I need to generate around 9-100 million non-repeating random numbers, ranging from zero to the amount of numbers generated, and I need them to be generated very quickly. Several answers to similar questions proposed simply shuffling an array in order to get the random numbers, and others proposed using a bloom filter. The question is, which one is more efficient, and in case of it being the bloom filter, how do I use it?
You don't want random numbers at all. You want exactly the numbers 0 to N-1, in random order.
Simply filling the array and shuffling should be very quick. A proper Fisher-Yates shuffle is O(n), so an array of 100 million should take well under a second in C or even Java, slightly slower in a higher-level language like Python.
You only have to generate N-1 random numbers to do the shuffle (maybe up to 1.3N if you use rejection sampling to get perfect uniformity), so the speed will depend largely on how fast your RNG is.
You'll never need to look up whether a number has already be generated; that will deadly be slow no matter which algorithm you use, especially toward the end of the run.
If you need slightly fewer than N total numbers, fill the array from 0 to N-1, then just abort the shuffle early and take the partial result. Only if the amount of numbers you need is very small compared to their range should you consider the generate-and-check-for-dups approach. In that case Bob Floyd's algorithm might be good.