1 unstable release

0.1.0 Jan 16, 2024

#376 in Data structures

Download history 4/week @ 2024-01-11 2/week @ 2024-02-15 9/week @ 2024-02-22 4/week @ 2024-02-29 10/week @ 2024-03-28 6/week @ 2024-04-04 45/week @ 2024-04-11

61 downloads per month

Apache-2.0

105KB
2K SLoC

Totally Not a PlayStation

tnaps, pronounced tiː-næps, is a pure entity-component-system (ECS) framework written in rust. It was written to test the hypothesis that the same techniques that allow games and their development process to scale with ECS will also allow interactive applications to be built quickly and with good performance. tnaps is a MVP to validate this hypothesis.

Entity-Component-System Overview

Per Wikipedia, the entity-component-system paradigm has been around at least since "Thief: The Dark Project" in 1998. The system embraces the maxim, "Composition over inheritance." Entities exist, they have components, and systems operate on said components. In tnaps, an entity is a bag of components. Arbitrary components can be added to or unbound from entities. It can be thought of as a pointer. A component is pure data, or data with a small number of localized methods that touch no other component. A system operates on a set of entities and their components and serves as the means by which entities with components interact with one another. Put another way that may be familiar to database folks, the entities are row pointers, the components are columns, and the system is a query engine.

The entity-component-system has a major advantage that can best be seen in this table:

Component Entity 1 Entity 2 Entity 3 ... Entity N
ABC
...
XYZ

The component ABC is bound to entities 1, 2, and N, while the component XYZ is bound to entities 2, 3, and N. A system that operates on ABC must necessarily visit entities 1, 2, and N, while a system that operates on XYZ must necessarily visit 2, 3, and N. A system that operates on both ABC and XYZ components need only touch entities 2 and N. And because entities 2 and N are different, the system can operate on them in parallel as if it operated on them sequentially.

This observation is at the heart of tnaps.

This table illustrates the two degrees of freedom tnaps exhibits. First, data is partitioned by component. A component is defined to be a self-contained unit. It could be a health meter, a position, or any other property of an entity. tnaps rewards small, isolated components. Second, data is partitioned by entity. Systems that support concurrency and parallelism may operate on different entities in parallel.

tnaps leverages rust's concurrency guarantees---namely send and sync marker traits---to provide a framework by which code can be made concurrent. Systems whose state is entirely maintained within components can be trivially parallelized with minimal code changes.

Getting Started

Let's jump right into an end-to-end example. Here is a complete implementation of a system that uses tnaps to process two components.

# use tnaps::{
#     system, system_parallel, ComponentChange, ComponentCollection,
#     ComponentRef, CopyOnWriteComponentCollection, CopyOnWriteComponentRef,
#     Entity as EntityTrait, MutableComponentCollection, MutableComponentRef,
#     NopPartitioningScheme, Partitioned, PartitioningScheme, ThreadPool,
# };
// Declare Entity to be a u128.
// Out of the box, tnaps supports u128, u64, and u32 entities.
type Entity = u128;

// Two sample componenents.
// They are just plain-old-data.
// Components can be practically any Rust type.
#[derive(Clone, Debug)]
struct ComponentAbc {
    x: u64,
    y: u64,
}

#[derive(Debug)]
struct ComponentXyz {
    z: f64,
}

// Let's declare a system that operates only on ABC.  It's totally
// permitted---and even desirable---for systems to have no state of their own.
// More complicated systems such as those we'll see later can have state, but it
// is subject to Rust's standard rules for concurrency.
struct SystemAbc;

// The system macro is the means by which we create a `run` method for our
// system.  We declare SystemAbc to have entity type Entity, and that it
// operates on a CopyOnWriteComponentCollection of type ComponentAbc.
system! {
    SystemAbc<Entity> {
        abc: CopyOnWriteComponentCollection<ComponentAbc>,
    }
}

// Our implementation of the system operates on an entity and a reference to a
// ComponentAbc.  Notice there's no mention of other entities.  We are operating
// on a single column from the above table.
impl SystemAbc {
    fn process(&self, entity: Entity, abc: &mut CopyOnWriteComponentRef<ComponentAbc>) {
        println!("processing: {}", entity);
    }
}

// Our system that operates entirely on a mutable component collection of
// ComponentXyz.
struct SystemXyz;

system! {
    SystemXyz<Entity> {
        xyz: MutableComponentCollection<ComponentXyz>,
    }
}

// Our implementation of SystemXyz.
impl SystemXyz {
    fn process(
        &self,
        entity: Entity,
        xyz: &mut MutableComponentRef<ComponentXyz>,
    ) {
        // We unbind entity two.
        if entity == 2 {
            xyz.unbind();
        }
        println!("processing: {}", entity);
    }
}

