zfs meta data slowness

[mike tancsa]

On 7/22/2020 9:26 AM, Eugene Grosbein wrote:
>
> It's hard to read due to wrapping plus, it is truncated.
> Maybe additional flag -d3 or similar will help, combined with dedirection < /dev/null > top.out
> so it won't use size of your terminal to wrap/truncate output.
>
> Also, make sure you invoke top while "zfs" command is running.
> Also, procstat -kk for pid of "zfs" command would be useful (but may occur pretty long).
>
> I suppose it blocks waiting for some kernel lock and procstat would show details.

.txt file is attached. I also ran procstat -kk <pid> in a loop for a bit
while it was running.

In this case the list too just under 5min to run.

time /sbin/zfs list -Hp -oname,creation -S creation -t snapshot >
/var/log/snapshots.out
2.311u 24.963s 4:55.13 9.2%     71+179k 2706576+26io 0pf+0w

Ideally, I would like to bias slightly ARC to hold this meta data a
little longer than evict it out of cache so fast.  I put in an nvme disk
hoping that would help, but it makes no difference or doesnt seem to
anyways.

    ---Mike