Notes on the Ask SGI Session - 19th February 2015
            =================================================

Q.  In 2014, DMFUG prepared a "wish-list" of desired features and
    submitted it to SGI.  Are you able to give a status update on
    this please?

A.  Looking at the top 4, all but #3 are likely to be released in
    June as part of ISSP 3.4

    1.  xfsdump scalability for large numbers of inodes

        The directory dumping phase has been improved.  A site with
        650M inodes (no SSD) had dumps change from 3 days to 22 hours.
        There are still a couple of bugs, but release is expected in
        4-8 weeks.

    2.  Control of daemon requests

        The control mechanism is in place.  Work on the UI is
        proceeding; information will be available in XML and JSON
        formats as well as text.  A PostgreSQL interface is planned.

    3.  Automatically prioritize DMF requests by type

        With the mechanisms described above in place, this is easier
        to implement now.

    4.  DMF clients for Windows

        Done.


Q.  When will DMF be available for SLES 12, SLES 11SP4 and RHEL 7?

A.  End of 2015 for DMF servers, and June 2015 for CXFS clients.


Q.  In XVM, does the new auto detection for multipathing and load
    balancing really work?  How can we tell?

A.  Sometimes.  It works, but it can interfere with lower level
    optimizations impacting some access patterns such as sequential
    reads.

    If /etc/failover2.conf exists, it overrides the automatic pathing.


Q.  What are the recommended IO performance monitoring tools and
    techniques?

A.  PCP and iostat


Q.  On a largish multi user NFS system. (eg a HPC machine with NFS
    data store, 300+ users).  How can you tell which user is causing
    an IO load? Any suggestions to help diagnose the cause of IO load?
    Usually the user is running a program on many machines (compute
    nodes).

A.  The SGI version of NFS is instrumented to collect statistics
    associated with a client hosts.  The SMC-IS interface provides a
    GUI to view the NFS client PCP statistics.  It may be possible
    to enhanced the implementation to collect statistics by user.
    Further investigation is required to determine if the enhancements
    are doable.

    A similar feature for SMB would require further investigation.


Q.  Windows users using Samba on DMF. Recalls of off-line files are
    very inefficient (one at a time). Users also don't have (or know)
    SSH access, so can't use dmget commands.  Are there any recommended
    practices to help out?

    eg. I have set SAM-QFS do prefetching of files. Or recalling all
    the files in a directory, when a couple of them are accessed.

A:  Mostly obsoleted by a presentation from SGI the previous day:
    http://hpc.csiro.au/users/dmfug/Meeting_Feb2015/Presentations/SGI_JN_DMF_on_Windows.pdf

    Unresolved areas include the CLI interface, and how wildcards
    should be handled.  On Windows, wildcards are interpreted by the
    program itself, and not by the shell which invokes it.


Q.  Does/will CXFS support VMs as clients?

A.  CXFS clients work in a VM environment (tested with VMware) but are
    neither officially supported nor unsupported.


					    Last updated 24th February 2015