Skip to content
/ tonk Public

Tonk : Reliable UDP (rUDP) Network Library and Infinite Window Erasure Code

License

Notifications You must be signed in to change notification settings

catid/tonk

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

26 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Tonk

Reliable UDP (rUDP) Network Library in C++

Tonk.h provides a simple C API for a powerful, multithreaded network engine. The Tonk network engine transports message data over the Internet using a custom protocol over UDP/IP sockets. Like TCP, Tonk delivers data reliably and in-order. The main improvement over TCP is that data sent with Tonk arrives with much less latency due to congestion control and head-of-line blocking issues.

Tonk is experimental: It's only been unit tested and hasn't been used for any major projects yet. There are probably bugs to fix. It currently only builds for Linux and Windows targets right now.

Tonk puts each connection into its own thread, so all processing related to each connection happens serially. The application is responsible for handling thread safety on the server side between multiple connections.

Tonk implements a novel HARQ transport based on SiameseFEC, meaning that a new type of forward error correction is used to protect data and avoid head-of-line blocking that causes latency spikes with TCP sockets. SiameseFEC is the world's first infinite-window erasure code, suitable for protecting reliable in-order data like a TCP stream, but less suitable for audio or video data protection, which does not need an infinite window.

So what's an infinite window erasure code?

An "erasure code" means when data is lost due to packetloss, there are extra redundant packets sent that can fill in for the loss without knowing ahead of time what to retransmit. This means that if the loss is smaller than the redundancy then there is no head of line blocking like in TCP. Typically there are multiple chances to recover from a loss with FEC before retransmission (ARQ) schemes get a chance to react to packetloss, and ARQ often waits two round trips or more to retransmit a lost message to avoid sending data unnecessarily. Meanwhile, FEC takes roughly 1-2% BW overhead (about 2x the loss rate) and eliminates the need for retransmissions in almost all cases. For cost comparison, Google's BBR protocol intentionally takes a 2% hit to throughput as part of its design.

A "streaming erasure code" is one that's suitable for data being generated live - close to when it is sent. For example, key presses in a SSH terminal being sent over a network cannot be predicted so they must be "streamed." An erasure code that's suitable for a file would be called a "block erasure code" and it would work well for a large file (block of data) broken up into pieces. There are a few types of streaming erasure codes:

(1) Generational block codes - This uses a block code as a streaming code. These will group sets of source data into block generations. The downside is that recovery can only happen at the block boundaries so this is often not much better than ARQ. My Shorthair project implements this.

(2) Sliding window codes - These will send redundant data for the last X seconds of original data. This is fine for video streaming but TCP style transfer does not care how old the data is. It's inefficient to calculate a whole convolution for every new recovery packet, so it cannot scale to higher data rates. My Cauchy Caterpillar project is a sliding window code.

(3) Infinite window codes - These never leave any old data behind and new recovery packets can recover data from back to the start of the connection. This is suitable for data that must be delivered reliably and in order. The main advantage over sliding window codes is they can reuse work from one calculation to the next, saving lots of CPU time. My Siamese project implements the tricky math for this.

Tonk takes Siamese and adds in the reliable UDP code so it's useful to people more directly. It has a bunch of other nice features too.

So what are those other nice features?

Tonk implements a new type of jitter-robust time synchronization protocol over the Internet that is more suitable for mobile networks than NTP/PTP. Tonk enables applications to compress timestamps down to 2 or 3 bytes.

Tonk implements a new type of congestion control strategy that attempts to guarantee low latency for real-time data while full-speed file transfers are being performed in the background. Tonk supports three levels of message prioritization, 12 different parallel data streams, unordered delivery, and unmetered unreliable delivery.

Tonk implements several types of fully automated Peer2Peer NAT traversal to allow mobile applications to peer with the aid of a rendezvous server. It also incorporates WLANOptimizer to reduce latency on some Windows PCs.

Tonk's philosophy is to trade extra CPU work and extra memory for faster data delivery, so it may not be suitable for your application. For example it uses streaming data compression with Zstd to compress packets so there is less to send, but this does take more CPU resources and requires memory. It also implements the fastest known erasure codes for this application and it has to send recovery packets in addition to normal data, which takes more CPU/network resources but will avoid head-of-line blocking and deliver the data faster. It uses a multithreaded design, which requires extra complexity and overhead but means that multiple connections can be processed in parallel to reduce the overall latency. The peer2peer design allows the application to cut out a network hop to the server and this requires a lot of additional complexity, but the result is significantly lower latency. At the same time to offset all of these additional costs, a lot of engineering effort went into choosing/designing/tuning the best algorithms to minimize overhead. TL;DR Tonk is all about reducing latency by taking advantage of faster CPUs.

