luiswirth
commented
2 months ago Replace bevy ECS with evenio
no more bevy :(?
Closed rj00a closed 1 month ago
luiswirth
commented
2 months ago Replace bevy ECS with evenio
no more bevy :(?
kehrazy
commented
2 months ago Replace bevy ECS with evenio
no more bevy :(?
kill bevy with FIRE! 🔥
OvermindDL1
commented
2 months ago Replace bevy ECS with evenio.
May I ask why? bevy_ecs outperforms it on its own benchmark everywhere except on random access, where bevy is twice the time (because it's a dual lookup for an entity->archetype->data), but of times so tiny they are just a pointer indirection cost anyway?
Here's my benchmark times using evenio's own benchmark with the latest version of both evenio and bevy_ecs (so git for evenio and 0.14.2 for bevy, no code changes needed for their benchmark code, it all just worked) for their benchmarks that have bevy in it:
add_components(honestly I expect bevy to be slightly slower here as archetypes styles often are due to more copies, so... weird?):
Timer precision: 70 ns add_components fastest │ slowest │ median │ mean │ samples │ iters ├─ spawn_many_comps_40_bevy 1.662 µs │ 44.46 µs │ 1.705 µs │ 2.952 µs │ 100 │ 1000 ╰─ spawn_many_comps_n_evenio │ │ │ │ │ ├─ 1 276.1 ns │ 1.73 µs │ 282.2 ns │ 305.4 ns │ 100 │ 1000 ├─ 2 463.1 ns │ 2.065 µs │ 471.1 ns │ 506.9 ns │ 100 │ 1000 ├─ 3 688.1 ns │ 2.177 µs │ 699.1 ns │ 735.2 ns │ 100 │ 1000 ├─ 5 1.189 µs │ 3.053 µs │ 1.212 µs │ 1.253 µs │ 100 │ 1000 ├─ 10 2.637 µs │ 6.394 µs │ 2.664 µs │ 2.762 µs │ 100 │ 1000 ├─ 15 2.67 µs │ 11.34 µs │ 4.419 µs │ 4.437 µs │ 100 │ 1000 ├─ 20 2.762 µs │ 11.69 µs │ 3.285 µs │ 3.384 µs │ 100 │ 1000 ├─ 25 3.339 µs │ 11.59 µs │ 3.916 µs │ 4.176 µs │ 100 │ 1000 ├─ 30 5.099 µs │ 16.01 µs │ 5.131 µs │ 5.406 µs │ 100 │ 1000 ╰─ 40 6.76 µs │ 20.27 µs │ 7.936 µs │ 8.12 µs │ 100 │ 1000
iter(I expected bevy to be faster here, and it indeed was in much of it, but there seems to be a weird bit of noise... somewhere, it's not scaling for either of them as it should on the mandelbrot stuff... Also I find it odd they only look up a single component...):
Timer precision: 70 ns iter fastest │ slowest │ median │ mean │ samples │ iters ├─ iter_mandelbrot_bevy │ │ │ │ │ │ ╰─ true │ │ │ │ │ │ ├─ 1 38.87 µs │ 106.5 µs │ 64.03 µs │ 66.08 µs │ 100 │ 100 │ ├─ 2 48.01 µs │ 130.8 µs │ 71.97 µs │ 72.85 µs │ 100 │ 100 │ ├─ 3 63.36 µs │ 133.7 µs │ 97.65 µs │ 98.05 µs │ 100 │ 100 │ ├─ 8 127 µs │ 213.1 µs │ 154.8 µs │ 156.5 µs │ 100 │ 100 │ ├─ 16 199.2 µs │ 598.1 µs │ 239.2 µs │ 277.9 µs │ 100 │ 100 │ ├─ 32 220.5 µs │ 828.7 µs │ 337.8 µs │ 349.6 µs │ 100 │ 100 │ ├─ 64 431.3 µs │ 673.7 µs │ 540 µs │ 538.6 µs │ 100 │ 100 │ ├─ 128 454.5 µs │ 1.673 ms │ 759.5 µs │ 773.4 µs │ 100 │ 100 │ ├─ 256 834.4 µs │ 2.127 ms │ 1.437 ms │ 1.459 ms │ 100 │ 100 │ ├─ 512 2.077 ms │ 3.302 ms │ 2.748 ms │ 2.764 ms │ 100 │ 100 │ ╰─ 1024 2.737 ms │ 9.691 ms │ 5.098 ms │ 4.412 ms │ 100 │ 100 ├─ iter_mandelbrot_evenio │ │ │ │ │ │ ╰─ true │ │ │ │ │ │ ├─ 1 56.97 µs │ 1.506 ms │ 111.4 µs │ 156.6 µs │ 100 │ 100 │ ├─ 2 63.31 µs │ 775.3 µs │ 186.8 µs │ 189.3 µs │ 100 │ 100 │ ├─ 3 62.99 µs │ 503.5 µs │ 78.32 µs │ 128.9 µs │ 100 │ 100 │ ├─ 8 74.49 µs │ 375.4 µs │ 141.4 µs │ 159.8 µs │ 100 │ 100 │ ├─ 16 90.65 µs │ 978.5 µs │ 229.7 µs │ 242.3 µs │ 100 │ 100 │ ├─ 32 110.7 µs │ 919.3 µs │ 243.2 µs │ 247 µs │ 100 │ 100 │ ├─ 64 145 µs │ 1.086 ms │ 337.8 µs │ 343.9 µs │ 100 │ 100 │ ├─ 128 227.7 µs │ 800.8 µs │ 449.3 µs │ 465.5 µs │ 100 │ 100 │ ├─ 256 411.8 µs │ 2.352 ms │ 636.3 µs │ 662.1 µs │ 100 │ 100 │ ├─ 512 687.1 µs │ 1.418 ms │ 709 µs │ 765.7 µs │ 100 │ 100 │ ╰─ 1024 1.291 ms │ 2.342 ms │ 1.357 ms │ 1.464 ms │ 100 │ 100 ├─ iter_simple_bevy │ │ │ │ │ │ ├─ false │ │ │ │ │ │ │ ├─ 1 7.806 ns │ 44.63 ns │ 11.79 ns │ 11.98 ns │ 100 │ 25600 │ │ ├─ 10 18.97 ns │ 51.59 ns │ 18.99 ns │ 19.31 ns │ 100 │ 51200 │ │ ├─ 100 64.47 ns │ 135.7 ns │ 64.48 ns │ 65.53 ns │ 100 │ 12800 │ │ ├─ 1000 630.3 ns │ 2.205 µs │ 630.5 ns │ 685.9 ns │ 100 │ 800 │ │ ├─ 10000 6.298 µs │ 25.85 µs │ 12.55 µs │ 10.02 µs │ 100 │ 100 │ │ ├─ 100000 53.55 µs │ 150.1 µs │ 125.1 µs │ 99 µs │ 100 │ 100 │ │ ╰─ 1000000 594 µs │ 1.416 ms │ 1.115 ms │ 1.034 ms │ 100 │ 100 │ ╰─ true │ │ │ │ │ │ ├─ 1 1.789 µs │ 51.56 µs │ 1.958 µs │ 5.421 µs │ 100 │ 100 │ ├─ 10 4.938 µs │ 41.98 µs │ 9.783 µs │ 11.13 µs │ 100 │ 100 │ ├─ 100 13.95 µs │ 74.89 µs │ 18.28 µs │ 21.41 µs │ 100 │ 100 │ ├─ 1000 20.77 µs │ 90.14 µs │ 30.4 µs │ 33.6 µs │ 100 │ 100 │ ├─ 10000 25.73 µs │ 186 µs │ 42.01 µs │ 44.17 µs │ 100 │ 100 │ ├─ 100000 67.79 µs │ 240.5 µs │ 111.2 µs │ 121.2 µs │ 100 │ 100 │ ╰─ 1000000 217.9 µs │ 937.3 µs │ 476 µs │ 477.7 µs │ 100 │ 100 ╰─ iter_simple_evenio │ │ │ │ │ ├─ false │ │ │ │ │ │ ├─ 1 30.02 ns │ 658.6 ns │ 30.38 ns │ 131.9 ns │ 100 │ 25600 │ ├─ 10 37.88 ns │ 86.93 ns │ 40.53 ns │ 41.42 ns │ 100 │ 25600 │ ├─ 100 151 ns │ 271 ns │ 151.4 ns │ 153.3 ns │ 100 │ 3200 │ ├─ 1000 1.284 µs │ 4.563 µs │ 1.286 µs │ 1.465 µs │ 100 │ 400 │ ├─ 10000 12.57 µs │ 33.79 µs │ 25.07 µs │ 19.54 µs │ 100 │ 100 │ ├─ 100000 106.8 µs │ 308.9 µs │ 220.3 µs │ 194.3 µs │ 100 │ 100 │ ╰─ 1000000 1.341 ms │ 4.775 ms │ 2.521 ms │ 2.478 ms │ 100 │ 100 ╰─ true │ │ │ │ │ ├─ 1 59.16 ns │ 115.6 ns │ 59.24 ns │ 64.57 ns │ 100 │ 12800 ├─ 10 12.66 µs │ 260.4 µs │ 34.37 µs │ 48.77 µs │ 100 │ 100 ├─ 100 20.55 µs │ 252.3 µs │ 31.27 µs │ 39.79 µs │ 100 │ 100 ├─ 1000 42.22 µs │ 201.3 µs │ 89.62 µs │ 94.8 µs │ 100 │ 100 ├─ 10000 58.51 µs │ 438.7 µs │ 118.2 µs │ 152.4 µs │ 100 │ 100 ├─ 100000 79.11 µs │ 710.3 µs │ 232.8 µs │ 257.6 µs │ 100 │ 100 ╰─ 1000000 289 µs │ 2.196 ms │ 765 µs │ 866.6 µs │ 100 │ 100
random_access(I expect bevy to be twice the cost here since 2 pointer indirections instead of one for the secondary index/cache, but again, these are so fast that the times are just noise, the test has to have a million entities to even get to the times it's at, which is literally just memory pressure timings at this point as they run literally no code beyond a black_box access in the tests, doing any work at all will flatten this):
Timer precision: 29 ns random_access fastest │ slowest │ median │ mean │ samples │ iters ├─ random_access_bevy 1.842 ms │ 3.721 ms │ 2.02 ms │ 2.18 ms │ 100 │ 100 ╰─ random_access_evenio 917.1 µs │ 1.279 ms │ 1.052 ms │ 1.056 ms │ 100 │ 100
dyc3
commented
2 months ago I can't find the discussion right now, but this has been discussed before. We are looking to replace bevy primarily because of development ergonomics. A lot of the code we write relies very heavily on events that trigger logic.
