Dear Vlad,
I'm sorry, I don't want to test this again on my system just yet! It caused too much instability for my users and I don't have enough resources for a development environment. The only other variables that changed before the crashes was the group metadata-cache[0], which I enabled the same day as the parallel-readdir and readdir-ahead options:
$ gluster volume set homes group metadata-cache
I'm hoping Atin or Poornima can shed some light and squash this bug.
Regards,
On Fri, Jan 26, 2018 at 6:10 AM Vlad Kopylov <vladkopy@xxxxxxxxx> wrote:
can you please test parallel-readdir or readdir-ahead gives
disconnects? so we know which to disable
parallel-readdir doing magic ran on pdf from last year
https://events.static.linuxfound.org/sites/events/files/slides/Gluster_DirPerf_Vault2017_0.pdf
-v
On Thu, Jan 25, 2018 at 8:20 AM, Alan Orth <alan.orth@xxxxxxxxx> wrote:
> By the way, on a slightly related note, I'm pretty sure either
> parallel-readdir or readdir-ahead has a regression in GlusterFS 3.12.x. We
> are running CentOS 7 with kernel-3.10.0-693.11.6.el7.x86_6.
>
> I updated my servers and clients to 3.12.4 and enabled these two options
> after reading about them in the 3.10.0 and 3.11.0 release notes. In the days
> after enabling these two options all of my clients kept getting disconnected
> from the volume. The error upon attempting to list a directory or read a
> file was "Transport endpoint is not connected", after which I would force
> unmount the volume with `umount -fl /home` and remount it, only to have it
> get disconnected again a few hours later.
>
> Every time the volume disconnected I looked in the client mount log and only
> found information such as:
>
> [2018-01-24 05:52:27.695225] I [MSGID: 108026]
> [afr-self-heal-common.c:1656:afr_log_selfheal] 2-homes-replicate-1:
> Completed metadata selfheal on ed3fbafc-734b-41ca-ab30-216399fb9168.
> sources=[0] sinks=1
> [2018-01-24 05:52:27.700611] I [MSGID: 108026]
> [afr-self-heal-metadata.c:52:__afr_selfheal_metadata_do]
> 2-homes-replicate-1: performing metadata selfheal on
> b6a53629-a831-4ee3-a35e-f47c04297aaa
> [2018-01-24 05:52:27.703021] I [MSGID: 108026]
> [afr-self-heal-common.c:1656:afr_log_selfheal] 2-homes-replicate-1:
> Completed metadata selfheal on b6a53629-a831-4ee3-a35e-f47c04297aaa.
> sources=[0] sinks=1
>
> I enabled debug logging for that volume's client mount with `gluster volume
> set homes diagnostics.client-log-level DEBUG` and then I saw this in the
> client mount log the next time it disconnected:
>
> [2018-01-24 08:55:19.138810] D [MSGID: 0] [io-threads.c:358:iot_schedule]
> 0-homes-io-threads: LOOKUP scheduled as fast fop
> [2018-01-24 08:55:19.138849] D [MSGID: 0] [dht-common.c:2711:dht_lookup]
> 0-homes-dht: Calling fresh lookup for
> /vchebii/revtrans/Hircus-XM_018067032.1.pep.align.fas on
> homes-readdir-ahead-1
> [2018-01-24 08:55:19.138928] D [MSGID: 0] [io-threads.c:358:iot_schedule]
> 0-homes-io-threads: FSTAT scheduled as fast fop
> [2018-01-24 08:55:19.138958] D [MSGID: 0] [afr-read-txn.c:220:afr_read_txn]
> 0-homes-replicate-1: e6ee0427-b17d-4464-a738-e8ea70d77d95: generation now vs
> cached: 2, 2
> [2018-01-24 08:55:19.139187] D [MSGID: 0] [dht-common.c:2294:dht_lookup_cbk]
> 0-homes-dht: fresh_lookup returned for
> /vchebii/revtrans/Hircus-XM_018067032.1.pep.align.fas with op_ret 0
> [2018-01-24 08:55:19.139200] D [MSGID: 0]
> [dht-layout.c:873:dht_layout_preset] 0-homes-dht: file =
> 00000000-0000-0000-0000-000000000000, subvol = homes-readdir-ahead-1
> [2018-01-24 08:55:19.139257] D [MSGID: 0] [io-threads.c:358:iot_schedule]
> 0-homes-io-threads: READDIRP scheduled as fast fop
>
> On a hunch I disabled both parallel-readdir and readdir-ahead, which I had
> only enabled a few days before, and now all of the clients are much more
> stable, with zero disconnections in the days since I disabled those two
> volume options.
>
> Please take a look! Thanks,
>
> On Wed, Jan 24, 2018 at 5:59 AM Atin Mukherjee <amukherj@xxxxxxxxxx> wrote:
>>
>> Adding Poornima to take a look at it and comment.
>>
>> On Tue, Jan 23, 2018 at 10:39 PM, Alan Orth <alan.orth@xxxxxxxxx> wrote:
>>>
>>> Hello,
>>>
>>> I saw that parallel-readdir was an experimental feature in GlusterFS
>>> version 3.10.0, became stable in version 3.11.0, and is now recommended for
>>> small file workloads in the Red Hat Gluster Storage Server documentation[2].
>>> I've successfully enabled this on one of my volumes but I notice the
>>> following in the client mount log:
>>>
>>> [2018-01-23 10:24:24.048055] W [MSGID: 101174]
>>> [graph.c:363:_log_if_unknown_option] 0-homes-readdir-ahead-1: option
>>> 'parallel-readdir' is not recognized
>>> [2018-01-23 10:24:24.048072] W [MSGID: 101174]
>>> [graph.c:363:_log_if_unknown_option] 0-homes-readdir-ahead-0: option
>>> 'parallel-readdir' is not recognized
>>>
>>> The GlusterFS version on the client and server is 3.12.4. What is going
>>> on?
>>>
>>> [0]
>>> https://github.com/gluster/glusterfs/blob/release-3.10/doc/release-notes/3.10.0.md
>>> [1]
>>> https://github.com/gluster/glusterfs/blob/release-3.11/doc/release-notes/3.11.0.md
>>> [2]
>>> https://access.redhat.com/documentation/en-us/red_hat_gluster_storage/3.3/html/administration_guide/small_file_performance_enhancements
>>>
>>> Thank you,
>>>
>>>
>>> --
>>>
>>> Alan Orth
>>> alan.orth@xxxxxxxxx
>>> https://picturingjordan.com
>>> https://englishbulgaria.net
>>> https://mjanja.ch
>>>
>>>
>>> _______________________________________________
>>> Gluster-users mailing list
>>> Gluster-users@xxxxxxxxxxx
>>> http://lists.gluster.org/mailman/listinfo/gluster-users
>>
>>
> --
>
> Alan Orth
> alan.orth@xxxxxxxxx
> https://picturingjordan.com
> https://englishbulgaria.net
> https://mjanja.ch
>
>
> _______________________________________________
> Gluster-users mailing list
> Gluster-users@xxxxxxxxxxx
> http://lists.gluster.org/mailman/listinfo/gluster-users
Alan Orth
alan.orth@xxxxxxxxx
https://picturingjordan.com
https://englishbulgaria.net
https://mjanja.ch
_______________________________________________ Gluster-users mailing list Gluster-users@xxxxxxxxxxx http://lists.gluster.org/mailman/listinfo/gluster-users