Re: Speed improvements in ZFS

On 9/4/23, Alexander Leidinger <Alexander@leidinger.net> wrote:
> Am 2023-08-28 22:33, schrieb Alexander Leidinger:
>> Am 2023-08-22 18:59, schrieb Mateusz Guzik:
>>> On 8/22/23, Alexander Leidinger <Alexander@leidinger.net> wrote:
>>>> Am 2023-08-21 10:53, schrieb Konstantin Belousov:
>>>>> On Mon, Aug 21, 2023 at 08:19:28AM +0200, Alexander Leidinger wrote:
>>>>>> Am 2023-08-20 23:17, schrieb Konstantin Belousov:
>>>>>> > On Sun, Aug 20, 2023 at 11:07:08PM +0200, Mateusz Guzik wrote:
>>>>>> > > On 8/20/23, Alexander Leidinger <Alexander@leidinger.net> wrote:
>>>>>> > > > Am 2023-08-20 22:02, schrieb Mateusz Guzik:
>>>>>> > > >> On 8/20/23, Alexander Leidinger <Alexander@leidinger.net>
>>>>>> > > >> wrote:
>>>>>> > > >>> Am 2023-08-20 19:10, schrieb Mateusz Guzik:
>>>>>> > > >>>> On 8/18/23, Alexander Leidinger <Alexander@leidinger.net>
>>>>>> > > >>>> wrote:
>>>>>> > > >>>
>>>>>> > > >>>>> I have a 51MB text file, compressed to about 1MB. Are you
>>>>>> > > >>>>> interested
>>>>>> > > >>>>> to
>>>>>> > > >>>>> get it?
>>>>>> > > >>>>>
>>>>>> > > >>>>
>>>>>> > > >>>> Your problem is not the vnode limit, but nullfs.
>>>>>> > > >>>>
>>>>>> > > >>>> https://people.freebsd.org/~mjg/netchild-periodic-find.svg
>>>>>> > > >>>
>>>>>> > > >>> 122 nullfs mounts on this system. And every jail I setup has
>>>>>> > > >>> several
>>>>>> > > >>> null mounts. One basesystem mounted into every jail, and then
>>>>>> > > >>> shared
>>>>>> > > >>> ports (packages/distfiles/ccache) across all of them.
>>>>>> > > >>>
>>>>>> > > >>>> First, some of the contention is notorious VI_LOCK in order
>>>>>> > > >>>> to
>>>>>> > > >>>> do
>>>>>> > > >>>> anything.
>>>>>> > > >>>>
>>>>>> > > >>>> But more importantly the mind-boggling off-cpu time comes
>>>>>> > > >>>> from
>>>>>> > > >>>> exclusive locking which should not be there to begin with --
>>>>>> > > >>>> as
>>>>>> > > >>>> in
>>>>>> > > >>>> that xlock in stat should be a slock.
>>>>>> > > >>>>
>>>>>> > > >>>> Maybe I'm going to look into it later.
>>>>>> > > >>>
>>>>>> > > >>> That would be fantastic.
>>>>>> > > >>>
>>>>>> > > >>
>>>>>> > > >> I did a quick test, things are shared locked as expected.
>>>>>> > > >>
>>>>>> > > >> However, I found the following:
>>>>>> > > >>         if ((xmp->nullm_flags & NULLM_CACHE) != 0) {
>>>>>> > > >>                 mp->mnt_kern_flag |=
>>>>>> > > >> lowerrootvp->v_mount->mnt_kern_flag &
>>>>>> > > >>                     (MNTK_SHARED_WRITES | MNTK_LOOKUP_SHARED |
>>>>>> > > >>                     MNTK_EXTENDED_SHARED);
>>>>>> > > >>         }
>>>>>> > > >>
>>>>>> > > >> are you using the "nocache" option? it has a side effect of
>>>>>> > > >> xlocking
>>>>>> > > >
>>>>>> > > > I use noatime, noexec, nosuid, nfsv4acls. I do NOT use nocache.
>>>>>> > > >
>>>>>> > >
>>>>>> > > If you don't have "nocache" on null mounts, then I don't see how
>>>>>> > > this
>>>>>> > > could happen.
>>>>>> >
>>>>>> > There is also MNTK_NULL_NOCACHE on lower fs, which is currently set
>>>>>> > for
>>>>>> > fuse and nfs at least.
>>>>>>
>>>>>> 11 of those 122 nullfs mounts are ZFS datasets which are also NFS
>>>>>> exported.
>>>>>> 6 of those nullfs mounts are also exported via Samba. The NFS
>>>>>> exports
>>>>>> shouldn't be needed anymore, I will remove them.
>>>>> By nfs I meant nfs client, not nfs exports.
>>>>
>>>> No NFS client mounts anywhere on this system. So where is this
>>>> exclusive
>>>> lock coming from then...
>>>> This is a ZFS system. 2 pools: one for the root, one for anything I
>>>> need
>>>> space for. Both pools reside on the same disks. The root pool is a
>>>> 3-way
>>>> mirror, the "space-pool" is a 5-disk raidz2. All jails are on the
>>>> space-pool. The jails are all basejail-style jails.
>>>>
>>>
>>> While I don't see why xlocking happens, you should be able to dtrace
>>> or printf your way into finding out.
>>
>> dtrace looks to me like a faster approach to get to the root than
>> printf... my first naive try is to detect exclusive locks. I'm not 100%
>> sure I got it right, but at least dtrace doesn't complain about it:
>> ---snip---
>> #pragma D option dynvarsize=32m
>>
>> fbt:nullfs:null_lock:entry
>> /args[0]->a_flags & 0x080000 != 0/
>> {
>>         stack();
>> }
>> ---snip---
>>
>> In which direction should I look with dtrace if this works in tonights
>> run of periodic? I don't have enough knowledge about VFS to come up
>> with some immediate ideas.
>
> After your sysctl fix for maxvnodes I increased the amount of vnodes 10
> times compared to the initial report. This has increased the speed of
> the operation, the find runs in all those jails finished today after ~5h
> (@~8am) instead of in the afternoon as before. Could this suggest that
> in parallel some null_reclaim() is running which does the exclusive
> locks and slows down the entire operation?
>

That may be a slowdown to some extent, but the primary problem is
exclusive vnode locking for stat lookup, which should not be
happening.

-- 
Mateusz Guzik <mjguzik gmail.com>