ECSOS - C++ Entity Component System based on Ordered Sets

A data-oriented header-only C++14 library

Bas Geertsema

During the development of my game I have created an entity-component-system (ECS). I decided to extract the functionality into a generic library as it might be useful for other (game) developers as well. The library is named ECSOS and is now available on GitHub.

ECSOS is a small header-only C++14 library that provides a data-oriented entity-component-system (ECS) based on ordered sets. By default it uses the boost flat_set container, which stores set elements (components) in a cache-friendly contiguous array, although other (custom) containers might be used.

The library is now available on GitHub.

Why should I use it?

Entity-Component-Systems are commonly used in games where entities consist of one ore more components. The use of an ECS provides some advantages: components can be removed and added during runtime and components can efficiently be processed by systems operating on this data because of the compact data-layout. This is in contrast with Object-Oriented design in which entity types are determined during compile time and where entities are large heterogeneous objects with different datasizes, which leads to pointer-chasing in the heap which is often slower than straight-forward traversal of arrays.

Some characteristics of using ordered sets in an ECS are:

  • fast traversal of a single component set and unions of component sets
  • reasonably fast lookup of individual components
  • convenient value semantics to store and copy state (no pointers involved)

How does it work?

Entities and identities

Each component contains an identifier (typically an integer, but can be any type) that defines the entity that it belongs to. Entities are implicitly defined by their components. Therefore only components are ever stored: there are no distinct entity objects stored in the system. For example, if you would have two component of types Transform and RigidBody, and both components would have an identifier of ‘2’ then these components would implicitly be part of the same entity. Note that components are typically very simple data structs.

Finding and traversing entities

A common problem in ECS is how to work on entities that must contain at least two or more predefined components. For example, let’s say you have a component type Transform that defines the world transformation of an entity and is used by the renderer. Furthermore, you also have another component type RigidBody that contains the transform of its collision object and is used by the physics engine. In each frame the physics engine updates the positions in the RigidBody component. But, just before rendering, you would like to propagate the new positions of the RigidBody component to the Transform component such that the entity is rendered properly. Because the components are disjoint in memory you somehow have to figure out how to lookup the related Transform component for each RigidBody component that is associated with the same entity.

In ECSOS, this is solved by making use of the fact that the sets are known to be ordered. For a single entity, you can easily lookup another component of the same entity based on the entity identifier. This is a binary search and thus has a lookup cost of, at worst, O(log N).

However, in many scenarios you would want to iterate over all entities that contain components in a specific list of component types. For example if you would like to iterate over all entities that have at least two components of both type Transform and RigidBody. The iteration over these entities is done by having multiple iterators, one for each component set, that advances through all components and skips over any components of which the entity id is not found in any of the other set(s). This might sound complicated but fortunately this is abstracted away by the ECSOS library, as you will see. In your application you will typically work with a intuitive single iterator that works on the union of component sets.

The following example will make it clear. In the following example we iterate over all entities that contain both a component of type Transform and a component of type RigidBody. These components are respectively stored in the sets transforms and bodies:

for (auto it = entities_begin(transforms, bodies); it != entities_end(transforms, bodies); ++it) {
    Transform& transform = get<Transform>(*it);
    RigidBody& body = get<RigidBody>(*it); 
}

Equivalently, we can use the auto range loop:

for (auto entity : entities(transforms, bodies)) {
    Transform& transform = get<Transform>(entity);
    RigidBody& body = get<RigidBody>(entity);            
}

The examples above show that the usage of ECSOS is quite straight-foward and intuitive.

Accessing components

Individual components can be accessed by using the get<> function, as can be seen in the examples above. The entity variable in the example above holds a light-weight temporary object of the template type entity<>. A entity<> object contains pointers to its components and can therefore efficiently be copied to other functions as an argument. For example:

void HandleTransformRigidBody(entity<Transform, RigidBody> e)
{
}

entity<Transform, RigidBody> e;
HandleRigidbodyByRef(e);	// this is ok, because the entity object is cheap to copy

An entity object can implicitly be converted to subsets if possible. For example, the entity in the example above contains a Transform, RigidBody and Character component but can also be given as an argument to the following function, that expects an entity containing only a Transform and Character component.

void HandleRigidbodyTransform(entity<Transform, Character> e)
{
}

entity<Transform, RigidBody, Character> e;
HandleRigidbodyByRef(e);	// implicitly converted to a subset entity

The entity object also implicitly converts to references of any of its components, for example:

void HandleRigidbodyByRef(const RigidBody& body)
{
}

entity<Transform, RigidBody> e;
HandleRigidbodyByRef(e);	// implicitly converted to a reference to its RigidBody component

ECSOS has a very straight-forward, convenient and type-safe approach to using entities and components in your system.

How fast is it?

Fair question, but it is hard to give a simple answer. The performance depends heavily on your usage patterns, data layout, component sizes, etc. So I cannot give you absolute values. However, you can make educated guesses on its performance for your use case based on the characteristics of ECSOS.

What makes ECSOS fast:

  • Components are stored in contiguous arrays. Traversing is cache-friendly and works well with pre-fetching.
  • Component sets are disjoint in memory. Systems only load the components in cache that they actually need to do their work.
  • Components (state) can easily be copied and processed in parallel. Because there are no pointers to keep track of. For example, you can trivially copy and retain the whole state of the ECS for a couple of frame in a computergame to allow for pipelined execution.
  • Searching the component for a particular entity is a relatively fast binary search. Which is faster than a linear search. However, it is obviously not as fast as a direct pointer and might also be slower than a lookup in a hashtable.

What might make ECSOS not so fast:

  • When iterating a union of sets, all components in the sets are read to extract the entity identifier. This means that for union sets in which there is a large ratio between the component sets (i.e. there are a large number of elements in set A, and only a few in set B) a lot of data is read and subsequently skipped over. Technically we could do a smart binary search if the ratio is very high and this might be faster, but this is currently not implemented. The impact of this is largely determined by the ratio of your component sets in the union sets that you use.
  • Removing or inserting an element in the middle of a ordered set means that other elements must be moved to fill the gap. Although this reduces to a comfortable memmove operation it is still slower than using unordered sets or hashmaps.

That being said, this system has served me well on my (in progress) development of a RTS game which has ‘only’ a couple of hundred entities that do not change a lot over time. Especially the option to create relatively fast copies of the whole state has been proven very convenient for, amongst others, pipelining, parallel processing, deserialization and testing.

So the bottomline is that you really have to experiment and measure for yourself whether this would give you enough performance for your needs.

Are other (custom) containers supported?

Yes, you can specialize the ecs::is_union_base_set trait for any other (custom) container. It is assumed that the container keeps its elements ordered at all times. See here the specialization for boost::container::flat_set. As another example it can also work with std::set, however the iterators of std::set do not provide mutable references so I found it much less usable in practice.

Where can I get it?

ECSOS is available on GitHub.

comments powered by Disqus