ZFS large recordsize use case

[Elliott]

I have been benchmarking pools with different recordsize for a workload of sequential read and write for very large files. I find that throughput increases with recordsize up to 1M. I tried larger sizes of 2M, 4M, and 16M, but it seems like 1M is the sweet spot. I noticed the warning:

Code:

# sysctl -d vfs.zfs.max_recordsize
vfs.zfs.max_recordsize: Maximum block size.  Expect dragons when tuning this.

With other filesystems, I have seen throughput increase all the way through 16M block sizes. I'm curious, did the ZFS team choose this intentionally and optimize around 1M blocks?

 

[sretalla]

I'm not sure if @Allan Jude has any comment on that, but I imagine it may have something to do with optimising the parity calculation.