"Oddness" in head since around r343678 or so

[Niclas Zeising]

On 2019-02-10 16:35, Niclas Zeising wrote:
> On 2019-02-08 10:27, Alexander Leidinger wrote:
>> Hi,
>>
>> I recently noticed some generic slowness myself. I experienced this 
>> during replacing disks in a raidz by bigger ones. Long story short, 
>> check top -s if you have vnlru running for a long period at high 
>> CPU... If yes increase kern.maxvnodes (I increased to 10 times). Note, 
>> we should improve the admin page in the FAQ, the vnlru entry could 
>> need a little bit more hints and explanations.
>>
>> If you encounter the same issue we have probably introduced a change 
>> somewhere with an unintended side effect.
>>
>> Bye,
>> Alexander.
>>
> 
> Hi!
> I'm seeing this as well, on 13-CURRENT.  I updated a computer from the 
> last January snapshot (30 or 31 of January, I can't remember) and it 
> seems disk IO is very slow.  I remember having a svn checkout taking a 
> very long time, with the SVN process pegged at 100% according to top.  I 
> can't see the vnlru process running though, but I haven't looked 
> closely, and I haven't tried the maxvnodes workaround.  Something has 
> changed though.
> This is systems using ZFS, both mirror and single disk.  Gstat shows 
> disks are mostly idle.
> 
> I know this is a lousy bug report, but this, and the feeling that things 
> are slower than usual, is what I have for now.
> Regards

Hi!
I did some more digging.  In short, disabling options COVERAGE and 
options KCOV made my test case much faster.

My test:
boot system
create a new zfs dataset (zroot/home/test in my case)
time a checkout of https://svn.freebsd.org/base/head, putting the files 
in the new zfs dataset.

This is in no way scientific, since I only ran the test once on each 
kernel, and using something on the network means I'm susceptible to 
varying network speeds and so on, but.
In this specific scenario, using a kernel without those options, it's 
about 3 times faster than with, at least on the computer where I ran the 
tests.

I noticed in the commit log that the coverage and kcov options has been 
disabled again, albeit for a different reason.  Perhaps they should 
remain off, unless the extra runtime overhead can be disabled in 
runtime, similar to witness.
Regards
-- 
Niclas