Comment 33 for bug 1654517

Hello Colin, nope.

It looks like ZFS has some sort of contention in waiting the ZIL to finish the async callbacks from scheduled work, even when more spindles exist. Not sure if its related to slab merging or getting the objects from either slab/kmalloc or directly from page allocs. I think its a locking contention due to zfs scalability iself, to be honest :\. I'm getting, sometimes, soft lockups here and there (all waiting on on zfs sync - txg_wait_synced). I'm using a raid-z with 4x480G SSDs and an extra SSD as L2ARC disk.

Its fairly easy to reproduce, i'm basically compiling kernel tress all day in this machine, using the zfs pools as ccache's cache and as the compiled objects repository. I haven't explored tuning it though, considering its cache is not kept together with the pagecache (like changing zfs_dirty_data_sync or zfs_sync_taskq_batch_pct to make zfs cache flush less async, which might help).

Cheers!