The most used "benchmark" for "randomness" of pseudorandom generators is George Marsaglia's Diehard suite. At the moment the most accepted pRNG is Mersenne Twister; but it is not very suitable for microcontrollers - it's too memory and processing time hungry.

However, it is difficult to tell, what is good or bad it pRNG's; this depends on the intended purpose. A LFSR is adequate in many applications.

But even if you would have a perfect true (physical) random generator, with improper use you can spoil it completely. What is "proper use" is not quite trivial to define.

An example: Using the whole length of LFSR as a pRNG output after each step (shift-by-one-bit) is definitively bad - clearly there is too much correlation between successive outputs.

Isn't this an another topic deserving to become a FAQ...created cooperatively within a wiki-sh environment, perhaps?...

Jan Waclawek