Learning Center
Plans & pricing Sign in
Sign Out
Get this document free



Cluster Configuration Challenges
             Software on Nodes
 User can define “Package Sets” *
 User applies a package set to a group of
 Basically works today with one to one
  associations between package sets and
  node groups
 All collected package set data is stored in
* Package Set: a saved set of OSCAR packages
     Configurations on Nodes
 Works today in the one to one
  configuration to package set relationship
 this is transitively one to one configuration
  to node group
 All collected configuration data can be
  stored in ODA
                 Node Groups
 Due to package set exclusive package set and
  configuration associations with nodes, exclusive
  node groups are needed
 A non-exclusive grouping capability will certainly
  be necessary in the future
    – C3, other packages
    – Storage nodes, compute nodes, user access points,
    – Capability will be very useful to admins, package
            Arising Problems
1) Having two types of node grouping is confusing
   (exclusive and non-exclusive)
2) Configurations for some packages may want to
   be cluster wide, not limited to the package set
   associated nodes. However, some may need
3) Package Sets and Configurations are LOCKED
   to each other. This means that all nodes with
   a shared configuration must have the SAME
   Package Set applied.
4) $OSCAR_SERVER is configured in saved
   Configurations. What if two Configurations
          Arising Problems… cont.
5)   Support is needed for client/server packages
     that need to use something other than
     –   server package is in rpm format
         a) Install server rpm in post_install
         b) Install server rpm everywhere, enable in post_install (could
            be overkill, especially for diskless nodes)
         c) Make special disk image just to reflect a single rpm
            difference per client/server package (results in many
            images with redundant data, and LOTS more work to
            manage for the sysadmin!)
6)   Node names not available to package’s during
     their configuration steps. Package authors
     need the flexibility this would provide.
    Solutions, and their problems…
1)   Make many SIS images behind the scenes to reflect client/server
     rpm differences, and overlapping package sets
    Really slick! PBS server node could be rebuilt with minimal
     post_install… it has it’s OWN unique image
    Grossly inefficient in time for image creation and disk space
    Having many images significantly reduces advantages gained by
     using images in the first place.
    Solves no Configuration related problems
    Extremely complicated to make a system which would adjust to
     changes in configuration along the way, change all necessary
     images, delete them, merge them, etc.

No good!
    Solutions, and their problems…
2)   Go back to LUI type install (like Kickstart) and install
     nodes by rpm lists
    Solves the “many image” problem for client/server
     packages – node definitions and differences are merely
     “lists”, not images
    Inefficient, no hope for multicast, no multi-distro
     product (LUI is dead)
    Solves no Configuration related problems
    Doesn’t leverage SIS. We’re locked into SIS, and we
     know we want it.

No Good!
    Solutions, and their problems…
3)   Continue on course and save a Configuration per Package Set.
     The package could have LOCAL (node group wide) or GLOBAL
     (cluster-wide) configuration options, saved per Package Set.
    Keeps us on course, not as much backtracking
    Still have to have exclusive and non-exclusive node groups
    Adds much complexity to the configuration for package author
    Configurations still bound to package sets. i.e. Solves problem
     #2, but not #3. (see Arising Problems slides)
    Configuration routine must be gone through by the user N times
     (where N=number of occurrences in package sets)

 But wait… let’s step back once.
Take a simplistic view of what the user actually needs to do:
 Define nodes
 Choose software for nodes (or node groups)
 Configure software
Suppose we allow the user to bind a specific package’s saved “Configuration” to node
     directly, instead of via linking to a Package Set, which is bound to nodes… well, that
     would mean that a configuration could exist for a node without that package installed
     for it. That wouldn’t work- could have a conflict.
What if… we didn’t define Package Sets at all and only configured? Application of the
     saved “Configuration” for the package to a node would IMPLY a selection itself. No
     Package Selection needed. No Package Sets needed. Different packages’ saved
     “Configurations” can now overlap different groups of nodes!
Well what about the image then? We can’t have overlapping dynamic configurations and
     have different images. It’s like having our cake and eating it too, essentially. That
     conflict will cost a slight change in our API- if the operations currently done on the
     image were done during post_install instead… we could have a SINGLE BASE IMAGE.
     Currently, this means eliminating just 4 scripts, <300 lines.
(ls packages/*/scripts/post_client_rpm_install, post_rpm_install)
    [Insert more description here]
   Jeremy draws some windows on the
     Solutions, and their problems…
4)      Combine Package Selection and Configuration into the same step- just Configuration. Node
        definition would occur BEFORE configuration.
       Simplifies wizard… what’s a Package Set? Who needs it? User just picks software and
        matches it to nodes; It’s the actual end-of-the-day, bottom-line process without all the junk in
        the middle.
       Non-exclusive groups all around! No exclusive required with only one base image.
       Node names are finally available during the Configuration step.
       Image based maintenance is preserved as an option to users.
       While still gaining the benefit of images, OSCAR is not so dependent on them, and is much
        closer to a state where it can install on pre-existing machines.
       Treating the $OSCAR_SERVER the same as other nodes becomes incredibly easier, allowing us
        to treat it just like other nodes (all I need to do is ssh to it and sync it)
       Add/Remove package is greatly simplified. Because packages would not be using image
        dependent scripts, the use of images is not crippling. A package would be essentially “Added”
        in the post_install anyway!
       Requires an API change… post_rpm_install and post_clients can’t apply. It is the price of
        granularity, and using a single image.
       Post_install is a little heavier. (note the original base install is also lighter)


To top