// A system that operates on both ABC and YXZ.
struct SystemAbcXyz;

system! {
    SystemAbcXyz<Entity> {
        abc: CopyOnWriteComponentCollection<ComponentAbc>,
        xyz: MutableComponentCollection<ComponentXyz>,
    }
}

// Our implementation of SystemAbcXyz.
impl SystemAbcXyz {
    fn process(
        &self,
        entity: Entity,
        abc: &mut CopyOnWriteComponentRef<ComponentAbc>,
        xyz: &mut MutableComponentRef<ComponentXyz>,
    ) {
        // We unbind entity two.
        if entity == 2 {
            xyz.unbind();
        }
        println!("processing: {}", entity);
    }
}

fn main() {
    let mut collection_abc = CopyOnWriteComponentCollection::from_iter(vec![
        (1u128, ComponentAbc { x: 10, y: 20 }),
        (3u128, ComponentAbc { x: 42, y: 43 }),
    ]);
    let mut collection_xyz = MutableComponentCollection::from_iter(vec![
        (2u128, ComponentXyz { z: std::f64::consts::PI }),
        (3u128, ComponentXyz { z: std::f64::consts::E }),
    ]);
    let sys1 = SystemAbc;
    let sys2 = SystemXyz;
    let sys3 = SystemAbcXyz;
    // Execute sys1 against abc.
    let (changes_abc,) = sys1.run(&mut collection_abc);
    assert!(changes_abc.is_empty());
    // Execute sys2 against xyz
    let (changes_xyz,) = sys2.run(&mut collection_xyz);
    assert!(!changes_xyz.is_empty());
    collection_xyz.apply(changes_xyz);
    // Execute sys3 against both abc and xyz.
    let (changes_abc, changes_xyz) = sys3.run(&mut collection_abc, &mut collection_xyz);
    collection_abc.apply(changes_abc);
    collection_xyz.apply(changes_xyz);
}

As may be intuitive, this will run sys1 against entities 1 and 3, sys2 against entities 2 and 3, and sys3 against their intersection: entity 3.

This is the power of the tnaps framework: It takes the intersection of entities and components and efficiently selects only those entities that are relevant to a system.

This example showcases the majority of features of tnaps; it's intended to be lightweight and consumable within a few hours.

We can make our example parallel as follows (only for sys3):

# use std::sync::Arc;
#
# use tnaps::{
#     system, system_parallel, ComponentChange, ComponentCollection,
#     ComponentRef, CopyOnWriteComponentCollection, CopyOnWriteComponentRef,
#     Entity as EntityTrait, MutableComponentCollection, MutableComponentRef,
#     NopPartitioningScheme, Partitioned, PartitioningScheme, ThreadPool,
# };
#
# type Entity = u128;
#
# #[derive(Clone, Debug)]
# struct ComponentAbc {
#     x: u64,
#     y: u64,
# }
#
# #[derive(Debug)]
# struct ComponentXyz {
#     z: f64,
# }
// A system that operates on both ABC and YXZ, in parallel.
struct SystemAbcXyz;

system_parallel! {
    SystemAbcXyz<Entity> {
        abc: CopyOnWriteComponentCollection<ComponentAbc>,
        xyz: MutableComponentCollection<ComponentXyz>,
    }
}

// Our implementation of SystemAbcXyz.  Unchanged from our first example.
impl SystemAbcXyz {
    fn process(
        &self,
        entity: Entity,
        abc: &mut CopyOnWriteComponentRef<ComponentAbc>,
        xyz: &mut MutableComponentRef<ComponentXyz>,
    ) {
        // We unbind entity two.
        if entity == 2 {
            xyz.unbind();
        }
        println!("processing: {}", entity);
    }
}

fn main() {
    let partitioning: Arc<dyn PartitioningScheme<Entity>> = Arc::new(NopPartitioningScheme) as _;
    let collection_abc = CopyOnWriteComponentCollection::from_iter(vec![
        (1u128, ComponentAbc { x: 10, y: 20 }),
        (3u128, ComponentAbc { x: 42, y: 43 }),
    ]);
    let mut collection_abc = Partitioned::from(&partitioning, collection_abc.partition(&*partitioning));
    let mut collection_xyz = MutableComponentCollection::from_iter(vec![
        (2u128, ComponentXyz { z: std::f64::consts::PI }),
        (3u128, ComponentXyz { z: std::f64::consts::E }),
    ]);
    let mut collection_xyz = Partitioned::from(&partitioning, collection_xyz.partition(&*partitioning));
    let sys3 = Arc::new(SystemAbcXyz);
    // Execute sys3 against both abc and xyz.
    let thread_pool = ThreadPool::new("example", 4);
    let waiter = sys3.run(&thread_pool, &mut collection_abc, &mut collection_xyz);
    let (changes_abc, changes_xyz) = waiter();
    collection_abc.apply(changes_abc);
    collection_xyz.apply(changes_xyz);
}

As we can see, our implementation of our system didn't change. Our only changes were to use the system_parallel macro instead of system and to wrap each entity as a Partitioned collection. We used a NopPartitioningScheme, but it's just as valid to use another partitioning scheme.

