by rottegift » Thu Dec 10, 2015 1:36 am
The logbias property is mainly felt in zfs_log_write() where it's used to decide whether to write to a separated log vdev or not. If you haven't configured a slog, setting logbias either way is going to do very little.
If you have configured a slog, then sync writes to a dataset with logbias=throughput will NOT write to the slog. This frees up the slog for other datasets' sync writes. It also means that applications trying to do a sync write to that dataset will not see their write (or fsync or whatever) call return as promptly.
There are cases where a slog will help an SSD-based pool, but those typically involve workloads that are highly synchronous (many NFS writers is the typical case) and a slog vdev whose latency is even lower than that of the SSDs (e.g. a battery-backed DRAM drive).