On Tue, Apr 7, 2020 at 8:16 PM Strahil Nikolov <hunter86_bg@yahoo.com> wrote:

Hi Gianluca,

The positive thing is that you can reproduce the issue.

I would ask you to check your gluster version and if there are any updates - update the cluster.

I'd prefer to stick on oVirt release version of Gluster if possible

Also check the gluster's op-version, as this limits some of the features.

What do you mean by thgis?

If there are none - enable trace logs in gluster ( https://access.redhat.com/documentation/en-us/red_hat_gluster_storage/3.1/html/administration_guide/configuring_the_log_level ), start volume peofiling, reproduce the issue and then reduce the log level (it's generating a lot of logs) and stop the profiling.

Once that info is collected, some of the Gluster members can check the situation.

Best Regards,
Strahil Nikolov

Ok. I think that the INFO level set on the different layers outlined problem somehow related to shardin.

Related to this, I found no official docs on Gluster web site after 3.7 version... where is it?

Only information I found was in Red Hat Gluster Storage 3.5 Administration Guide, but I would expect something more upstream...

In particular in my case where I have only one host and the gluster volumes are single brick based, do you think I can try to disable sharding and verify if using new disks with it disabled and oVirt thin provisioned disks let the problem go away?

Also, I found some information about sharding block size.

Apparently the only supported size on Red Hat Gluster Storage is 512MB, but oVirt sets it to 64MB....?

I also found a bugzilla about passing from 128MB to 64MB in oVirt 4.1.5:

https://bugzilla.redhat.com/show_bug.cgi?id=1469436

Now I see that by default and so also in my environment I have:

features.shard on
features.shard-block-size 64MB
features.shard-lru-limit 16384
features.shard-deletion-rate 100

Not clear how to cross-check for all the values above in docs....

Can I safely set

features.shard off

In Red Hat Gluster storage admin guide it is considered only for replicated gluster volumes (and so also ditributed-replicated...)

But in my case with single host and single beick for volumes I think it doesn't give any particular benefit, isn't it?

I found this interesting post about sharding and connection between image files and shard files:

http://opensource-storage.blogspot.com/2016/07/de-mystifying-gluster-shards.html

Gianluca