Specific to Linux, the NFS client uses standard filesystem caching
which has a few pros and cons of it's own.
Native GlusterFS uses up application space RAM and is a hard-set
number that you must define. In our studio, our standard rollout is a
32GB RAM workstation, so native GlusterFS clients are told to use
quite a bit of RAM for cache. When we connect up smaller VMs, often
they have less RAM overall than we assign just for cache, and as such
use NFS instead.
-Dan
----------------
Dan Mons - R&D Sysadmin
Cutting Edge
http://cuttingedge.com.au
On 29 April 2015 at 20:01, Kingsley <gluster at gluster.dogwind.com>
wrote:> On Tue, 2015-04-28 at 17:09 -0700, Dave Warren wrote:
>> Bandwidth is also a consideration, the FUSE client will upload multiple
>> copies based on the replica setting for the volume, so if the client is
>> connected at 100Mb/s or over wifi, and the servers are cross-connected
>> on a 10Gb/s backplane, having the client upload multiple copies vs
>> having the NFS server handle the replicas may have an impact on very
>> large files.
>>
>> Finally, NFS seems to have a lighter CPU footprint on the client, at
the
>> possible cost of higher server CPU load, although this is anecdotal
>> (from my own experience), and probably a mixed bag.
>
> Oh I think I get it - the Gluster server daemons are also NFS server
> daemons, so mounting via NFS you're still talking to a Gluster daemon
on
> the server ... and when writing to the server, the servers then handle
> the replication themselves?
>
> I thought people were just putting the path to a brick in /etc/exports
> and then just using nfsd that came with the OS, which I presume would
> then break things.
>
> In that case, would this be a reasonable summary of mounting via NFS
> instead of using the native fuse client?
>
> NFS pros:
> * faster
> * lighter weight on client resources (CPU, bandwidth)
> * available to clients that have a standard NFS client but cannot
> install gluster-fuse
>
> NFS cons:
> * more load on servers
> * writes complete on the client before the data is replicated
> (which is faster, but less secure)
> * if the server that the client happens to be connected to goes
> down, the client loses access to the volume (whereas the fuse
> client recovers and continues writing to the remaining notes)
>
> Does that seem about right?
>
> --
> Cheers,
> Kingsley.
>
> _______________________________________________
> Gluster-users mailing list
> Gluster-users at gluster.org
> http://www.gluster.org/mailman/listinfo/gluster-users