I'd been trying to track down a behavioural difference between
non-crc and crc enabled filesystems that was resulting non-crc
filesystem executing prefetch almost 3x faster than CRC filesystems.
After amny ratholes, I finally stumbled on the fact that btree
format directories are not being prefetched due to a missing magic
number check, and it's rejecting all XFS_BMAP_CRC_MAGIC format BMBT
buffers. This makes prefetch on CRC enabled filesystems behave the
same as for non-CRC filesystems.
The difference a single line of code can make on a 50 million inode
filesystem with a single threaded prefetch enabled run is pretty
amazing. It goes from 3,000 iops @ 50MB/s to 2,000 IOPS @ 800MB/s
and the cache hit rate goes from 3% to 49%. The runtime difference: