Notes from meeting with Yan, Jeff, Ben, Mike, Craig

Discussed concept of using Workbench in conjunction with ROGER system to provide interactive access to data.

The basic idea is to install Workbench in OpenStack to provide an alternative to the traditional "login node" approach:

  • Interactive access to data
  • Shared home/scratch space
  • Integrated user accounts/groups via LDAP
  • Ability to launch jobs via TORQUE or Hadoop/Spark
  • Allow users to launch visualization tools, do development on the system
  • Integration with Globus.

workbench-roger

The ROGER system is a single set of nodes that serve distinct purposes: OpenStack, TORQUE/PBS and Hadoop (as well as GPFS/Globus). 

  • OpenStack
  • TORQUE: Traditional batch cluster
  • Hadoop cluster: HDFS over GPFS. Has Twitter and BigRaster data

Use cases:

  • Developer ("Henry") currently accesses TORQUE login node to run qsub jobs and Workbench for visualization.  Two different user accounts, two different home directories.
  • Adding users to Globus requires admin to map Globus identity to user on ROGER

Issues:

  • Auth
  • FS/permissions
  • Scratch/user space
  • Best way to submit jobs
  • Docker root problem

Trust:

  • System might be considered if it could be configured securely by Admin. (No Docker root, PAM/SSD in all containers)
  • We might then be able to mount the /home and /scratch spaces directly

Discussed similar projects in CyberGIS:

  • "Bastion" servers to provide access to OpenStack projects using PAM/SSSD and mounting the /home directory from ROGER
  • TopoLens
    • Uses community account (via NCSA identity)
    • Deployed via Docker in OpenStack
    • Data mounted via NFS
    • Uses parimico to submit jobs to TORQUE
    • Jupyerhub uses topolens API
    • Deployed on Swarm
  • Hydroshare
    • Working on iRODS support
  • Twitter data on HDFS available now with some geoanalytics
  • Also "BigRaster" data in HDFS

With any HPC cluster, there are for key problems: 

  • Security
  • Data management
  • Job submission
  • Software environment

Ideally HPC moves to cloud models (e.g., object storage, API based access)

Batch systems: SLURM, TORQUE-PBS, Condor, HTCondor=MOAB

XSEDE moving to Globus Auth and role-based auth.


  • No labels