AES Modes

[Brian Gladman]

Eric Young wrote:
> Quoting Brian Gladman <brg at gladman.plus.com>:
> 
> 
>>Ian Grigg wrote:
>>
>>
>>>Jack Lloyd also passed along lots of good comments I'd
>>>like to forward (having gained permission) FTR.  I've
>>>edited them for brevity and pertinence.
>>
>>[snip]
>>
>>> >>I'm obviously being naive here ... I had thought that the combined 
>>>mode would
>>> >> be faster, as it would run through the data once only, and that AES 
>>>seems to
>>> >> clip along faster than SHA1.
>>>
>>>AFAIK all of the modes that use only one block cipher invocation per 
>>>block of
>>>input are patented. EAX+CCM both use two AES operations per block, and
>>>byte-for-byte SHA-1 is 2-5x faster than AES (at least in the 
>>>implementations
>>>I've seen/used/written), so using AES+HMAC is probably going to be 
>>>faster than
>>>AES/EAX or AES/CCM. The obvious exception being boxes with hardware AES 
>>>chips
>>>and slow CPUs (eg, an ARM7 with an AES coprocessor), where AES will of 
>>>course
>>>be much faster than SHA-1.
>>
>>Maybe my C implementation of SHA1 is hopeless but I get SHA1 on an x86 
>>at about 17 cycles per byte (over 100,000 bytes) and AES in C at 21 
>>cycles per byte.
>>
>>So I would put these two algorihms at about the same speed in C. In 
>>consequence I rather suspect that the 'two encryptions per block' cost 
>>might also apply to combined modes when AES is used with HMAC-SHA1.
> 
> Are you running on a P4?  ASM for sha1 on a P4 takes about 11.9 cycles 
> per byte.  The P4 is a very touchy x86 implementation.
> On most other architectures I nearly always see a bit less than 2 times faster
> sha1 vs AES.  On AMD64, asm, I have
> AES-cbc at 12.2 cycles per byte and sha1 at 6.8.  This is about
> as good a CPU as it gets (IPC near 3 for both implementations).

The SHA1 figure is for a P3 using VC++ set to generate code that will 
run on all Pentium family machines.  I have not optimised the C code for 
any particular machine. 17/12 for C/ASM is a bit worse than I would have 
hoped for but is not that bad.

I would not be surprised to see an average AES/SHA1 speed comparison in 
the 1.5:2.5 range but I was a bit surprised to see Jack's 2.0:5.0 range.

I will have to see if VC++ can be coaxed down from 17 cycles per byte 
for SHA1 without giving up on code that runs on all Pentium compatible 
machines :-)

    Brian Gladman


---------------------------------------------------------------------
The Cryptography Mailing List
Unsubscribe by sending "unsubscribe cryptography" to majordomo at metzdowd.com