Sequential IO on ZFS exhibits factor >13 slowdown when not aligned to record boundaries.
It seems that using dd to write a large sequential file with bs=131072 is quick while bs=65536 exhibits a factor of >13 slowdown. This has been observed on both Linux and Solaris 10u10. This appears to be a write-hole caused by an absence of write coalescing. Full details are available at the ZFSOnLinux issue:
This is not something that the ZFSOnLinux project should tackle by itself, so I decided to file an upstream issue.