##Discussed:##

  • Running the VM compute nodes for HPC cluster
  • Short term: running on engage1 OpenStack (quicker, homogeneous tests)
    Long term: also running on Kaizen OpenStack (politically looks good, cross campus lines connectivity and resource sharing)
  • Avoids the short term issues of the MeetMe switch, but there are still issues bridging networks within MIT.
  • We should be able to do both, and they should work the same way (up to network cfg issues).
  • Make sure we get an account on Liberty OpenStack to verify the configuration to have a similar setup. (This is not a substitute!)

  • Path to project goal: Make one of these VMs visible (by some route) to the main engage1 cluster and federate with existing slurm master to receive jobs.
    Idea:

    • slurm.conf can just include an arbitrary public IP for a virtual node so that it’s known and visible to the slurm master for the job assignment.

    • Can set up a private LDAP mirror for VM for security purposes.

      • one way mirror only for verification and no modification—can’t mess up mirror
      • This syncs up users frm the master

    • OSG route:

      • Current implementation:

        • proxy host user ID in engage1.
          • Simplifies user id story if we only target OSG first—only one user ID.
          • Takes work from Open Science workflow cloud, runs it on SLURM cluster.
        • runs when there are idle cycles, can get killed.
        • Uses virtual file system.

      • We could have it be the only user, work 24/7.

      • Issue: OSG runs could get killed every 15 minutes, if they aren’t checkpointing, they restart and make no progress.

        Soln: Runs in virtual machine, can checkpoint before they get killed.

      • OSG services can be suspended depending on if other work comes along. (Gives a way to use idle cycles)
        Remember: Cloud, how do we support bursty needs, vs. steady scientific computing.

      • May want to mirror their CVMFS world - gets us out of LDAP world, but there are security concerns about giving access to CVMFS.

  • Initial requirements for OSG route:

    • Need to get OSG account.
    • CVMFS - namespace in the OSG, based on HTTP perhaps OR We could set up our own server to host our own stuff
    • End to end story: we’re going to use OSG to support these end-to-end workflows, similar to what other people may do.
    • One user id in our SLURM piece: OSG proxy.

##To-do##

  • Get salt-cloud work on github

    Update:

    • config files and recipie for salt-cloud and salt updated on github
    • updated wiki with the steps for use

  • Take cloud-stack stuff, be able to instantiate a flexible number of VMs. This is a reference set up anyone can “checkout and run.”

    Update:

    • salt-cloud map config file can be used to specify the number of VMs to be provisioned with their specifc configuration profiles.
    • salt-cloud and salt-master installation and config files from git can be used to setup a salt VM cluster.