| |||||||||||||
Good pieceGood piecePosted Nov 20, 2012 21:42 UTC (Tue) by epa (subscriber, #39769)In reply to: Good piece by man_ls Parent article: LCE: Don't play dice with random numbers
I imagine that the HAVEGE test involved replacing the CPU-nondeterminism part with a source that returns all 1s. However this is not the last stage - it goes through a PRNG afterwards. So all the supposed testsuite was showing was that the PRNG wasn't totally awful - it did not say anything about the quality of the input to that PRNG. That is my uninformed guess based on reading the article.
(Log in to post comments)
Good piece Posted Nov 20, 2012 21:46 UTC (Tue) by dlang (✭ supporter ✭, #313) [Link]
But it does say a lot about the quality of the tests that the result is passing.
If the tests can't tell the difference between it being fed all '1's and random input, then the tests are pretty worthless in terms of evaluating it.
Remember that the claim is that this is far better than a PRNG by itself, so to show this, you need to have a test that fails on a PRNG, but passes with the HAVEGE hardware input.
Either that or you need to explain why HAVEGE is better, not just say "see, we studied it and it passed all the tests", which is what HPA is complaining that most of the papers on it boil down to.
Good piece Posted Nov 20, 2012 22:02 UTC (Tue) by mkerrisk (subscriber, #1978) [Link]
> If the tests can't tell the difference between it
> being fed all '1's and random input, then the > tests are pretty worthless in terms of evaluating it.
To be clear: it is not the *tests* that are being fed all 1s. It is HAVEGE, which then anyway produces good enough PRNG outputs that the test pass.
|
|||||||||||||