A bit of benchmarking ...

[Andreas Schneider]

On Friday, 1 March 2019 22:56:31 CET Jeremy Allison wrote:
> On Fri, Mar 01, 2019 at 08:34:59PM +0100, Stefan Metzmacher wrote:
> > >>> It would be wonderful to move from any home-grown
> > >>> crypto to GnuTLS for everything !
> > >> 
> > >> I've tested AES-GCM with Samba Crypto (Intel AES NI) and GnuTLS.
> > >> 
> > >> TL;DR
> > >> 
> > >> Samba Crypto (Intel AES NI):  10 min
> > >> GnuTLS:                       12 sec
> > >> 
> > >> https://hackmd.io/s/S1bJnG4IN
> > > 
> > > Wow, that is a .. significant change. Any ideas
> > > on why the difference is so great ?
> > 
> > Doing xor and lshift/rshift operations in C is what kills us (even when
> > using aesni).
> 
> Thanks for the info Metze, much appreciated !

GnuTLS has several optimized versions of AES GCM implementation, depending on 
what features the CPU provides like AES NI, SSE3 and AVX.

I've requested a better API for AEAD ciphers (CCM and GCM) so we can get rid 
for that allocations and memcpy().

https://gitlab.com/gnutls/gnutls/issues/718

I've opened a RHEL bug too. This way we might get some resources to implement 
in the next month.

Currently I have some issues with AES128 CFB8. Not sure if there is a bug in 
the nettle implementation or in GnuTLS or I have to do something special.


Cheers,


	Andreas