My guess would be that the implementation of Linux's RAID6 is such that it requires a complete recalculation of the parity blocks with each write. Sounds like a massive write-performance hit (due to the requirement of needing to read that block on each disk) to me. But since RAID6 can be implemented in any way (the specification requires only that it be able to survive two simultaneous disk failures) and I don't feel like crawling through source code, I could easily be guessing wrongly.