• artemis-odb 0.4.0, 0.9.0, 1.0.1, 1.5.0 (prel.)
• ashley 1.6.0
• gdx-artemis 0.5.0
• retanizer 0.2.0

• All benchmarks measure throughput; higher score is better.
• Attempt to be as similar as possible across frameworks.
• Run with entityCount: 1024, 4096, 16384 and 65536.
• Share the same random seed.
• Raw, unprocessed results, here: https://github.com/junkdog/entity-system-benchmarks/tree/master/results
• Additional data:
  • OpenJDK Runtime Environment (IcedTea 2.5.5) (7u79-2.5.5-0ubuntu1)
  • Intel(R) Core(TM) i7-4700HQ CPU @ 2.40GHz

• Have 3 entity systems updating a position component.
• An Entity Deleter system: each 100th tick an entity is deleted and a new entity is created the next tick.
• There are 3 types of iteration benchmarks:
  • baseline: position isn't updated, not included in charts.
  • plain: position component is the default component type.
  • pooled: pooling all components.

• Chooses from up to 144 different entity compositions.
• Each tick, removes/inserts 1/4 of the total entityCount.
• When creating entities, a random composition is selected.
  • The composition is randomly chosen, but shared across all frameworks/benchmarks.

• Have 3 BaselinePositionSystems.
• Each tick, toggles the Position component for 1/4 all entities.
  • Entities are chosen from a pre-shuffled list of entity ids.
• Two types of benchmarks:
  • entity_edit: Toggling achieve by calling entity.edit().create/remove
  • transmuter: Toggles using the EntityTransmuter class.

As usual, writing microbenchmarks is tricky business - take the numbers with a grain of salt. These benchmarks should not be interpreted as benchmarking each framework as a whole, though they aim to give a hint of each framework's performance.

• Overall, ashley-realted benchmarks may be improved; I'm not very familiar with the framework.
• I never got pooling working with Ashley-1.2.0. Ashley-1.0.1:s pooling benchmarks had an error, but considering it's now so old, I'd rather fix the benchmark for the most recent version.
• artemis-odb's PackedComponent benchmarks only have a single component implementation, skewing the benchmark results in its favor.

Artemis-odb-0.7.0 added support for automatically rewriting the generated class files of EntitySystems in order to achieve higher performance. The source code itself remains identical. More about it on the artemis-odb wiki: Hotspot Optimization.