Question

Almost every pseudorandom generator in C/C++ (Mersenne, …) uses some kind of internal state, usually a short vector of bytes. My question is, when such a random generator is used and shared in a multithreaded environment is it “much” better to have it thread-safe or letting “race conditions” to occur can only increases randomness?

I know this question is extremely hard to answer rigorously but will appreciate any opinions.

Answer 1

Concurrent write

It highly depends on the kind of internal state. If every possible bit pattern is a valid representation of internal state, and will therefore occur at some point of the random number sequence, then having write races should be no problem. But many random number generators, including the Mersenne you quoted, have a period which is not a power of 256, therefore have some state patterns which are never reached in single-threaded operation and might cause problems in multi-threaded operation.

Concurrent read

But there is an even better reason to make the rng thread-safe: otherwise two processes might read the same state before either one can update it. This can lead to two processes sharing exactly the same random number, which can lead to all kinds of bizarre problems, depending on your application. You can make it thread-sdafe either using mutexes or thread-local state.