As a follow up, our MDS was locking up under load so I went ahead and tried
it. It seemed that some directories were getting bounced around the MDS
servers and load would transition from one to the other. Initially my guess
was that some of these old clients were sending all requests to one MDS
server, but the performance wasn't stabilizing or getting better, so I went
back to one. We had a number of clients that stopped being able to work in
some directories, but others were fine. We had to reboot all the nodes that
were affected in this manner. We are going to move low performance clients
to FUSE and working to upgrade the kernel on the rest. Until then, we will
just have to fight with these clients that won't release their caps.
I've set the cache memory to 200 GB and the reservation to 50% and that
seems to help a bit. Any other tricks that might help these stubborn
clients release their caps?
Thanks,
Robert LeBlanc
----------------
Robert LeBlanc
PGP Fingerprint 79A2 9CA4 6CC4 45DD A904 C70E E654 3BB2 FA62 B9F1
On Mon, May 4, 2020 at 7:51 PM Gregory Farnum <gfarnum(a)redhat.com> wrote:
On Sat, May 2, 2020 at 3:12 PM Robert LeBlanc
<robert(a)leblancnet.us>
wrote:
If there was a network blip and a client was having trouble
reconnecting, do you
think reducing the ranks to 1 would allow them to
connect? At which point the ranks could be increased again.
Or is it a matter of the client kernel panicking so any kind of
reconnection
won't work?
It's client issues. It shouldn't kernel panic, but it also won't
connect with the existing mount, and IIRC you can't do anything on the
server side to make it do so once it gets stuck. You'd have to reboot
the server or possibly do a umount -l -f and then mount again.
-Greg
Thanks
----------------
Robert LeBlanc
PGP Fingerprint 79A2 9CA4 6CC4 45DD A904 C70E E654 3BB2 FA62 B9F1
On Fri, May 1, 2020 at 7:37 PM Robert LeBlanc <robert(a)leblancnet.us>
wrote:
>
> Thanks guys. We are so close to the edge that we may just take that
chance,
usually the only reason an active client has to reconnect is
because we have to bounce the MDS because it's overwhelmed.
> ----------------
> Robert LeBlanc
> PGP Fingerprint 79A2 9CA4 6CC4 45DD A904 C70E E654 3BB2 FA62 B9F1
>
>
> On Fri, May 1, 2020 at 4:00 AM Paul Emmerich <paul.emmerich(a)croit.io>
wrote:
>>
>> I've seen issues with clients reconnects on older kernels, yeah. They
sometimes get stuck after a network failure
>>
>> Paul
>>
>> --
>> Paul Emmerich
>>
>> Looking for help with your Ceph cluster? Contact us at
https://croit.io
>>
>> croit GmbH
>> Freseniusstr. 31h
>> 81247 München
>>
www.croit.io
>> Tel: +49 89 1896585 90
>>
>>
>> On Thu, Apr 30, 2020 at 10:19 PM Gregory Farnum <gfarnum(a)redhat.com>
wrote:
>>>
>>> On Tue, Apr 28, 2020 at 11:52 AM Robert LeBlanc <robert(a)leblancnet.us>
wrote:
>>> >
>>> > In the Nautilus manual it recommends >= 4.14 kernel for multiple
active
>>> > MDSes. What are the potential
issues for running the 4.4 kernel with
>>> > multiple MDSes? We are in the process of upgrading the clients, but
at
>>> > times overrun the capacity of a
single MDS server.
>>>
>>> I don't think this is documented specifically; you'd have to go
>>> through the git logs. Talked with the team and 4.14 was the upstream
>>> kernel when we marked multi-MDS as stable, with the general stream of
>>> ongoing fixes that always applies there.
>>>
>>> There aren't any known issues that will cause file consistency to
>>> break or anything; I'd be more worried about clients having issues
>>> reconnecting when their network blips or an MDS fails over.
>>> -Greg
>>>
>>> >
>>> > MULTIPLE ACTIVE METADATA SERVERS
>>> > <
https://docs.ceph.com/docs/nautilus/cephfs/kernel-features/#multiple-active…
>>> >
>>> > The feature has been supported since the Luminous release. It is
>>> > recommended to use Linux kernel clients >= 4.14 when there are
multiple
>>> > active MDS.
>>> > Thank you,
>>> > Robert LeBlanc
>>> > ----------------
>>> > Robert LeBlanc
>>> > PGP Fingerprint 79A2 9CA4 6CC4 45DD A904 C70E E654 3BB2 FA62 B9F1
>>> > _______________________________________________
>>> > ceph-users mailing list -- ceph-users(a)ceph.io
>>> > To unsubscribe send an email to ceph-users-leave(a)ceph.io
>>> >
>>> _______________________________________________
>>> ceph-users mailing list -- ceph-users(a)ceph.io
>>> To unsubscribe send an email to ceph-users-leave(a)ceph.io