I'd like to enlist some old servers (e.g, G5 HP servers) for TensorFlow distributed computation. They're running Windows 2012 r2 on the bare-metal. I'm thinking about creating a Ubuntu virtual machine guest on the servers with VirtualBox or Hyper-V to operate as the TensorFlow cluster workers.
Has anyone done any performance benchmarking on a similar set-up (a) between different virtualization software, and (b) versus bare-metal?
Are there any customizations / configurations that should be used to maximize performance on this set-up?
I note that I was unable to install modern operating systems (e.g., Ubuntu 16 and 19 and CentOS7) on the bare-metal because of antiquated RAID controllers that can't be disabled (hence why I'm exploring virtualization).