Some of our random write benchmarks on a fragmented pool show that
single-threaded portion of sync process (txg_sync_thread) can use
up to 45% of CPU time. Most of it is consumed by metaslab_sync()
and metaslab_sync_done(), during which time the pool is not doing
anything else.
While metaslab_sync() is not trivial to parallelize due to having
single spacemap log, metaslab_sync_done() is doing only per-metaslab
accounting and they can run in parallel. Even better, we can run
them while waiting for vdev label update and cache flush I/Os.
With this patch on my test system similar test randomly writing 12
100GB files with 4KB blocks shows IOPS increase from 176K to 220K.
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <alexander.motin@TrueNAS.com>
Closes#18622
Alexander Motin
5fea0c838a