[Linux-cluster] is it possible an active-active NFS server?

[Gordan Bobic]

On 06/28/2010 01:28 PM, ESGLinux wrote:
>
>
> 2010/6/28 Gordan Bobic <gordan at bobich.net <mailto:gordan at bobich.net>>
>
>     On 06/28/2010 12:26 PM, ESGLinux wrote:
>
>
>
>         2010/6/28 Gordan Bobic <gordan at bobich.net
>         <mailto:gordan at bobich.net> <mailto:gordan at bobich.net
>         <mailto:gordan at bobich.net>>>
>
>
>             On 06/28/2010 11:11 AM, ESGLinux wrote:
>
>                 Hi All,
>
>                 I´m going to mount an active-active file server and my first
>                 idea is to
>                 mount a NFS service with luci but now I have the doubt
>         if is it
>                   possible. with luci i have allways mounted Active-Passive
>                 services. So,
>                 my question is that.
>
>                 Any other aproach to get an Active-Active file Server?
>
>
>
>             Not with NFS, since NFS has no feature to have multiple
>             servers/share. But there is no reason you can't connect half
>         of the
>             clients to the other server.
>
>
>         I haven't realized about it, it could be a solution.
>
>         one thing, I have been investigating about it, and  I have
>         thought it
>         could be possible using Linux Virtual Server (administered with
>         piranha), what do you think about it?
>
>
>     I think you need to start to list your requirements in a coherent
>     manner first, in terms of performance, features, and redundancy. The
>     solution you should be looking for will be more obvious then.
>
>
>
> Hi again,
>
> you are right this is a bit confusing (my customer said me: I need a
> cluster file server, its your problem... :-/ ). Now I´m investigating
> how to do it.
>
> what basically I need is:
>
> I need to access the files in HA and it must be scalable. If the load is
> a problem I could be able to add another node to solve the load problem,
> (so I thought in a Active-active, because with Active-Pasive only there
> is one node active so the load problem is still there)

Whether it will scale is dependant almost exclusively on your access 
pattern. If you can group your cluster file system accesses so that 
nodes hardly ever access the same file system subtrees then it will 
scale reasonably well. If you are going to have nodes randomly accessing 
the file system paths, then the performance will take a nosedive, and 
get progressively slower as you add nodes.

This will scale linearly:
Node 1 accessing /my/path/1/whatever
Node 2 accessing /my/path/2/whatever

This will scale inversely (get slower):
Node 1 accessing /my/path
Node 2 accessing /my/path

Cluster file systems are generally slower at random access than 
standalone file systems, so you are likely to find that having a 
standalone failover (active-passive) solution is faster than a clustered 
active-active solution, especially as you add nodes.

So the question really comes down to access patterns. If you are going 
to have random access to lots of small files (e.g. Maildir), the 
performance will be poor to start with and get worse as you add nodes 
unless you can engineer your solution so that access for a particular 
subtree always hits the same node. OTOH for large file operations, the 
bandwidth will be more dominant than random access lock acquisition 
time, so the performance will be OK and scale reasonably as you add nodes.

Note that this isn't something specific to GFS - pretty much all cluster 
file systems behave this way.

Gordan