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

[Paul Campbell]

On Thursday, 4 October 2018 4:43:05 PM NZDT Victor Sun (孫國偉) wrote:
> My onerng.conf is default value
> ONERNG_START_RNGD="1"
> ONERNG_MODE_COMMAND="cmd0"
> ONERNG_VERIFY_FIRMWARE="1"
> ONERNG_AES_WHITEN="1"
> ONERNG_URANDOM_RESEED="0"
> ONERNG_ENTROPY=".93750"
> 
> I capture the file is by this command
> /sbin/onerng.sh daemon ttyACM0
> dd if=/dev/random of=/media/sf_Downloads/OneRNG bs=256 count=4M
> iflag=fullblock &
> 
> OneRNG is a 1024MB file and I split it 256*4 / 512*2 /1024*1 for dieharder
> test

so running it thru /dev/random you are IMHO measuring the validity of the 
kernel CSPRNG far more than that of OneRNG - I think that if you did it 
without OneRNG you're likely to get roughly the same result (randomly) but an 
awful lot slower.

(bear in mind when testing random systems like this - you are getting random 
data, results will be different every time, even random test failures. 
Theoretically there's some tiny tiny chance you'll get a file full of 0s and it 
would still be random (insert that XKCD cartoon here). The larger the sample 
you take the more chance you are getting something close to the system's 
actual performance)

BTW ONERNG_ENTROPY is the correction factor mentioned in the previous email - 
the actual value we've measured from real OneRNGs is better than 0.93750 but 
we're being a little paranoid here on purpose, if you want to be even more 
paranoid, you can reduce this value closer to 0 - 0.5 will pull 2 bits from 
onerng for ever bit /dev/random produces)

The way that we test OneRNGs is essentially what you're doing, but instead we 
kill rngd and then dd directly from /dev/ttyACM0 rather than /dev/random 
(which passes through the kernel CSPRNG)

	Paul