[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: new to VPN
Paul Koning wrote:
>>>>>>"Joe" == Joe Touch <touch@ISI.EDU> writes:
>>>>>
>
> Joe> 1) For some algorithms, such as DES, hardware can be 10-100x
> Joe> faster than software. For others, e.g., MD5, software and
> Joe> hardware have similar speeds.
>
> If you have a software implementation of MD5 that runs at several
> gigabits per second, I'd love to see it...
Well, for grins I dug up the scripts I used to measure MD5 for Sigcomm
back in 1995. I ran them on the following:
Dell Precisin 530 PC
dual XEON 2.4Ghz processors (only 1 of which is used)
1GB RAMBUS RAM
The Sigcomm paper predicted, on a processor with at least 2-way
parallelism (e.g., the XEON), that the limit would be:
My measurements of optimized code (as postted at ?? in 1995) are:
unoptimized (RFC code): 1.03 Gbps +/-0.06
optimized, main memory: 1.604 Gbps +/-0.008
optimized, cache memory: 1.627 Gbps +/-0.001
The 3.06 Ghz processors should achieve over 2 Gbps easily (see below).
That would indeed be multigigabit; this is, IMO, reasonably close. Sure,
you can have custom hardware do better, but vs. riding the PC CPU curve,
there isn't much of a win.
FWIW:
In RFC1810 I approximated 2/(3N) bits/second, where N=ns/instruction.
Ten turns of Moore's law later (7.5 years), it appears that:
2.4 Ghz XEON = 0.4 ns/inst
2 / (3 x 0.4) = 1.666 Gbps predicted
1.627 measured
or roughly 2.4% off... ;-)
Joe