Specs faster than Bevy for numerical simulation?

[ElliotB256]

I'm using the two examples listed in this repo. Stats are for 1a6b9fd.

The minimal example consists of a velocity-verlet integration of non-interacting particles in a harmonic trap. It consists of three systems: (i) integrate position, (ii) update force, (iii) integrate velocity. The benchmark integrates the following particle numbers for N steps:

pub const PARTICLE_NUMBER : i32 = 100_000;
pub const STEP_NUMBER : i32 = 1_000;

I'm performing my profiling on a workstation with an i7-8700 CPU, which has 6 cores and 12 threads. The profiling is performed using Intel VTune, Microarchitecture Exploration runs.

Results

My tests show bevy currently underperforms specs by at least an order of magnitude.

Total time

bevy: 7.417s

specs: 0.576s

CPU Load

bevy:

specs:

Core Utilisation

bevy:

specs:

Summary

bevy:

specs:

Disclaimer

I might well be an idiot. It's my first time using bevy, there could be something very obviously incorrect with this. I'd still be grateful to hear that :)

[ElliotB256]

Discussions still ongoing.

In the meantime, added some examples to compare against legion, using 39a2723.

Total Time

legion: 1.5s

CPU Load

Core Utilisation

Summary

[ElliotB256]

It was foolish of me to not separate the entity creation time from the loop run time, so I'll add that for each below:

ECS	create entities	run loop
bevy	0.65	6.78
specs	0.03	0.54
legion	0.82	0.70

TheRawMeatball took a look and made some changes, I have not yet added those bench numbers

[ElliotB256]

Updates to bevy example, now running on 9d56a6f.

Total time

bevy: 1.12

Note: Graphs below use a longer sim of 5xN steps.

CPU Load

Core Utilisation

Summary

[ElliotB256]

Including new results:

ECS	create entities	run loop
bevy	0.01	1.11
specs	0.03	0.54
legion	0.82	0.70

[minghuaw]

Since bevy 0.6 was just released, I have run the benchmark with bevy updated to 0.6 and have everything else kept the same. Unfortunately, on my machine (i7-7700hq, 16g, Ubuntu 21.10), I did not see a significant improvement for this particular benchmark.