Hi Prasun,
pNFS was recently released in a "tech-preview" form. With multiple
MDS-es or even an all-symmetric arch (every ganesha node can act as both
DS and MDS, which will also be a supported config) you could potentially
see improvements (due to increased throughput) but, from our experiments
so far (and this is certainly work in progress) there is much improved
perf for large files and not so much for small file i/o (using
distributed iozone workloads). That said, we have not done a lot of work
playing around with the tuneables as yet (glusterfs side options and
things like vm-background ratio and vm-dirty ratio tuning on the OS
side) for small file pNFS access, so it is pending exercise at this
point. We can confirm only when these performance experiments are
completed. But if you happen to have large files as well, it should be a
win. And the best part is, it comes at zero extra cost to try it out.
If you have the bandwidth, you should probably give pnfs a spin with the
latest glusterfs versions if you want very quick answers. It is quite
easy to setup and we could lend a helping hand there. And maybe you will
pleasantly surprise us ;-) :)
Anand
On 08/13/2015 12:18 PM, Prasun Gera wrote:> Thanks. For small files and random I/O, nfs has been recommended over
> fuse. Would pNFS, with multiple MDS'es in the future, be the
> recommended approach for small files ?
>
> On Wed, Aug 12, 2015 at 11:03 PM, Soumya Koduri <skoduri at redhat.com
> <mailto:skoduri at redhat.com>> wrote:
>
> It depends on the workload. Like native NFS, even with
> NFS-Ganesha, data is routed through the server where its mounted
> from. In addition NFSv4.x protocol adds more complexity and cannot
> be directly compared with NFSv3 traffic. However with pNFS, I/O is
> routed to data servers directly by the NFS clients which results
> in performance gain for larger I/O workloads. Also we do have plan
> to support multiple metadata servers going forward.
>
> Thanks,
> Soumya
>
> ----- Original Message -----
> From: "Prasun Gera" <prasun.gera at gmail.com
> <mailto:prasun.gera at gmail.com>>
> To: "Joe Julian" <joe at julianfamily.org <mailto:joe
at julianfamily.org>>
> Cc: gluster-users at gluster.org <mailto:gluster-users at
gluster.org>
> Sent: Thursday, August 13, 2015 2:55:17 AM
> Subject: Re: [Gluster-users] ganesha BFS
>
> And do either of them perform better than fuse mounts ? With
> native nfs, all data is routed through the server where it's
> mounted from, which makes HA and load balancing difficult. For
> pNFS, there is a single metadata server. How does that affect HA
> and load ? I thought one of the main goals of gluster was
> decentralized metadata. Where do the four options (fuse, native
> nfs, nfsv4, pnfs ) stand in terms of benefits and disadvantages ?
>
>
> On Wed, Aug 12, 2015 at 11:26 AM, Joe Julian <
> joe at julianfamily.org <mailto:joe at julianfamily.org> >
wrote:
>
>
>
> nfs-ganesha is a much more feature rich nfs server that uses
> libgfapi to access the gluster volume in userspace. This userspace
> solution avoids the context switches like the native gluster nfs
> does, but adds support for pnfs/nfsv4 and udp.
>
> From the development standpoint, they have a full set of
> developers working only on and focused only on their nfs server
> whereas the gluster version was implemented as a stop-gap to
> provide a solution where the kernel nfs re-share was failing.
>
> I think nfs-ganesha is a better solution. There is integration
> work being done in glusterfs to make its use seamless, so I
> suspect that's the long-term nfs solution that will eventually
> replace gluster's native nfs.
>
>
> On 08/12/2015 09:54 AM, paf1 at email.cz <mailto:paf1 at
email.cz> wrote:
>
>
>
> Hello Dears,
>
> can anybody explain advanteges / disadvantages of Ganesha NFS ??
> Will U reccomend me go through this way ??
> ( 4 node glusterFS )
> regs.
> Pavel
>
>
>
> _______________________________________________
> Gluster-users mailing list Gluster-users at gluster.org
> <mailto:Gluster-users at gluster.org>
> http://www.gluster.org/mailman/listinfo/gluster-users
>
>
> _______________________________________________
> Gluster-users mailing list
> Gluster-users at gluster.org <mailto:Gluster-users at
gluster.org>
> http://www.gluster.org/mailman/listinfo/gluster-users
>
>
> _______________________________________________
> Gluster-users mailing list
> Gluster-users at gluster.org <mailto:Gluster-users at
gluster.org>
> http://www.gluster.org/mailman/listinfo/gluster-users
>
>
>
>
> _______________________________________________
> Gluster-users mailing list
> Gluster-users at gluster.org
> http://www.gluster.org/mailman/listinfo/gluster-users
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://www.gluster.org/pipermail/gluster-users/attachments/20150813/6cd1fdd8/attachment.html>