As discussed here: http://comments.gmane.org/gmane.os.illumos.zfs/4085, I would like to propose a fourth option to zfs sync settings (along with standard, always and disabled).
The new option could be named as "delayed", or "throughput", and would fit between standard and disabled.
It would work as follows: when you issue an fsync call, it wouldn't trigger a ZIL write or a txg commit, instead the call would wait until the data gets written onto stable storage with the normal txg flush (happening at, or near the txg timeout setting).
This setting would help applications which need throughput and safety, but don't need low latency, on hardware, which don't have a fast SLOG device or high-IOPS VDEVs (like a mirror of two SATA disks).
It's a middle ground between low latency to the application (ie current fsync()) and complete indifference (using the words of Chris Siebenmann from the above discussion).
I've got positive feedback on the idea, that's why I open this issue. Hopefully somebody can pick this up sometime.
No data to display