I’ve jumped right in to the deep end with my game project (which is going under the working title of “Encounters” – I needed to call the solution something!) and one of the core design goals of the project is the multiplayer focus of the game.
I’m a huge fan of Id Software. John Carmack is probably my favourite “celebrity computer programmer”. He’s incredible smart and consistent and has made some of both the most influential and my own favourite games.
Standing on the shoulders of giants - The “Quake” model
I’ve always thought the “Quake model” of single player and multiplayer game development to be a good one. For those that don’t know, one of the simple design tenants of the Quake engine is that everything is a multiplayer game. When you play the campaign in the original Quake what the game actually does is start a local game server which you then connect to. Due to the proliferation of the Quake engine and other engines that have been inspired by it’s design this became quite a common way to build first person shooters with multiplayer support from the offset. It reduced the implementation of multiplayer to simply having a second player connect to the active session.
Standing on the shoulders of giants, I’ve decided that this is the model I wish to follow for Encounters, especially seeing as the primary work-in-progress game design desires 4-player coop as the main campaign (with NPC assistants if you play with fewer humans). Because of the way supporting multiplayer from the offset effectively means that all the games “thinking” is done in the server component (collision detection, cheat prevention, state management..) I decide that it’d be the best place to start the implementation.
Prototyping
So far so good, I spent an hour or two last night designing a simple state managing game server and retrofitting it to my previous prototype (an arena with a player-controlled unit moving around it) and moving all the logic and validation to the server side.
At this point the “game server” was a singleton class that the game accessed though a fake “proxy” class (left empty for eventual network implementation) and it worked pretty well. The collision detection worked on the “server side” when the server was running in the same application domain and as a proof of concept everything was quite sound. I’m still left facing a few issues surrounding the frequency of syncing with the server and client side prediction, but they’re all relatively well solved problems in gaming (and there’s certainly some prior art to take inspiration from in XNA tutorials around the web).
Building a network stack
Once the proof of concept was working I started looking into the networking support provided by XNA and hit the mother of all roadblocks. Because I’m targeting both the 360 and Windows, networking becomes significantly more complicated. See, Microsoft offer no direct network access on the 360, via XNA, or even to their licensed partners (internet hearsay claims). They provide Xbox Live APIs as part of the XNA framework however, which seemed like a decent solution until I realised that these networking APIs function only on a 360 as part of Xbox live.
Not so multiplatform really
So I’ve started thinking around the problem. I do a lot of work with WCF in my day-job writing distributed systems so my obvious inclination was to provide a WCF client-server architecture for the Windows targeted version of the game. This adds an additional set of issues. For a start it looks as though the WCF assemblies (System.ServiceModel) are probably not available on the 360. You’d think that’s not much of a problem if you’re going to use the Live APIs, but then you realise that if you’re using a version of .Net 3.5 prior to the first service pack, you need to annotate any classes you want to transfer via WCF with the DataContract attribute. Which is in the System.ServiceModel namespace.
I’m not sure (as I currently don’t have a subscription to deploy my test code onto my 360) what version of the framework the 360 is running, but I’d hazard a guess that it’s pre SP1. I’m going to have to further research this problem, because if you can use SP1 on the 360 (which supports serialization without any attributes on the classes) then we’re home free.
Multiplatform networking code using Inversion of Control
In the interim I’ve devised a cunning solution (or perhaps workaround, I’m not sure). See, all of these issues lead me to need two entirely separate network stacks for my game. One that supported the Windows, and the other that supported Xbox live, without contaminating my data model with mark-up that’s useable in only one or the other of the stacks.
Castle Windsor to the rescue.
This is purely speculative as I haven’t attempted to implement the 360 network stack yet, but I intend to use the Castle Windsor IoC (inversion of control) container to load a separate network stack at runtime, depending on the platform.
My idea is that the game knows about two interfaces IGameServer and IGameServerConnection. I then create two entirely separate assemblies, one with the networking code for the PC, the other with the Live networking code. The PC version contains a class called WcfGameServer which implements IGameServer (I’ve already written this code), and a class called WcfGameServerConnection which implementes IGameServerConnection (and acts as a hand crafted WCF proxy). The 360 implementation will feature XBL counterparts to these classes.
As far as the game is concerned, it’ll use the Windsor container to load an instance of IGameServer, call the StartListening(); method, then use Windsor to load an instance of IGameServerConnection() and call JoinGame() on that connection. This way, the specific networking implementation is entirely removed from the game and hidden behind these two simple interfaces.
I’ve currently got a good way through implementing the WCF version of this model though I’ve hit a few snags on the way. Because I’m deliberately developing in a pre-SP1 environment for the sake of this exercise, WCF doesn’t like serializing the game model to send over the wire. As a result I’ve had to produce an inelegant hack to work around my desire to keep the System.ServiceModel assembly reference clear of my game model.
Trying to keep it light
I’ve marked up my model with Serializable attributes (.NET 1.1, that’ll be fine) and inside my WCF implementation of IGameServerConnection I’m marshalling all my data into byte[]’s before sending it over the wire. This isn’t ideal as it requires the WCF implementation to manually deserialize the byte[]’s into their correct data types in the service implementation and the client library, but it does work. Unfortunately at the moment these byte arrays are being stored as XML before being sent over the wire (as is the default WCF way) so I’m going to need to force WCF to binary serialize all it’s data rather than bloat my packets (and as a result, the game latency). I’m using Net.Tcp so it’s pretty lightweight as far as protocols go but I suspect I’ll need to do some additional fine tuning to make the WCF implementation viable.
Either way, I’ve got a good feeling about the model, subtle hacks aside, and I think this is quite a good way to target both platforms with minimal impact to your game code.
So this is day 3 of development (I’m sure I’m going to loose count pretty quickly).
[Footnote]
It occurs to me that I got a few fundamental things wrong when this post was originally written regarding the requirements for .Net 3.5 SP1. These mistakes almost complicated the design of the networking stack. Read more in my follow up here.