Spawning child entities is extremely slow #17301
Comments
aevyrie
added
A-ECS
BenjaminBrienen
added
S-Needs-Design
|
I had a lot of trouble replicating this performance issue on The fix is to recognize that when the children we are adding to an entity have just been spawned, we can skip all of these checks. There is no way for the hierarchy to become malformed, because these entities are brand new. In the linked benchmark, the 1 million entity case was projected to take 7.5 hours, but completes in 111ms with fixes. |
|
Another note: the linked PR doesn't fully fix this issue, it only resolves one set of cases that are easy to optimize for (and common). In the case where an entity is not new, adding this as a child still scales exponentially after the change. This is because to check if an entity already exists, bevy currently does a linear scan of a vector, which scales linearly with the length of the vector. Instead, we should probably be using a hashmap once the vector is large enough to spill onto the heap. This would make these checks constant time. |
|
As a heads up, if we land something in the "non-fragmenting relations" category (which I have an incoming PR for), that would fully resolve this. I'm not opposed to landing your optimizations first though, in the interest of ensuring some form of improvement lands in the next release. |
What does this mean? The relations live in a central store or something? It's still possible for these optimizations to be missed, unless I'm missing something. Edit: #16920 After skimming this, it still appears possible to hit the same performance issue, it completely depends on the implementation. I don't see how the design would prevent this from happening. |
You're right it isn't inherent to the space. My particular implementation does not suffer from the problem. The Normal per-child removal does still require a scan to find the entity. But with the new |
|
Thanks for the details, that all makes sense. |
This adds support for one-to-many non-fragmenting relationships (with planned paths for fragmenting and non-fragmenting many-to-many relationships). "Non-fragmenting" means that entities with the same relationship type, but different relationship targets, are not forced into separate tables (which would cause "table fragmentation"). Functionally, this fills a similar niche as the current Parent/Children system. The biggest differences are: 1. Relationships have simpler internals and significantly improved performance and UX. Commands and specialized APIs are no longer necessary to keep everything in sync. Just spawn entities with the relationship components you want and everything "just works". 2. Relationships are generalized. Bevy can provide additional built in relationships, and users can define their own. **REQUEST TO REVIEWERS**: _please don't leave top level comments and instead comment on specific lines of code. That way we can take advantage of threaded discussions. Also dont leave comments simply pointing out CI failures as I can read those just fine._ ## Built on top of what we have Relationships are implemented on top of the Bevy ECS features we already have: components, immutability, and hooks. This makes them immediately compatible with all of our existing (and future) APIs for querying, spawning, removing, scenes, reflection, etc. The fewer specialized APIs we need to build, maintain, and teach, the better. ## Why focus on one-to-many non-fragmenting first? 1. This allows us to improve Parent/Children relationships immediately, in a way that is reasonably uncontroversial. Switching our hierarchy to fragmenting relationships would have significant performance implications. ~~Flecs is heavily considering a switch to non-fragmenting relations after careful considerations of the performance tradeoffs.~~ _(Correction from @SanderMertens: Flecs is implementing non-fragmenting storage specialized for asset hierarchies, where asset hierarchies are many instances of small trees that have a well defined structure)_ 2. Adding generalized one-to-many relationships is currently a priority for the [Next Generation Scene / UI effort](#14437). Specifically, we're interested in building reactions and observers on top. ## The changes This PR does the following: 1. Adds a generic one-to-many Relationship system 3. Ports the existing Parent/Children system to Relationships, which now lives in `bevy_ecs::hierarchy`. The old `bevy_hierarchy` crate has been removed. 4. Adds on_despawn component hooks 5. Relationships can opt-in to "despawn descendants" behavior, meaning that the entire relationship hierarchy is despawned when `entity.despawn()` is called. The built in Parent/Children hierarchies enable this behavior, and `entity.despawn_recursive()` has been removed. 6. `world.spawn` now applies commands after spawning. This ensures that relationship bookkeeping happens immediately and removes the need to manually flush. This is in line with the equivalent behaviors recently added to the other APIs (ex: insert). 7. Removes the ValidParentCheckPlugin (system-driven / poll based) in favor of a `validate_parent_has_component` hook. ## Using Relationships The `Relationship` trait looks like this: ```rust pub trait Relationship: Component + Sized { type RelationshipSources: RelationshipSources<Relationship = Self>; fn get(&self) -> Entity; fn from(entity: Entity) -> Self; } ``` A relationship is a component that: 1. Is a simple wrapper over a "target" Entity. 2. Has a corresponding `RelationshipSources` component, which is a simple wrapper over a collection of entities. Every "target entity" targeted by a "source entity" with a `Relationship` has a `RelationshipSources` component, which contains every "source entity" that targets it. For example, the `Parent` component (as it currently exists in Bevy) is the `Relationship` component and the entity containing the Parent is the "source entity". The entity _inside_ the `Parent(Entity)` component is the "target entity". And that target entity has a `Children` component (which implements `RelationshipSources`). In practice, the Parent/Children relationship looks like this: ```rust #[derive(Relationship)] #[relationship(relationship_sources = Children)] pub struct Parent(pub Entity); #[derive(RelationshipSources)] #[relationship_sources(relationship = Parent)] pub struct Children(Vec<Entity>); ``` The Relationship and RelationshipSources derives automatically implement Component with the relevant configuration (namely, the hooks necessary to keep everything in sync). The most direct way to add relationships is to spawn entities with relationship components: ```rust let a = world.spawn_empty().id(); let b = world.spawn(Parent(a)).id(); assert_eq!