Topic
  • 4 replies
  • Latest Post - ‏2009-02-24T06:57:55Z by Theeraph
Theeraph
Theeraph
110 Posts

Pinned topic Linux (for multimedia task) - a member of GPFS cluster vs NFS mount?

‏2009-01-30T09:16:54Z |
Hi,

We will propose a few JS22 blade AIX servers as GPFS NSD server nodes (with DS4800 disk) for a customer running multimedia work...

Their workstations are Linux (about 30-40 in number). Should we include them as GPFS node (non-NSD server) or should we just NFS mount the file system from GPFS cluster?

If we include it in GPFS cluster:
  • Need to pay GPFS license
+ Automatic workload balance (if our NSD servers are balanced)
+ No recovery action needed if NSD server fails (next NSD server in the list will takeover)

If we use NFS mounted:
+ No GPFS license
  • Need to do manual load balance (client group A mount from node X, client group B mount from node Y, ...)
  • Need to do manual mount/unmount if NSD server fails

How about the performance(receive data directly from NSD server via GPFS interconnect vs NFS mount)?

How many % do we get from NFS mount vs included it in the cluster? (eg. 70% performance?)

Also, are there other things that we may need to consider?

Thank you very much,
Theeraphong
Updated on 2009-02-24T06:57:55Z at 2009-02-24T06:57:55Z by Theeraph
  • tomerperry
    tomerperry
    18 Posts

    Re: Linux (for multimedia task) - a member of GPFS cluster vs NFS mount?

    ‏2009-01-30T13:32:22Z  
    > Theeraph wrote:
    > Hi,
    >
    > We will propose a few JS22 blade AIX servers as GPFS NSD server nodes (with DS4800 disk) for a customer running multimedia work...
    >
    > Their workstations are Linux (about 30-40 in number). Should we include them as GPFS node (non-NSD server) or should we just NFS mount the file system from GPFS cluster?
    >
    > If we include it in GPFS cluster:
    > - Need to pay GPFS license
    > + Automatic workload balance (if our NSD servers are balanced)
    > + No recovery action needed if NSD server fails (next NSD server in the list will takeover)
    >
    > If we use NFS mounted:
    > + No GPFS license
    > - Need to do manual load balance (client group A mount from node X, client group B mount from node Y, ...)

    Actually you can use round robin DNS for that purpose that will load balance between the nodes.

    > - Need to do manual mount/unmount if NSD server fails

    Well, if you'll use Linux, you'll be able to use CNFS that will take care of failover etc. for you, see:
    http://www.redbooks.ibm.com/abstracts/redp4400.html
    and GPFS documentation.

    >
    > How about the performance(receive data directly from NSD server via GPFS interconnect vs NFS mount)?
    >
    > How many % do we get from NFS mount vs included it in the cluster? (eg. 70% performance?)
    >
    > Also, are there other things that we may need to consider?
    >

    Well, the answer varies a lot wrt the access pattern of the clients. NSD model will give you larger block size ( in compare to the 32k of current Linux NFS) etc. -

    > Thank you very much,
    > Theeraphong
  • Theeraph
    Theeraph
    110 Posts

    Re: Linux (for multimedia task) - a member of GPFS cluster vs NFS mount?

    ‏2009-02-04T10:03:46Z  
    > Theeraph wrote:
    > Hi,
    >
    > We will propose a few JS22 blade AIX servers as GPFS NSD server nodes (with DS4800 disk) for a customer running multimedia work...
    >
    > Their workstations are Linux (about 30-40 in number). Should we include them as GPFS node (non-NSD server) or should we just NFS mount the file system from GPFS cluster?
    >
    > If we include it in GPFS cluster:
    > - Need to pay GPFS license
    > + Automatic workload balance (if our NSD servers are balanced)
    > + No recovery action needed if NSD server fails (next NSD server in the list will takeover)
    >
    > If we use NFS mounted:
    > + No GPFS license
    > - Need to do manual load balance (client group A mount from node X, client group B mount from node Y, ...)

    Actually you can use round robin DNS for that purpose that will load balance between the nodes.

    > - Need to do manual mount/unmount if NSD server fails

    Well, if you'll use Linux, you'll be able to use CNFS that will take care of failover etc. for you, see:
    http://www.redbooks.ibm.com/abstracts/redp4400.html
    and GPFS documentation.

    >
    > How about the performance(receive data directly from NSD server via GPFS interconnect vs NFS mount)?
    >
    > How many % do we get from NFS mount vs included it in the cluster? (eg. 70% performance?)
    >
    > Also, are there other things that we may need to consider?
    >

    Well, the answer varies a lot wrt the access pattern of the clients. NSD model will give you larger block size ( in compare to the 32k of current Linux NFS) etc. -

    > Thank you very much,
    > Theeraphong
    Hi,

    ! Actually you can use round robin DNS for that purpose that will load balance between the nodes.

    Did you mean DNS on Linux? (I learnt from SOFS class that Active Directory can do this...)

    ! Well, if you'll use Linux, you'll be able to use CNFS that will take care of failover etc.

    Too bad, we will be using GPFS on AIX (since we are very comfortable with AIX...)

    Thank you very much,
    Theeraphong
  • HajoEhlers
    HajoEhlers
    253 Posts

    Re: Linux (for multimedia task) - a member of GPFS cluster vs NFS mount?

    ‏2009-02-11T20:33:44Z  
    • Theeraph
    • ‏2009-02-04T10:03:46Z
    Hi,

    ! Actually you can use round robin DNS for that purpose that will load balance between the nodes.

    Did you mean DNS on Linux? (I learnt from SOFS class that Active Directory can do this...)

    ! Well, if you'll use Linux, you'll be able to use CNFS that will take care of failover etc.

    Too bad, we will be using GPFS on AIX (since we are very comfortable with AIX...)

    Thank you very much,
    Theeraphong
    Hi Theeraphong,
    i ran in the same issue - no cNFS for AIX.

    Currently i thinking about using Linux NSD clients as Gateway servers. ( GPFS <-> cNFS ) where the LX Server will using 10GB ethernet.
    This solution has also the benefit of
    - A high available GPFS core cluster on AIX and easily to scale
    - A high available GPFS Gateway (cNFS) cluster on LX and easily to scale as well.

    So in case something is wrong with the LX gateway servers you do not have to worry about the GPFS core cluster. and v.v.

    hth
    Hajo
  • Theeraph
    Theeraph
    110 Posts

    Re: Linux (for multimedia task) - a member of GPFS cluster vs NFS mount?

    ‏2009-02-24T06:57:55Z  
    Hi Theeraphong,
    i ran in the same issue - no cNFS for AIX.

    Currently i thinking about using Linux NSD clients as Gateway servers. ( GPFS <-> cNFS ) where the LX Server will using 10GB ethernet.
    This solution has also the benefit of
    - A high available GPFS core cluster on AIX and easily to scale
    - A high available GPFS Gateway (cNFS) cluster on LX and easily to scale as well.

    So in case something is wrong with the LX gateway servers you do not have to worry about the GPFS core cluster. and v.v.

    hth
    Hajo
    Hi,

    Thank you for your advices...

    However, we cannot do it here since the customer only want homogeneous environment...

    Anyway, I am interested in your ideas of doing this. Please share your experiences after you implemented it. (My email is theeraph@th.ibm.com...)

    Thank you,
    Theeraphong