3 ms·
One problem with simple CRC calculations is the question of whether it's correct to assume a random distribution of errors: if hardware failures produce cluster
by acdha 10y ago
One problem with simple CRC calculations is the question of whether it's correct to assume a random distribution of errors: if hardware failures produce clustered failures you might find that a particular device or model produces errors at much higher rates.
ftp://ftp.cis.upenn.edu/pub/mbgreen/papers/ton98.pdf
“First, a non-uniform distribution of data makes failure of
the TCP checksum far more likely than one would naively
expect. The undetected splice rate in our data for the 16-bit
TCP checksum over real data is comparable to uniform data
with a 10-bit checksum.
Second, checksum distributions on modest amounts of
real data are substantially different from the distributions one would anticipate for uniformly distributed data. This skewed distribution does result in significantly higher failure rates of the TCP checksum. In particular, if a router or host has a buffering problem that causes adjacent packets to be merged, the TCP checksum might fail .1% of the time rather than the 0.0015% of the time that purely random data distribution would suggest.”
https://users.ece.cmu.edu/~koopman/pubs/paultisch05_dsn_crc_ultradependable.pdf https://users.ece.cmu.edu/~koopman/pubs/paultisch05_dsn_crc_...
http://conferences.sigcomm.org/sigcomm/2000/conf/paper/sigcomm2000-9-1.pdf http://conferences.sigcomm.org/sigcomm/2000/conf/paper/sigco...