On Mon, Jan 7, 2019 at 9:37 AM Eduardo A. Bustamante López <dual...@gmail.com> wrote: > On Mon, Jan 07, 2019 at 08:15:12AM +0100, Ole Tange wrote: > > On Mon, Jan 7, 2019 at 12:08 AM Chet Ramey <chet.ra...@case.edu> wrote: > > > On 1/5/19 3:12 PM, Eduardo A. Bustamante López wrote: > > > > On Fri, Dec 28, 2018 at 10:24:50AM +0100, Ole Tange wrote: > > > > (...) > > > >> Patch attached. > > : > > > > - What is the performance impact (roughly 2X slower) > > > > The impact is not from Salsa20/ChaCha20, though, if you compare > > patched vs. unpatched code. > > > > On my system BASH_RANDOM_16 is a tiny bit faster than the original, > > while BASH_RANDOM_32 is a tiny bit slower. > > I find that really hard to believe. The new RNG is clearly doing more work, so > it should be slower.
It does more work, yes, but it also produces more numbers and the work it does is simpler: When it does 20 rounds it produces 16 or 32 numbers. The calculations in each round are parallelizable and are mostly bit shifting. https://en.wikipedia.org/wiki/Salsa20 says it uses 3.95 cpu cycles per byte. Given that the old version did 1 division, 1 modulo, and 2 multiplications, I would say it is not unreasonable that they are comparable in speed. > And that's OK. I just wanted to know by how much. > > What numbers are you seeing and how did you measure them? I ran your test script on 3 versions: * the original source. Git [3669b0bacc6] * the patched source with BASH_RANDOM_32 * the patched source with BASH_RANDOM_16 and the results were within 1% of each other. (I can, however, reproduce your finding that $RANDOM is twice as slow on 5.0 as on 4.4, but it is not due to the generator). /Ole
