I've done some moderately careful performance apples to apples
comparisons for various crypto functions with not very surprising
results. In general:
AES-NI < RC4 < ChaCha20 < AES (software)
The two AES functions, the Botan software version and the DJ Bernstein
NI (new instructions) version, were each computed with an CBC/CTS
wrapper. All functions were then tested with a Transform wrapper.
The specific numbers from a Win8 Intel Desktop were for 10 MB encode /
decode:
AES-NI: .040 seconds
RC4: .060
ChaCha20: .080
AES: .100
Each number was the floor of three runs, code compiled for Release.
Some notes:
1. As Sean pointed out, the AES instructions are common on Intel
processors. Not so for AMD, however, which only supports AES in
their high end server chips. My HP AMD mini-tower, for example,
doesn't have the AES instruction set. This means that the crypto
systems must do a runtime check for AES hardware support.
2. Simple comparisons between hardware and software AES are next to
meaningless. By the time you handle block chaining and cipher text
stealing, the necessary overhead dilutes the raw performance difference.
3. The Botan AES function was 11X the performance as the AES
implementation I had been using. Details on request.
4. The Botan crypto library has a good license, good performance, and
is a pain in the butt to work with. One of their apparent goals was
to use every possible C++ feature. Pulling AES out of their rich
and bloated infrastructure took more days than I care to think about.
5. The Botan AES-NI implementation doesn't seem to work, at least not
when extracted. Don't know why.
6. The DJ Bernstein AES-NI was as simple and elegant as the Botan
library was complex. It was C that could be dropped into a C++
wrapper without change. And it has a better license.
7. The ChaCha20 implementation is probably less than optimal. I'm
prepared to let it stay that way.
Bottom line: I'm inclined to agree with Sean that AES-NI always wins
and software AES doesn't lose by much, but the difference is nowhere as
much as hyped.
------------------------------------------------------------------------------
Firebird-Devel mailing list, web interface at
https://lists.sourceforge.net/lists/listinfo/firebird-devel
