I have been designing network protocol accelerator hardware and software for the last few years and it occured to me that some relatively cheap custom hardware could be used to accelerate online games. I got the idea from thinking about how you would add realistic voice chat to a game without overburdening the network. I worked up a simulation modeling an arena with several hundred players and the numbers are quite striking. I can also see how other features could be accelerated in this way. I am wondering if custom hardware is being used anywhere now. It needs to be custom because to get the acceleration you need to interpret the game protocol directly. With prgrammable gate arrays you can quickly reprogram if the game protocol changes. Is anyone using custom hardware now?

It would seem to me that if the component you are looking to accelerate can be separated enough from the rest of the game to allow it to be built using custom hardware, then most designers would probably just throw another standard off-the shelf processor at the job. After all, the greatest cost by far in software projects such as MMRPGs is salaries to developers so a custom hardware chip would have to be extremely efficient (basicly replace whole game servers) in order to make up for the increases in development and integration costs that it would bring over standard hardware.

Henry

I haven''t heard of anyone using custom hardware in this area. You might try selling the idea directly to them... I can definitely see this being useful. There are things you can do with custom hardware that just "throwing another processor at" couldn''t do.

I am having a hard time trying to imagine a MMRPG-related task that could not be accomplished cheaper by using standard hardware components than by using custom hardware. After all, the development and integration costs of even a small custom hardware component could in many cases cost more than whole multi-processor game servers. Multiple MMRPG companies could possibly benefit from doing joint development on solutions for common features in that industry, but such cooperation might quickly make those developments ''non-custom'' too

Henry

Henry: Unfortunately some causes of lag can''t be fixed by just adding more "standard hardware components". Getting information to the processor or processors in a format they can use immediately cuts down lag. Custom routers and protocol accelerators do exactly this.

Magmai: Peer-to-peer can never be the solution to this type of problem because it doesn''t scale. Peer-to peer voice chat for example creates an exponential explosion of network traffic. Client-server traffic grows linearly. In the architecture I have modeled you get several advantages (like packet filtering and duplication) in addition to the prime advantage to having a single logical expansion node. If you want to increase the quality, for example, you add bandwidth to one logical node and its associated processing elements. You can also easily adjust quality depending on load. To increase quality in a peer-to-peer network you have as many expansion nodes as clients. In hardware design we always look to minimize the number of expansion nodes. It is almost a definition of the quality of a design.

Another problem with the peer-to-peer model is compression. Each node in a peer-to-peer network must decompress the audio from every other node before it can be mixed and any other functions performed (like filtering a sound behind a wall) before the user can hear it. In a client-server model each client only has to encode and decode a single stream. Hardware compression and decompression can be added to the server. You can’t force each client to add this hardware.

You also have the problem of having to send information to the client you may not want to send just so the audio can be modeled correctly. Cheaters always seem to find some way to use this information.

As for “A special server netcard that could be programmed to decrypt, descramble, parse packets, and DMA them into a system RAM buffer might be useful...” Been there, done that.

There is no need to send all the data to every client. Each client has one stream to the server and one from the server. Let the server figure out what the client needs to hear, mix it and send it. Client decompresses, adds prerecorded sounds and plays it.

MaxClientBandwidth = 2*n
MaxServerBandwidth = 2*m*n

where
n = bits per second
m = number of clients

using your 5k number and 100 clients:
MaxClientBandwidth = 2*5k = 10k
MaxServerBandwidth = 2*5k*100 = ~1M

Server workload is dependant on the quality of the simulation. This simulation could also be accelerated by custom hardware. There are lots of neat optimizations for the simulation also.

No idea where you got you numbers from. My understanding is that MMORPG worlds are broken up over many servers. Each server handles one or more ''zones''. You would have to do the same division of labor for audio. In fact I would use the existing load balancing scheme for this division. I would add servers just for audio processing. That way using the custom protocols accelerators packet duplication feature you just poach the information you already have.