Re: [Tsvwg] [SCTP checksum problems]

[Craig Partridge <craig@aland.bbn.com>]

Hi Julian:

I read your ID quickly and may have more comments later.

I think the bit error study is flawed.  The key issue here is that
the TCP, and Fletcher checksum (and I think Adler) will catch all
1 bit errors.  Indeed, the TCP checksum will catch up to 16 distinct
bit errors, provided that the bit offsets are all distinct, modulo 16.

Expressing this simply, assuming my system did its Poisson calculations
right:

    * we don't have to worry about single bit errors in a chunk of
      data, the checksum will always catch it.

    * the probability of 2 bit errors occuring in a single 1KB packet
      is about 1 in 3* 10^15.  The probability that the two bits will
      have the same offset mod 16 is around .008 (if I did the distribution
      right) and the chance is 50% that the two bits will offset each other
      (i.e. they'll flip opposite ways).  Ergo, your chance of an
      undetected error is about 1 in 8*10^17.  I believe (without having
      done all the match) that is as good or better than you'll get for
      CRC32 in almost all cases.  Since the checksums are cheaper to compute,
      they're a win.

    * more than 2 bit errors per chunk is so rare we probably don't care

So the big issue is burst errors (and, incidentally, burst erasures).

Thanks!

Craig