[ih] End to end checksums

[Lawrence Stewart]

As an aside, the PARC Universal Packet included a 16 bit software checksum calculated by one’s complement addition of 16 bit words plus a 1-bit left cycle. The objective was to give equal protection to all bits.  No one would call the Alto “fast” but the checksum cost was deemed acceptable.

The decision about whether to have end-to-end checks on communications and storage systems depends both on the probability of undetected errors and on the data volume.

For example, hard drives generally have undetected read error rates of 10^-14.  That probably once seemed pretty good, but it is only a dozen terabytes!  Essentially all file systems and all communications protocols should have end to end integrity checks.

Of course the situation is made worse by skewed error rates.  Some particular devices will have error rates much higher than average, so you cannot depend on that 10^-14 anyway.
 
-L