That's all it takes to convert a system from non-concurrent to concurrent behavior.

Component Collections

tnaps' design centers around components and their collections. A singular component collection is intended to map from entities to components in an efficient manner that allows us to join multiple components together in a single system. tnaps ships four component collections out of the box, that reasonably cover more use cases. It is always possible to manually implement the ComponentCollection trait to implement a custom collection.

The included collections are:

  • Copy On Write: This component collection returns all its changes for use with the apply collection call. It will then rewrite the entire collection to maintain order of the components according to their entities.

  • Insert Optimized: This component collection is optimized for workloads that bind components to entities or update entities infrequently. Unlike the copy-on-write collection, which amortizes update cost, the insert-optimized collection exposes a method for quick inserting of single elements, such that the average cost of inserting an element is greater than other collections, but there is no amortization of the cost.

  • Mutable: This component collection is useful for when components won't be inserted frequently, but will instead be updated in-place. Binding and unbinding of components from entities is only slightly (nanoseconds) more expensive than the copy-on-write collection and is significantly less costly than the insert-optimized collection.

  • Partitioned: This component collection wraps another type of component collection and fragments it into partitions according to a partitioning scheme. This is useful because different threads can operate on different partitions concurrently. Consequently, the partitioned collection is a prerequisite for parallel systems.

Common Patterns

This section outlines some common patterns to get the most out of tnaps.

  • When working with systems that operate on different components, it is desirable to order the components by increasing cardinality. The fewer items in a component, the more it can prune entities from consideration in other components. For example, it might make sense to order the player-oriented components before the bullet-oriented components in a system that tries to detect when players get hit by a bullet.

  • The copy-on-write component collection is Send and Sync, meaning that it can be used by multiple systems in parallel. This allows an application to aggregate changes from multiple parallel systems and apply them as one batch at the end of a main-loop iteration.

  • It's tempting to use the run_subset feature of systems to subselect entities on which to operate. This is marginally more efficient than using marker components to prune for a system. Create a marker component for each entity that would be passed to run_subset and trust the system to run in parallel. run_subset is unable to work on parallel systems.

  • run_subset is useful for enumerating entities out of order. This allows a sequential system (perhaps collision detection) to pre-sort the entities by a predicate and then be guaranteed that the entities will be enumerated in arbitrary order. This can be a win for systems that need to consider entities with locality that doesn't match the total order across entities.

  • Binding new components to entities using an insert-optimized map has a cost. Unless the entity needs to be immediately bound, or there are few updates per tick, it is more efficient to use a copy-on-write or mutable component collection and apply all updates as a single batch.

  • All entities exist all the time, they just may not be bound to any components, and thus omitted from consideration. This eliminates a central entity registry.

  • It is possible to make an entity type that's not an integer. This is an anti-pattern within tnaps. Use the built-in entity types for most applications.

  • Generate entities using modular arithmetic. Pick a number that is co-prime to the number of entities possible, start at an arbitrary random number, and then use wrapping_mul to enumerate the entities. By the properties of modular arithmetic, all entities will be enumerated before recycling. To efficiently pick a coprime, select small primes other than two at random and multiply them together until the next-selected prime would overflow your entity type. By properties of the prime factorization, the number will be co-prime to two to the one-hundred-twenty-eight.

  • Alternatively, use u128 as an entity and generate a UUID.

  • Use sync components to allow in-place updates, even for the CopyOnWriteComponentCollection.

  • FastEntityMap is slower to construct, but significantly faster to query. There's an open TODO to make all component collections generic over EntityMap.

  • Consider the main loop to encode a DAG of data flow. This will help model systems better and is akin to a data flow engine.

  • Systems can be viewed as database-like "joins" of components. Where different entities must interact, there exists a minimal system that joins their component data. The process method of the system should collect the data for the join, and then the system should perform the join either in the process call, or in a separate call after run or run_parallel completes.

  • "Signaling" can be accomplished by having listeners that get joined against notifiers.

  • "Zones" of entities can be made where each zone has an independent set of collections and systems. To move entities between zones, have a system that completely unbinds an entity from one zone's collections and binds it in all the requisite component collections in the other zone. There's an open TODO to unify this with partitioning.

  • Input is a component.

  • Create marker components to join entities against recent activity.

  • Persistent systems should detect changes to components and persist the changes. This can be used to get durability of the entire entity-component-system application. It is an anti-pattern to try and keep all state from the database in tnaps. Instead, retain primary keys and copy them out to systems that return references into the database.

Design Choices

There are many design choices intentionally baked into tnaps to make it both functional and small. Functionality is preferred over size, but being able to keep the entire framework in one's head has its advantages.

