Replace shuffled vector with an LCG-based shuffled access pattern.

[ssbr]

This should reduce/remove measurement artifacts from the shuffle vector. Since the shuffle vector is, itself, the same size as the buffer being tested for cache, it can lower the measured value by a factor of 2. In fact, it can reduce it by more than 2, because the buffer is of char and the shuffle vector is of int32 or even int64 -- meaning it could cut it by a whole order of magnitude!

Doing some concrete tests on my machine, with the new shuffle we can see a latency increase that really ramps up around 1MB (size of L2 on my laptop):

But with the old algorithm, it starts substantially earlier -- the median stabilizes to a new high value at aroung 100K, a size corresponding to nothing in particular as far as I'm aware:

[mmdriley]

This isn't a linear congruential generator.

In an LCG the output is iteratively fed back into the input to create a pseudorandom sequence. Here, though, we're just passing the numbers 0..n-1 into this function expecting to get a permutation of 0..n-1 back.

This works because we're generating the finite cyclic group of integers under addition modulo m. That cyclic group is generated by any element coprime to the modulus. Here m is 2⁸, so any odd number is coprime. ref.

We added the 100 to avoid mapping 0 to 0. Unlike an LCG, it's not necessary to ensure we cover the range.