Usage

CMake can be used to generate build files for Linux or Windows.

The main documentation is in the tonk.h header file.

The tonk project is a dynamic link library (DLL) providing the Tonk C API. The tonk_static project is a static linkage version of Tonk.

There is also a tonkcppsdk project that builds a C++ static library containing a loader that will load the Tonk DLL and also defines the C++ SDK for Tonk.

There is a C# wrapper in the sdk/csharp directory. Some simple C# example projects are in the sdk/csharp/tests directory.

There is a LAN advertising and direct file transfer demo in the FileSender and FileReceiver projects.

There's a Peer2Peer file transfer demo as well. The P2PFileRendezvousServer project contains the server that must be on a public server. You can provide the DNS hostname and file names as arguments to the P2PFileReceiver and P2PFileSender applications. These demo applications will establish a peer2peer connection via the rendezvous server and transfer a file while reporting RTT and OWD conditions periodically.

The test_server and test_p2p_client projects are unit test projects that may have some useful example code.

There is a unit tester that uses MauProxy to simulate a file transfer during all sorts of nasty network conditions (packetloss, re-ordering, corruption, etc), which also tests a bunch of the classes.

Appropriate applications:

  • Data rate limit of 20,000,000 bytes per second, per connection
  • Low-latency messaging
  • Mobile file (e.g. video/image) upload
  • Lossless video/audio streaming
  • Real-time multiplayer gaming
  • Mobile VPN data accelerator
  • Peer2Peer mobile file transfer, chat, or data streaming

Inappropriate applications:

  • High rate file transfer > 20 MB/s: Forward error correction does not work at these higher rates, so using a custom UDP-based congestion control works better
  • Lossy real-time audio/video conference or chat: SiameseFEC has an infinite window, so using Cauchy Caterpillar is more appropriate: https://github.com/catid/CauchyCaterpillar
  • High security applications: Tonk does not implement any PKI or encryption

Message delivery:

  • Unordered unreliable message delivery
  • Unordered reliable message delivery
  • Multiple reliable in-order channels (6)
  • Multiple reliable in-order low-priority channels (6)
  • Breaks up large messages into pieces that fit into datagrams
  • Detects and rejects duplicate packets (incl. unreliable)
  • File transfer API for sending large buffers or files from disk (See tonk_file_transfer.h)

Latency features:

  • 0-RTT connections: Server accepts data in first packet
  • Lower latency than TCP congestion control (10 milliseconds)
  • Enables file transfer coexisting with low latency applications
  • Message prioritization delivers the most important data first
  • Forward Error Correction (FEC) to reduce median latency
  • Ordered reliable data is compressed with Zstd for speedier delivery

Extras:

  • Jitter-robust time synchronization for mobile applications
  • NAT hole punching using UPnP for Peer2Peer connections
  • NAT hole punching using STUN + NATBLASTER (with rendezvous server)
  • Automatically opens up any needed ports in the Windows firewall
  • Detects and rejects data tampering on the wire
  • SYN-cookies enabled during connection floods to mitigate DoS attacks
  • Fast obfuscation (encryption without security guarantees)
  • Reduced idle traffic using ack-acks

Not provided (Non-goals):

  • Data security (Requirements are too application-specific)
  • TCP fallback if UDP-based handshakes fail (Application can do this)
  • Unreliable-but-ordered delivery (Can be done using Unreliable channel)
  • MTU detection (Assumes 1488 byte frames or larger)

Runtime flag options:

These options can be set per-socket but not per-connection. They can be set differently on each side of a connection.

  • FEC usage can be reduced to save some CPU overhead
  • UPnP port forwarding can be turned on
  • Datagram compression can be turned off
  • Random padding is off by default and can be turned on
  • Automatic bandwidth control can be disabled and forced to a fixed maximum send rate

Credits

Tonk uses the Siamese library for its forward error correction (FEC), selective acknowledgements (SACK), and jitter buffer. Tonk uses the Asio library for portable networking and strands. Tonk uses the Cymric library for secure random number generation. Cymric uses BLAKE2b and Chacha. Tonk uses the Zstd library (modified) for fast in-order packet compression. Tonk uses the t1ha library for fast hashing.

Software by Christopher A. Taylor mrcatid@gmail.com

Please reach out if you need support or would like to collaborate on a project.

About

Tonk : Reliable UDP (rUDP) Network Library and Infinite Window Erasure Code

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published