  • Each component instance is bound to exactly one entity. This is a fundamental assumption of tnaps and enables its efficient pruning of entities and components. It is always possible to shove shared state behind an Arc to have components that share state across entities.

  • The interface for systems was intentionally designed to process one entity at a time. This allows parallelization across entities, without leaking any details of said parallelization. In the future, SIMD-like systems will be possible.

  • tnaps assumes all component collections will be sorted according to an entity map. This enables efficient scanning of all components to find the intersection of entities that exist in two or more component collections.

  • No main is included with tnaps. This is highly application dependent. It will look like a main function with a loop that repeatedly runs systems on components. The similarities stop there, so tnaps ships nothing.

  • No systems are included with tnaps. Common components for large-scale interactive applications will ship separately.

  • Entities always exist. Consequently it is never an error to bind new components to entities. This assumption enables performance.

  • Rust was chosen for tnaps because of its safety features, largely Send + Sync.

  • Most tnaps tests use random, property-based testing.

  • Entities are chosen to be pure, and devoid of meaningful data. Consequently they are lightweight and easy to copy. This is why Entity must implement Copy.

  • Components should be pure data. Different data should be different components. Multiple sparse collections should be as efficient to process as one giant, inheritance-based collection. There are no requirements on component in general (except Debug and Clone where appropriate).

  • Systems should be code plus system state. tnaps places no assumptions on what they look like. This is freedom, not a restriction.

  • tnaps' system and component interface allows flexible scheduling. It is an open TODO to make the scheduling dynamically repartition collections according to execution time of the partitions.

Napkin Computations

Let's pretend we're in a bar, working on a napkin, to figure out what's possible for tnaps.

For starters, let's clear the air: One million is not a big number for a computer.

ENTITIES = 1_000_000

And modern computers are ridiculously large for low cost. For $8-$11/hour we can get a machine with more than 100 cores and hundreds of gigabytes of memory.

Concretely, we have two reference points that can be leased from Amazon AWS:

c7i.48xl i4i.32xl
CPU Cores 192 128
Memory 384 GB 1024 GB
Network 50 Gbit/s 75 Gbit/s
Attached SSD N/A 8 x SSD
Price/Hour $8.568 $10.982

These machines are ridiculously over-powered and cost relatively little compared to the costs of a regular software-oriented business. If we can fit our entire application in tnaps (our central hypothesis), these machines should be able to fit the bill.

Let's start with compute. tnaps is built around a loop that executes once per tick. What's a tick? Let's make it variable and see how much compute time falls out for a c7i.48xl:

Tick Interval Compute Per Entity
1/60 second 3.2s 3.2µs
1/30 second 6.4s 6.4µs
1 second 192s 192µs
5 seconds 950s 960µs
15 seconds 48min 2.88ms
60 seconds 192min 11.5ms

I've waited longer than a minute for my rideshare and food delivery apps to acknowledge my order. What are they doing with that time?

Of course, I'm glossing over durability. Let's continue using our c7i.48xl machine. It provides 40Gbit/s of bandwidth to Amazon EBS.

Recovery Volume Recovery Time
64 B/entity 12.8ms
128 B/entity 25.6ms
256 B/entity 51.2ms
512 B/entity 102ms
1 KB/entity 204ms
... ...
64 KB/entity 13.1s
1 MB/entity 209s

Of course, these are perfect numbers, and 209s of downtime is nothing to write-off. Three minutes of downtime (which will happen on every application restart) might or might not be acceptable. But the full state doesn't have to materialize on startup. It can materialize slowly as requests trickle in.

Wrapping Up

I hold firmly to my hypothesis that entity-component-system is THE way to build apps where data fits on a single machine. Looking at SEC filings for common interactive apps in the gig economy, I believe 1e6 active entities to be a reasonable target point for applications. There's unlikely to ever be that many active entities in a single regulatory domain or geo-fenced area.

tnaps provides an interface in which components and systems are both trivial to implement and easy to test. Testing a system is as easy as instantiating the system and calling process on it in the same way that tnaps itself would.

Circling back to my hypothesis, there is one location that sets the data flow of the application (main). From there, separate concerns can be cleanly isolated into systems over re-usable components. Adding a functionality can be done by adding the data (components) and implementing the code (system). Critically, this only requires coordination with other systems that touch the same components, and Rust + tnaps takes care of that coordination for the most part. You know without a doubt what components are touched by what systems by looking for the run calls into your systems. Similarly, you know what data changes because of guarantees made by the type system.

Looking Forward

I want to validate the hypothesis behind tnaps, and take away the assumption of operating on a single-computer domain without persistence.

Disclaimer

As the title of this document would imply, tnaps is not affiliated with Sony. I love my PlayStation and tnax and tnan don't have the same ring as tnaps.

No runtime deps