[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: Remove little-used algorithms from IKEv2
> I found it to be about 15% faster than SHA-1, and that on a big endian
> machine.
OTOH, if your big-endian machine can dance the little-endian dance better,
you get noticeable speedup. The UltraSPARC-tuned MD5 is 30% faster than the
non-UltraSPARC-tuned MD5. Using your 15% improvement as a base...
M is MD5
S is SHA-1
M' is tuned MD5
M = 0.85S
M' = 0.70M
M' / 0.70 = 0.85S
M' = 0.70 * 0.85S
M' = 0.59S
If you want to see what I mean precisely, utter this on a sun4m (e.g. SPARC
20) machine running Solaris 7 or later:
dis -F MD5Transform /kernel/misc/md5
and utter this on a sun4u (e.g. any UltraSPARC box)
dis -F MD5Transform /platform/sun4u/kernel/misc/md5
and note the instruction count savings.
> That number makes sense given the structure of the two
> algorithms. So, somewhat better, yes. "Far better", no.
See my math above for SW and dancing the little-endian dance.
> In hardware implementations, the two tend to be pretty close, and
> usually faster than the encryption transform so it doesn't matter
> which you chose as far as performance goes.
Now _this_ I'll buy.
Dan