JackCrumpLeys
commented
2 months ago I personally don't think any developers are motivated to continue this rewrite. At the moment @rj00a has mysteriously disappeared and I seem to be the main person still working on it. My development effort is currently focused on 1.21 and after that I would rather add features then undertake a rewrite. If someone with a lot of spare time and motivation shows up this might happen but it seems unlikely.
giantcow
commented
2 months ago The evenio docs eludes to some of the reasoning https://docs.rs/evenio/latest/evenio/tutorial/#why-evenio
OvermindDL1
commented
2 months ago A lot of the code we write relies very heavily on events that trigger logic.
The evenio docs eludes to some of the reasoning https://docs.rs/evenio/latest/evenio/tutorial/#why-evenio
bevy_ecs also handles events, though it does it in a more controllable way with less overhead as they are batched and run at-once in a system with parallel support. What it looks like evenio does from my little of experimenting with it is it calls event handlers immediately, which is equivalent to bevy_ecs's ability to run a system on demand (many times in a frame if you so wish), except bevy's has less virtual dispatch overhead as the archetypes can be immediately known, statically if so wished (or even dynamically dispatched if you wanted to go that way, ala' how evenio does it).
I experimented with evenio but I didn't see any feature it had that bevy_ecs lacked, and bevy_ecs has a great many features that it lacks, and it had no performance boost to offset it either (and indeed you lose out in a great amount of performance in many areas with multi-threading as it can only multi-thread 'within' a system (if the user chooses to), not 'across' systems, and you cannot often multi-thread within a system and even most of the time when you can it's not worth it due to the task dispatch overhead and the low amount of iterations within most systems).
But yeah, in short if you want to run something to handle an 'event' now then bevy has run-on-demand systems for that, but honestly you shouldn't except in specialized cases (which do exist), that's needless performance overhead for no reason otherwise and makes it harder to follow 'when' something happens.
dyc3
commented
1 month ago I'm going to close and unpin this issue since the main driver behind it (@rj00a) is MIA for now. We can always reopen it later.
IMO, the changes suggested in this issue (particularly the switch to evenio) are a non-starter because it requires too much effort. We are better off continuing to iterate on the current version.
Here are the steps that need to be taken to rewrite the project and get it updated to the latest version of the game, since folks were asking.
Should be done roughly in this order, one or more PRs per check.
criteriontodivanfor benchmarking.evenioand depend on that.Arcs for easy sharing and fast lookups. Registry modification is then possible if the refcount is 1.compio.valence_anvil,valence_command.valence_protocol. Might need a rewrite.