(world.entity(a).get::<Children>().unwrap(), &[b]); ``` There are also convenience APIs for spawning more than one entity with the same relationship: ```rust world.spawn_empty().with_related::<Children>(|s| { s.spawn_empty(); s.spawn_empty(); }) ``` The existing `with_children` API is now a simpler wrapper over `with_related`. This makes this change largely non-breaking for existing spawn patterns. ```rust world.spawn_empty().with_children(|s| { s.spawn_empty(); s.spawn_empty(); }) ``` There are also other relationship APIs, such as `add_related` and `despawn_related`. ## Automatic recursive despawn via the new on_despawn hook `RelationshipSources` can opt-in to "despawn descendants" behavior, which will despawn all related entities in the relationship hierarchy: ```rust #[derive(RelationshipSources)] #[relationship_sources(relationship = Parent, despawn_descendants)] pub struct Children(Vec<Entity>); ``` This means that `entity.despawn_recursive()` is no longer required. Instead, just use `entity.despawn()` and the relevant related entities will also be despawned. To despawn an entity _without_ despawning its parent/child descendants, you should remove the `Children` component first, which will also remove the related `Parent` components: ```rust entity .remove::<Children>() .despawn() ``` This builds on the on_despawn hook introduced in this PR, which is fired when an entity is despawned (before other hooks). ## Relationships are the source of truth `Relationship` is the _single_ source of truth component. `RelationshipSources` is merely a reflection of what all the `Relationship` components say. By embracing this, we are able to significantly improve the performance of the system as a whole. We can rely on component lifecycles to protect us against duplicates, rather than needing to scan at runtime to ensure entities don't already exist (which results in quadratic runtime). A single source of truth gives us constant-time inserts. This does mean that we cannot directly spawn populated `Children` components (or directly add or remove entities from those components). I personally think this is a worthwhile tradeoff, both because it makes the performance much better _and_ because it means theres exactly one way to do things (which is a philosophy we try to employ for Bevy APIs). As an aside: treating both sides of the relationship as "equivalent source of truth relations" does enable building simple and flexible many-to-many relationships. But this introduces an _inherent_ need to scan (or hash) to protect against duplicates. [`evergreen_relations`](https://github.com/EvergreenNest/evergreen_relations) has a very nice implementation of the "symmetrical many-to-many" approach. Unfortunately I think the performance issues inherent to that approach make it a poor choice for Bevy's default relationship system. ## Followup Work * Discuss renaming `Parent` to `ChildOf`. I refrained from doing that in this PR to keep the diff reasonable, but I'm personally biased toward this change (and using that naming pattern generally for relationships). * [Improved spawning ergonomics](#16920) * Consider adding relationship observers/triggers for "relationship targets" whenever a source is added or removed. This would replace the current "hierarchy events" system, which is unused upstream but may have existing users downstream. I think triggers are the better fit for this than a buffered event queue, and would prefer not to add that back. * Fragmenting relations: My current idea hinges on the introduction of "value components" (aka: components whose type _and_ value determines their ComponentId, via something like Hashing / PartialEq). By labeling a Relationship component such as `ChildOf(Entity)` as a "value component", `ChildOf(e1)` and `ChildOf(e2)` would be considered "different components". This makes the transition between fragmenting and non-fragmenting a single flag, and everything else continues to work as expected. * Many-to-many support * Non-fragmenting: We can expand Relationship to be a list of entities instead of a single entity. I have largely already written the code for this. * Fragmenting: With the "value component" impl mentioned above, we get many-to-many support "for free", as it would allow inserting multiple copies of a Relationship component with different target entities. Fixes #3742 (If this PR is merged, I think we should open more targeted followup issues for the work above, with a fresh tracking issue free of the large amount of less-directed historical context) Fixes #17301 Fixes #12235 Fixes #15299 Fixes #15308 ## Migration Guide * Replace `ChildBuilder` with `ChildSpawnerCommands`. * Replace calls to `.set_parent(parent_id)` with `.insert(Parent(parent_id))`. * Replace calls to `.replace_children()` with `.remove::<Children>()` followed by `.add_children()`. Note that you'll need to manually despawn any children that are not carried over. * Replace calls to `.despawn_recursive()` with `.despawn()`. * Replace calls to `.despawn_descendants()` with `.despawn_related::<Children>()`. * If you have any calls to `.despawn()` which depend on the children being preserved, you'll need to remove the `Children` component first. --------- Co-authored-by: Alice Cecile <[email protected]>
Bevy version
0.15, has existed for as long as constructing
ParentandChildrenhas been locked down.What you did
Attempt to optimize spawning large batches of children in a custom command for
big_space.What went wrong
Profiling shows that applying system commands dominates system time. Spawning 320k entities took 31s.
The problem appears to be that even when spawning a brand new entity, bevy is doing reparenting checks for every child when its
Parentcomponent is added, and again for every child when the children are added to the parent entity'sChildrencomponent. These checks are completely unnecessary - the entities are new and cannot yet exist in the hierarchy.By doing some
transmuteshenanigans, I was able to see what a niave solution would look like if I could simply construct theParentandChildrenentities myself. The performance improvement was stark: what previously took 31s now took 83ms.I was also able to work around this without unsafe by using reflection to construct a
ParentandChildren, bypassing the private constructors.The text was updated successfully, but these errors were encountered: