Must have features

• GPU selection for training jobs
  • Select next available n GPUs (Select multiple GPUs for training #92)
  • Select specific n GPUs (Add ability to select which GPU to run on #19)
  • Testing

Nice to have features

• Restart jobs (this would help: Snapshot on signal BVLC/caffe#2253)
  • Pause jobs
  • Restart jobs after pausing or aborting (Add ability to continue training process after aborting #24)
    • Choose a different GPU when restarting
• Auto-scaling of dataset jobs to make most of available resources