Some kind of single point will always be there I guess. Because even if we go with the distributed filesystem, it will be mounted to the access VM and this access VM will be providing NFS/CIFS protocol access. So this machine is single point of failure (indeed we would be running two of them for active-passive HA setup. In case of distributed filesystem approach the failure of the access VM would mean re-mounting the filesystem on the passive access VM. In case of "monster VM" approach, in case of the VM failure it would mean reattaching all block volumes to a new VM.
On Mon, Nov 12, 2018 at 3:40 PM Ashley Merrick <singapore@xxxxxxxxxxxxxx> wrote:
My 2 cents would be depends how H/A you need.Going with the monster VM you have a single point of failure and a single point of network congestion.If you go the CephFS route you remove that single point of failure if you mount to clients directly. And also can remove that single point of network congestion.Guess depends on the performance and uptime required , as I’d say that could factory into your decisions.On Mon, 12 Nov 2018 at 10:36 PM, Premysl Kouril <premysl.kouril@xxxxxxxxx> wrote:Hi Kevin,I should have also said, that we are internally inclined towards the "monster VM" approach due to seemingly simpler architecture (data distribution on block layer rather than on file system layer). So my original question is more about comparing the two approaches (distribution on block layer vs distribution on filesystem layer). "Monster VM" approach being the one where we just keep mounting block volumes to a single VM with normal non-distributed filesystem and then exporting via NFS/CIFS.Regards,Prema_______________________________________________On Mon, Nov 12, 2018 at 3:17 PM Kevin Olbrich <ko@xxxxxxx> wrote:Hi Dan,_______________________________________________ZFS without sync would be very much identical to ext2/ext4 without journals or XFS with barriers disabled.The ARC cache in ZFS is awesome but disbaling sync on ZFS is a very high risk (using ext4 with kvm-mode unsafe would be similar I think).Also, ZFS only works as expected with scheduler set to noop as it is optimized to consume whole, non-shared devices.Just my 2 cents ;-)KevinAm Mo., 12. Nov. 2018 um 15:08 Uhr schrieb Dan van der Ster <dan@xxxxxxxxxxxxxx>:We've done ZFS on RBD in a VM, exported via NFS, for a couple years.
It's very stable and if your use-case permits you can set zfs
sync=disabled to get very fast write performance that's tough to beat.
But if you're building something new today and have *only* the NAS
use-case then it would make better sense to try CephFS first and see
if it works for you.
-- Dan
On Mon, Nov 12, 2018 at 3:01 PM Kevin Olbrich <ko@xxxxxxx> wrote:
>
> Hi!
>
> ZFS won't play nice on ceph. Best would be to mount CephFS directly with the ceph-fuse driver on the endpoint.
> If you definitely want to put a storage gateway between the data and the compute nodes, then go with nfs-ganesha which can export CephFS directly without local ("proxy") mount.
>
> I had such a setup with nfs and switched to mount CephFS directly. If using NFS with the same data, you must make sure your HA works well to avoid data corruption.
> With ceph-fuse you directly connect to the cluster, one component less that breaks.
>
> Kevin
>
> Am Mo., 12. Nov. 2018 um 12:44 Uhr schrieb Premysl Kouril <premysl.kouril@xxxxxxxxx>:
>>
>> Hi,
>>
>>
>> We are planning to build NAS solution which will be primarily used via NFS and CIFS and workloads ranging from various archival application to more “real-time processing”. The NAS will not be used as a block storage for virtual machines, so the access really will always be file oriented.
>>
>>
>> We are considering primarily two designs and I’d like to kindly ask for any thoughts, views, insights, experiences.
>>
>>
>> Both designs utilize “distributed storage software at some level”. Both designs would be built from commodity servers and should scale as we grow. Both designs involve virtualization for instantiating "access virtual machines" which will be serving the NFS and CIFS protocol - so in this sense the access layer is decoupled from the data layer itself.
>>
>>
>> First design is based on a distributed filesystem like Gluster or CephFS. We would deploy this software on those commodity servers and mount the resultant filesystem on the “access virtual machines” and they would be serving the mounted filesystem via NFS/CIFS.
>>
>>
>> Second design is based on distributed block storage using CEPH. So we would build distributed block storage on those commodity servers, and then, via virtualization (like OpenStack Cinder) we would allocate the block storage into the access VM. Inside the access VM we would deploy ZFS which would aggregate block storage into a single filesystem. And this filesystem would be served via NFS/CIFS from the very same VM.
>>
>>
>> Any advices and insights highly appreciated
>>
>>
>> Cheers,
>>
>> Prema
>>
>> _______________________________________________
>> ceph-users mailing list
>> ceph-users@xxxxxxxxxxxxxx
>> http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
>
> _______________________________________________
> ceph-users mailing list
> ceph-users@xxxxxxxxxxxxxx
> http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
ceph-users mailing list
ceph-users@xxxxxxxxxxxxxx
http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
ceph-users mailing list
ceph-users@xxxxxxxxxxxxxx
http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
_______________________________________________ ceph-users mailing list ceph-users@xxxxxxxxxxxxxx http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
