not possible? Is there any server in the middle that all packets need to route to, especially for peers behind firewall
> how good is the latency once connections are already established, say for a real time video call over Freenet, or if this is not possible?
We're aiming for the delay between a contract's state being modified and all subscribers being notified to be no more than 1 second, which should be acceptable for applications like IM.
If you were doing something like a video call you'd negotiate it over Freenet and then establish a direct connection for the video/audio for minimal latency.
> Is there any server in the middle that all packets need to route to, especially for peers behind firewall
Freenet uses UDP hole-punching to establish direct connections between peers even if both are behind firewalls. A new peer uses a public (non-firewalled) peer to join the network initially but once joined all further communications can be between firewalled peers.
The low-level transport is one of the final components we're working on prior to launch - we previously hoped to use something off-the-shelf but on close examination nothing fit our requirements.
We handle congestion by specifying a global maximum upload rate for the peer, which will be a fraction of the peer's total available bandwidth - the goal being to avoid congestion. In the future we'll likely use an isotonic regression to determine the relationship between upstream bandwidth usage and packet loss, so that we can adaptively choose an appropriate maximum rate.
This is a more detailed explanation of the transport protocol, it's quite a bit simpler than ECN but we'll refine it over time: [1]
At a higher level a peer's resource usage will be proportional to the number of connected peers, and so the peer will adaptively add or remove connections to stay under resource usage limits (which includes bandwidth).
Sure, but ultrapeers/supernodes/routers/etc are forms of centralization. Locally, the network is centralized around these supernodes, and they can be annoying/impossible to bypass. The "inner network" or backbone of supernodes, if exists, can also represent a central authority. Nothing necessarily wrong with any of this, but it can stretch the meaning of P2P if it really means P2central-authority2P.
Functionally there is almost no difference between me sending you an (anonymous, encrypted) message over Facebook versus over some sophisticated, large, hierarchical "P2P" network. We still have to trust the local authorities, so to speak.
How are Gnutella ultrapeers centralized? Nodes self-promote themselves to ultrapeers if they meet uptime and resource availability requirements.
It's a mechanism to allow low-spec'd peers to participate without getting crushed with traffic, and also an optimization to reduce churn in the routing tables. The ultrapeers are just long-uptime regular peers.
Former LimeWire dev here... which P2P networks use a fully meshed topology? LimeWire and other Gnutella clients just have a random mesh with a fixed number of (ultra)peers. If the network gets too large, then your constrained broadcast queries hit their hop count before reaching the edge of the network, but that seems fine.
Last I checked, Freenet used a variation on a random mesh.
Kademlia's routing tables take O(log(N)) space and traffic per-peer to maintain (so O(N log(N)) for global total network space and traffic). Same for Chord (though, twice as much traffic due to not using a symmetric distance metric like XOR).
There are plenty of "True" (non-centralized) P2P networks that aren't fully meshed.