This document is about: FUSION 1
SWITCH TO

Performance & Optimizations

Level 4

Interest Management

Fusion BR uses the Eventual Consistency replication mode that allows clients to receive data only about network objects they are interested in. This improves performance and reduces network traffic.

Every player defines a cone-shaped area based on camera direction to receive information about objects in front and close to them.

player aoi

Since the Area of Interest (AoI) is represented by a single box, the cone shape is achieved by using multiple AoIs with different sizes. The image above illustrates the area setup. The player only receives state for objects within these areas, other objects are skipped.

The project also contains special components to further reduce performance cost and network traffic:

StaticNetworkTransform
One-time position/rotation synchronization of static objects upon spawn. StaticNetworkTransform is used for item boxes. There could be hundreds of boxes in the level and usage of StaticNetworkTransform is more optimal than NetworkTransform.

NetworkAreaOfInterestProxy NetworkAreaOfInterestProxy is used for objects that don't need to synchronize position or rotation but still need to be filtered by the interest management system. For example, weapons that are parented under an owning Agent don't need to synchronize transforms, but their networked data should be received based on whether the weapon is in the player's area of interest or not. The position for AoI calculations is taken from a specified object. In the case of the weapon, it is based on the owning agent position. This feature has zero network traffic overhead.

NetworkCulling Small utility for local culling. When a NetworkObject doesn't get a new state from the state authority for a given period of time (outside of player’s AoI), the Networked object is culled.

Although the features above have a major impact on network traffic, there is still some space for further improvements:

  • Better area of interest cone limits based on level design
  • Precalculated visibility zones
  • Player A shooting at player B could be manually added to player B's list of interested objects to ensure the network data will be received and visible (e.g., incoming projectiles are visible to the player even though they were shot from far behind).

Additional info about the interest management can be found in the Fusion Interest Management section.

Benchmarks

The recommended maximum player counts for stable 60Hz simulation on headless dedicated server instance:

Configuration (Hardware + Build type) Player Count
Mac Mini M1 2020, Mac server (Silicon) 200
Core i9-10900K, Windows server 200
Multiplay Bare Metal (Intel® Xeon® E-2288G @ 3.70GHz), Linux server 200
Multiplay Virtual Machine (GCP N2 - Intel® Xeon® @ 2.80GHz), Linux server 60 - 80
multiplay bare metal (intel® xeon® e-2288g @ 3.70ghz) performance
Performance of Multiplay Bare Metal (Intel® Xeon® E-2288G @ 3.70GHz) over time, DeltaTime in milliseconds
multiplay virtual machine (gcp n2 - intel® xeon® @ 2.80ghz)
Performance of Multiplay Virtual Machine (GCP N2 - Intel® Xeon® @ 2.80GHz) over time, DeltaTime in milliseconds

The actual performance and the player count highly depend on the gameplay mode. Deathmatch, where all players are active all the time, is always more demanding than Battle Royale where eliminated players have low overhead. It is also affected by level design (map size, player density) and gameplay design (forcing players to constantly move/look/shoot has a different impact than 200 campers). Numbers above represent the worst case scenario - first minute after the spawn - all players actively moving, looking and shooting.

Note on Multiplay Virtual Machines: Another factor which affects the performance is the specification of currently allocated hardware. While one match (game server instance) can barely handle 60 players, another runs 100 players without any problems. This is caused by different hardware specifications. Running on a bare metal server for higher player counts and improves stability.

Profiling

In addition to common profiling tools, this sample provides tools for offline analysis of the game in release mode. When the game is launched with the -recordSession parameter, engine delta time and player count are recorded every frame and stored in a file in CSV format. Then this file can be processed by provided Python scripts (available in Assets/Extras.zip) which will generate HTML files with graphs.

The file is created at the Application.persistentDataPath path or at the path specified in the -dataPath XXX command.

profiling stats

The image above shows an increasing player count (up to 200) displayed by the blue line and engine delta time (being stable most of the time with few peaks above 16ms). Aside from DeltaTime and PlayerCount, you can create your own set of tracked properties with a custom StatsRecorder instance.

Pooling

In shooter projects, a large amount of object spawning is expected (projectiles, muzzle effect, impacts, hit effects,...). Since instantiation of a new object is expensive, everything in Fusion BR is pooled.

There are two types of pools:

  • NetworkObjectPool is used to pool the NetworkObjects. The NetworkObjectPool implements the INetworkObjectPool interface and is passed into the NetworkRunner.StartGame function in the Networking script. For more information, check out the Network Object Pool section in the Fusion Manual.

  • ObjectCache is a generic GameObject pool that's used for non-networked objects such as impact effects. It comes with a feature to return objects after a specified delay. ObjectCache is accessible in other parts of the code through the SceneContext.

Back to top