[onerng talk] Dieharder Test Failed with 256MB/512MB samples

[Paul Campbell]

On Thursday, 4 October 2018 12:37:37 PM NZDT Jim Cheetham wrote:
> I'd normally wait for Paul to answer you, but I didn't want you to think
> we're ignoring your email 

sorry, I'm quite snowed with other work at the moment, I hadn't forgotten to 
reply - I'd second what Jim has just said - and add a couple of things:

1) we fully admit that our RNG is not perfect, NO RNG is perfect, instead the 
trick is to quantify the imperfection and deal with it appropriately.

Running code through a whitener (a software RNG) is a good way to deal with 
temporal issues. We use an avalanche diode - it generates big and small, and 
in between, avalanches (pulses) - the big ones occasionally swamp the high 
freq data (the small ones) with low freq data ... (this is the temporal thing 
I worry most about)

We deal with not being perfect by telling the kernel RNG (thru rngd) that we 
are generating less that 1 bit/bit of entropy, so it reads more data from the 
OneRNG - you can tweak this by editing /etc/onerng.conf 

2) there was a bug we fixed quite recently (the latest software release) on 
some devices that resulted in some stretches of bad data - make sure you are 
using that latest code, which flushes the output pool before you use the device 
(this problem tended to make it fail badly, not slightly)

	Paul