search

kyren / turbulence

Networking library for games, multiplex reliable and unreliable streams over unreliable datagrams.
Apache License 2.0
262 stars 14 forks source link

readme

turbulence

We'll get there, but it's gonna be a bumpy ride.

Build Status Latest Version API Documentation

Multiplexed, optionally reliable, async, transport agnostic, reactor agnostic networking library for games.

This library does not actually perform any networking itself or interact with platform networking APIs in any way, it is instead a way to take some kind of unreliable and unordered transport layer that you provide and turn it into a set of independent networking channels, each of which can optionally be made reliable and ordered.

The best way right now to understand what this library is useful for probably to look at the MessageChannels test. This is the highest level, simplest API provided: it allows you to define N message types serializable with serde, define each individual channel's networking settings, and then gives you a set of handles for pushing packets into and taking packets out of this MessageChannels interface. The user is expected to take outgoing packets and send them out over UDP (or similar), and also read incoming packets from UDP (or similar) and pass them in. The only reliability requirement for using this is that if a packet is received from a remote, it must be intact and uncorrupted, but other than this the underlying transport does not need to provide any reliability or order guarantees. The reason that no corruption check is performed is that many transport layers already provide this for free, so it would often not be useful for turbulence to do that itself. Since there is no requirement for reliability, simply dropping incoming packets that do not pass a consistency check is appropriate.

This library is structured in a way that provides a lot of flexibility but does not do very much to help you actually get a network connection set up between a game server and client. Setting up a UDP game server is a complex task, and this library is designed to help with one piece of this puzzle.

What this library actually does

turbulence currently contains two main protocols and builds some conveniences on top of them:

1) It has an unreliable, unordered messaging protocol that takes in messages that must be less than the size of a packet and coalesces them so that multiple messages are sent per packet. This is by far the simpler of the two protocols, and is appropriate for per-tick updates for things like position data, where resends of old data are not useful.

2) It has a reliable, ordered transport with flow control that is similar to TCP, but much simpler and without automatic congestion control. Instead of congestion control, the user specifies the target packet send rate as part of the protocol settings.

turbulence then provides on top of these:

3) Reliable and unreliable channels of bincode serialized types.

4) A reliable channel of bincode serialized types that are automatically coalesced and compressed.

And then finally this library also provides an API for multiplexing multiple instances of these channels across a single stream of packets and some convenient ways of constructing the channels and accessing them by message type. This is what the MessageChannels interface provides.

Questions you might ask

Why would you ever need something like this?

You would need this library only if most or all of the following is true:

1) You have a real time, networked game where TCP or TCP-like protocols are inappropriate, and something unreliable like UDP must be used for latency reasons.

2) You have a game that needs to send both fast unreliable data like position and also stream reliable game related data such as terrain data or chat or complex entity data that is bandwidth intensive.

3) You have several independent streams of reliable data and they need to not block each other or choke off fast unreliable data.

4) It is impractical or undesirable (or impossible) to use many different OS level networking sockets, or to use existing networking libraries that hook deeply into the OS or even just assume the existence of UDP sockets.

Why do you need this library, doesn't XYZ protocol already do this (Where XYZ is plain TCP, plain UDP, SCTP, QUIC, etc)

In a way, this library is equivalent to having multiple UDP connections and bandwidth limited TCP connections at one time. If you can already do exactly that and that's acceptable for you, then you might consider just doing that instead of using this library!

This library is also a bit similar to something like QUIC in that it gives you multiple independent channels of data which do not block each other. If QUIC eventually supports truly unrleliable, unordered messages (AFAIK currently this is only a proposed extension?), AND it has an implementation that you can use, then certainly using QUIC would be a viable option.

So this library contains a re-implementation of something like TCP, isn't trying to implement something like that fiendishly complex and generally a bad idea?

Probably, but since it is designed for low-ish static bandwidth limits and doesn't concern itself with congestion control, this cuts out a lot of the complexity. Still, this is the most complex part of this library, but it is well tested and definitely at least works in the environments I have run so far. It's not very complicated, it could probably be described as "the simplest TCP-like thing that you could reasonably write and use".

You should not be using the reliable streams in this library in the same way that you use TCP. A good example of what shouldn't probably go over this library is something like streaming asset data, you should have a separate channel for data that should be streamed as fast as possible and will always be bandwidth rather than gameplay limited.

The reliable streams here are for things that are normally gameplay limited but might be spikey, and where you want to limit the bandwidth so those spikes don't slow down more important data or slow down other players.

Why is this library so generic? It's TOO generic, everything is based on traits like PacketPool and Runtime and it's hard to use. Why can't you just use tokio / async-std?

The PacketPool trait exists not only to allow for custom packet types but also for things like the multiplexer, so it serves double duty. Runtime exists because I use this library in a web browser connecting to a remote server using webrtc-unreliable, and I have to implement it manually on top of web APIs and that is currently not trivial to do.

Current status / Future plans

I've used this library in a real project over the real internet, and it definitely works. I've also tested it in-game using link conditioners to simulate various levels of packet loss and duplication and as far as I can tell it works as advertised.

The library is usable currently, but the API should in no way be considered stable, it still may see a lot of churn.

In the near future it might be useful to have other channel types that provide in-between guarantees like only reliability guarantees but not in-order guarantees or vice versa.

Eventually, I'd like the reliable channels to have some sort of congestion avoidance, but this would probably need to be cooperative between reliable channels in some way.

The library desperately needs better examples, especially a fully worked example using e.g. tokio and UDP, but setting up such an example is a large task by itself.

License

turbulence is licensed under any of:

at your option.