Is `/dev/urandom` suitable for simulation purpose?

Question

It seems that using plain C in unix-like systems, fread from /dev/urandom is the simplest way to extract high quality random bytes. I need to run a simulation that needs about 10k 32-bit random numbers per second, and it may run for several days. Is /dev/urandom okay to use for this purpose? How is the quality of random bytes from here when the entropy pool is depleted?

edit_1

While I'm now running 3 parallel diehard tests for /dev/urandom in my laptop, I got the following interesting lines. The test isn't yet complete.

#=============================================================================#
        test_name   |ntup| tsamples |psamples|  p-value |Assessment
#=============================================================================#
 diehard_parking_lot|   0|     12000|     100|0.99573896|   WEAK
        diehard_sums|   0|       100|     100|0.00116464|   WEAK
          sts_serial|   7|    100000|     100|0.99996076|   WEAK

Answer 1

In the underlying implementation of /dev/urandom is a CSPRNG, the output pool of which has a maximal period of less than 2^(26∗32) − 1, which is then fed into SHA-1 to produce output for /dev/urandom. As such, urandom can obviously produce the amount of random numbers you want, however it can not offer you reproducible results - you will have to cache the sequence you get yourself.

You do not have to worry about what happens when the entropy pool is estimated to be depleted, /dev/urandom will output whatever you request of it. The "theoretical attacks" the urandom(4) man page speaks of are nonexistent. (the "issue" is a huge misunderstanding of what "entropy estimation" is)

Many other PRNGs with large periods exist which reproducible seeding: the Mersenne Twister in C++, xorshift PRNGs, etc. You should be able to adapt any PRNG to the distribution which is suitable for your purposes.