I was only saying that it isn't really a good idea to present numbers like that without having a clear idea of what's causing the worst-case performance. Glancing over the code, and assuming the first bullet fired is always the first one to impact the ground, it looks like the code is always going to run std::vector::erase on the first element of the vector, which will always shift the 499 remaining elements. It's possible that's the reason for the worst-case performance, rather than vector being inherently a worse choice than list.
It was very counterintuitive and weird to realize that vector is almost always a better choice than list due to how powerful the CPU caches are. CPU caching is so good that taking full advantage of caching will usually outweigh algorithmic tweaks, like whether you're using a list or a vector. So the power of vector is that its elements are in contiguous memory, whereas the downfall of list is that they usually aren't.
It's difficult to prove that's true except in real-world scenarios. Unless a test is designed to introduce memory fragmentation, the cache performance penalty of using list won't really show up in the results. All of the list allocations will be pretty much contiguous in a test environment, so the tests will show performance on par with vector. But in the real engine, you'll be left scratching your head why everything's running so slowly, when the answer is because the elements are sprayed all across your memory space and it's causing the CPU cache to stall when they're accessed...
EDIT:
There is no actual rendering time or AI or anything else which would prevent us from seeing upwards of 100,000 "frames" per second.
When your FPS is higher than, say, 1,000, it becomes very difficult to pin down the reasons for performance fluctuations. An engine under load behaves differently than an artificial test, so you'll need to measure the performance impact in a real engine, especially when the framerates in the test are so high.
Your monitor is capped at 70 fps, or some high-end monitors can do ~130. Something like 100,000 fps isn't an accurate reflection of the behavior of the real system. Most of the discrepancy is due to memory fragmentation causing CPU thrashing, but your specific test might be running into some other discrepancies, like:
- clock() might not be a high resolution timer, so the worst-case timing might be due to discrepancies in the clock mechanism itself, especially at 100,000+ FPS. See http://stackoverflow.com/questions/6749621/high-resolution-t...
- since your test is running for 100+ seconds and your timing measurements are taking into account an entire frame, some frames are going to run interrupts. If an interrupt takes a long time to process, that load will show up in your timing measurements. The result is that vector (or list) will be unfairly pinned as the reason for the performance issue, when interrupts were the cause.
I'm not confident about the second one, but the point is that there are all kinds of conflating variables.