Game Networking Weekly #5

Unity rebrands “Netcode for Entities” as “Unity Netcode” and brings prediction, rollback and compressed state replication to GameObject workflows through its GhostObject APIs. This release post has a lot of breadcrumbs to follow for more details.

I particularly like one reply that makes their frame loop & hooks for prediction clear:

• MonoBehaviour FixedUpdate()
• MonoBehaviour Update()
• GhostBehaviour PredictionGatherInput() - once per frame
• Stepped together multiple times:
  • GhostBehaviour PredictionPhysicsUpdate() - x times per frame, full ticks only
  • Physic.Simulate() [sic]
  • GhostBehaviour PredictionUpdate() - n times per frame, can have partial ticks/batched if enabled

Explains Minecraft Bedrock’s move from RakNet to WebRTC and Open Collaboration’s NetherNet External Signaling (NXS).

A separate provider handles joins and issues short-lived encrypted tickets, while gameplay traffic goes directly to the server. Providers can route new players using server health and capacity. Their implementation multiplexes WebRTC connections onto one UDP port.

I haven’t found an explanation for why Mojang chose WebRTC data channels.

This one is more on the infra side. Riot compares server allocation strategies using recorded game workloads, leading to a reported 20% saving on hardware costs. Has some great data visualization.

Our intuition can be misleading when it comes to optimization. An algorithm that served your product well for a long time can work against you when circumstances change: in this case Riot’s operating environment moved from a data center-based pull model to a cloud-based push model for game server allocation. When this happens you have to measure again from (nearly) first principles. Put another way:

One accurate measurement is worth a thousand expert opinions.

For me the key takeaway from this article was not so much the specifics of their solution but the lesson of this paragraph.

We also wanted to use an algorithm that would not have to be changed every time we launch a new game mode. This decision proved necessary when we launched Swarm. This game mode completely obliterated the other alternatives and showed that Round Robin was the only algorithm that could withstand the havoc it unleashed on our game servers.

If your inputs are not fully under your control, requirements can change on you, and you may not have time to rearchitect when they do. Then you might want to stop chasing peak efficiency in favor of an algorithm that works well under a variety of conditions.

Cloudflare reports a shift toward numerous short DDoS bursts and increased UDP floods in Q3. Useful context for operating public game infrastructure.