RE: Security Gateways

[Jim McKinney <jmck+@ece.cmu.edu>]

Approved: sskm1

> Bernard Aboba wrote:
> 
> Integrity protection via new MACs such as UMAC is *not* 
> expensive. This is
> a myth. See http://www.cs.ucdavis.edu/~rogaway/umac/perf00.html 
>

Bernard,

Thanks for the pointer to some current research in this area.

That paper gave cycles/byte numbers for their algorithm, on 
a Pentium III and PowerPC.  For now, I'll comment on the PPC 
numbers, because that family is available as an embeddable core
that can be integrated into hardware, while the Pentium III is not. 
For 1500-byte messages, it looks like UMAC will take between 
2.7 and 3.7 CPU clocks per byte (estimates, as no numbers 
were given for the full UMAC function on PPC).  

That indicates that I'd have to run my core at 600-800 Mhz 
to handle a single full duplex Gigabit Ethernet port, assuming
I have the packets in memory that's fast enough to not stall the 
CPU, and that there's some way of getting packets into and out 
of that memory without interfering with the CPU.

I grant you that is "do-able," and it certainly is less expensive 
than HMAC-SHA1, but I wouldn't call it "cheap."  At least for 
some portion of the audience here, "cheap" may be measured by how 
many thousands of gates the implementation adds to the existing 
data path logic, and "fast" by how many gate delays were 
introduced into that path. 

So, I guess that puts me in agreement with Jim William's 
comments; security protocols MUST be supported, and 
supported in the context of a single standard, 
not a "de jure" standard and a "de facto" alternative.
And, if that means the standard must embrace non-monolithic 
"system level" compliance, then let's figure out how to 
do that properly, rather than declaring it illegal and hoping 
it will go away.

- milan