6 replies to this topic

[laylay]

I'm really not understanding network prediction, specifically for weapon firing.

For example, If I tell the server I'm starting to fire my weapon, it's going to simulate weapon firing until it gets told to stop weapon firing. I don't see how client timestamps of these two actions can solve that because it doesn't know the exact time duration the client held down fire until the stop firing messages comes in, this could come in late due to packet latency. So it's possible that server simulated weapon firing for 5 seconds but client did it for 4.9 seconds, that has a chance of popping off a weapon fire when its off cooldown on the server but the client didn't play the animation and effects. Ammo count is now out of sync (unless I then correct that prediction error but I still didn't see the weapon fire effects for the last fire and I think it would be too late to correct that and play the animation and effects)

[hplus0603]

it doesn't know the exact time duration the client held down fire

Simulation should use a fixed time step. For example, 30 steps per second, or 60 steps per second.
Weapon fire only starts and stops on time steps.
Input data should be time stamped with the time step it applies to.

[laylay]

I'm a bit confused. Should the weapon fire simulation run at a fixed time step on both client and server?

I can't wrap my head around how that fixes the problem. If input data is sent to the server to start firing and it happend at 10 seconds client time and ended at 11 seconds, what if the second input packet got to the server really late? The server would carry on simulating even though client isn't holding down fire any more.

[Inferiarum]

The packet being late means it is dropped. And of course this can lead to different results on the server and client which have to be corrected.

You have to run the client at the right offset relative to the server time such that only a small percentage of the packets is late. But of course with a long prediction window you get other problems

[laylay]

But aren't all packets late because they have to travel to the server? The problem is the latency, an input packet could come in quickly but the second input packet that has mouse button up could come in late enough to have the server fire the weapon more times than the client wanted it to.

If I timestamp packets, isn't that open for speed hacking because anyone could spoof that data.

Also I'm a bit wary of running the simulation at a fixed time step, wouldn't that make it frame rate dependant?

Edited by laylay, 15 September 2012 - 08:56 PM.

[wildbunny]

This problem is quite hard - you need to have the client tell the server how many fixed sized time-steps there were between key-actions, then the server can correct forward or backwards (by rewinding time).

[hplus0603]

I'm a bit confused. Should the weapon fire simulation run at a fixed time step on both client and server?

Yes.

I can't wrap my head around how that fixes the problem. If input data is sent to the server to start firing and it happend at 10 seconds client time and ended at 11 seconds, what if the second input packet got to the server really late? The server would carry on simulating even though client isn't holding down fire any more.

If the packet doesn't make it, then the server simulates firing, which means that the client will be told that it's wrong (corrected.) This is what it means to be server authoritative.
The alternative would be to delay the game for everybody until the input makes it, which is what some lock-step RTS games do.