Fw: [e2e] Re: [Tsvwg] Really End-to-end or CRC vs everything else ?

[Jim Williams]

----- Original Message -----
From: "David P. Reed" <dpreed at reed.com>
To: <vince_cavanna at agilent.com>; <vince_cavanna at agilent.com>;
<jonathan at DSG.Stanford.EDU>
Cc: <tsvwg at ietf.org>; <end2end-interest at postel.org>;
<sommerfeld at orchard.arlington.ma.us>; <joe_steinmetz at agilent.com>;
<matt_wakeley at agilent.com>
Sent: Friday, June 08, 2001 3:56 PM
Subject: RE: [e2e] Re: [Tsvwg] Really End-to-end or CRC vs everything else ?


> Because I proposed secure auth vs. checksum, I've gotten real curious
about
> relative performance of checksums, CRC32, MD5 and MD5-MAC.  Since there is
> an open source project called OpenCL that has implementations of CRC32 and
> MD5, I thought I'd investigate, and found that there were already
> benchmarks of that code.  Their somewhat surprising result is that MD5-MAC
> is slightly *faster* than CRC32 in that implementation.

This is interesting.  I would point out for the record
that hardware to compute MD5-MAC in real time at line
rate for 10Gb is not possible.  The nature for the
MD5 algorithm makes impossible to effectively pipeline
in that each stage of the computation must be
complete before the next can be begun, and so can't
be parallelized.  I would expect that for this reason
the MD5-MAC algorithm may eventually be replaced
by something more hardware friendly.  UMAC is parhaps
an example of a more parallelizable algorithm, which
also claims better software performance.

Computing CRC32 at line rate for 10Gb should not
be a problem with hardware.