• notify users ASAP
  • suggest creating very small cluster at beginning, then when we're happy "flip the switch"
  • user-notification of change
  • cross-queue fair-share settings
  • multiple levels of job preemption
  • "node reservation time" to allow larger MPI jobs to run
  • management control: members of various groups get queue management permissions, or node management permissions (e.g. to close nodes)
  • worry about openmpi being well tested (preliminary tests look OK) or if we need a new mpi
  • No labels