As I gain confidence in that PEAC thing, I think that soon should come the time to compare its behaviour with a big competition. This will be even easier once I get my Pi cluster in working order.

It's only fair to compare the 32-bit version with other 32-bit checksums, in particular Fletcher32 and CRC32.

Maybe I can also compare with some 16-bit algos with a "compacted" PEAC signature. ADD16, Fletcher16, CRC16 and IPsum would be good candidates.

This will also be an opportunity to compare the effect of the end-around carry, and settle the debate concerning the mixing method : whole word or half word ? XOR or ADD ?

Considering that on average, only one error will be detected for every 4 billion tests, this is going to be slow and require a smart algorithm running on tens of cores to collect data,

The next question is : what is the expected bit error model ? What is the buffer's length ?

Another thing to measure and compare is the histogram of the signatures for small blocks of small values.