ECS

Veloren uses an Entity Component System (ECS). This is a relatively new paradigm in game engine development that competes against traditional object-oriented design models that make heavy use of hierarchy, inheritance, and polymorphism. It encourages game developers to design data structures in a way that allows for efficient batch processing of data on modern CPU architectures with deep caches by storing batchable data contiguously in memory.

Traditionally, the representation of entities within a game are done like so:


#![allow(unused)]
fn main() {
struct Entity {
    position: Vec3<f32>,
    velocity: Vec3<f32>,
}
}

When represneting multiple entities, it's common to use an array-like data structure.


#![allow(unused)]
fn main() {
struct World {
    entities: Vec<Entity>,
}
}

However, this comes with problems. Modern CPUs have deep caches, meaning that they are significantly faster at accessing memory that is closer in the address space to memory they have already accessed. The most efficient storage formats pack data to be processed very close together in memory. Our current representation looks like the following:

| Entity 0            | Entity 1            | Entity 2            |
|---------------------|---------------------|---------------------|
| position | velocity | position | velocity | position | velocity |

Let us imagine a 'typical' operation that we'd like to perform: applying gravity to the entities. This involves adding 9.81 in vertical velocity to each of the entities. Unfortunately, the access pattern of this operation involves touching each velocity field, but skipping each position field. These gaps in our data make the operation slower. In technical jargon, we call this a lack of 'cache coherency'.

All is not lost.

A design pattern you might have heard of is called Struct Of Arrays, or SOA. It suggests that instead of packing the data associated with each entity together in one place, we should instead group data according to its purpose.


#![allow(unused)]
fn main() {
struct World {
    positions: Vec<Vec3<f32>>,
    velocities: Vec<Vec3<f32>>,
}
}

Now our representation is much more densely packed and our application of gravity to entities is more efficient because we no longer need to skip over the position data while iterating through entity velocities.

| Entity 0 | Entity 1 | Entity 2 |     | Entity 0 | Entity 1 | Entity 2 |
|----------|----------|----------| ... |----------|----------|----------|
| position | position | position |     | velocity | velocity | velocity |

ECS is this idea taken to its logical conclusion: entities as associated but separate collections of components, with each type of component stored in its own distinct storage buffer. This approach comes with more advantages too: because components are stored separately, it is easy and fast to add or remove components from an entity as the game is running, thereby altering the behaviour of the entity and its capabilities. This is a remarkably powerful technique and allows for what amount to the ability to radically alter the behaviour of the game and the entities within it while the game is running and without sacrificing performance.

The specific ECS crate that Veloren uses is SPECS. You can read more about ECS and the specifics of how SPECS works here. If you're looking to work on Veloren, I strongly recommend reading this resource from cover to cover (it's quite short).