Re: AES performance

[Breen Ouellette <the.man@breeno.net>]

Theo,

    One question that has been nagging in the back of my mind is, what 
is the quality of the new C3 like, in general?  The numbers you guys 
have been getting from the C3 for AES are incredible, but historically 
the C3 has had some reliability problems.  You guys seem to have been 
dealing with it a lot lately, so do you (or anyone else on the team) 
have any words about the overall quality of the new C3?

    As usual, you guys rock

    -Breen

Theo de Raadt wrote:

>Got a couple figures on AES performance... the cool one at the bottom.
>
>type             16 bytes     64 bytes    256 bytes   1024 bytes   8192 bytes
>aes-128-cbc      15985.10k    16924.93k    17238.12k    17319.74k    17339.74k
>  1.6GHz amd64 in 32bit mode
>
>aes-128-cbc      13323.36k    14403.55k    13225.65k    14529.80k    14654.16k
>  P3/1GHz
>
>aes-128-cbc      13090.59k    51065.12k   174593.45k   426600.92k   735548.02k
>  VIA C3 with the xcrypt-* instructions.  This is using a new diff
>  I have written which makes OpenSSL directly use the cpu instructions
>  if they are available, right in userland, without having to call to
>  /dev/crypto
>
>Two things are apparent.  First, this CPU feature is really cool.
>Second, the OpenSSL glue above crypto operation has pathetically high
>overhead...