infrastructure provisioning roll-up

[eyberg]

before diving into anything I'd like to confirm precisely what each provider is going to want and also keep in mind that depending on which provider you are deploying to we might have different booting methods

• boot with GRUB #273 - boot with grub
  does not seem to be needed - out boot/boot2 should boot on any provider after boot2: read from disk using BIOS #480 (which is actually a prerequisite for booting on all providers)

• boot target on gce #275 - boot to GCE

  • virtio-scsi: https://cloud.google.com/compute/docs/images/building-custom-os

• boot on aws #276 - boot on AWS

  • Firecracker - https://groups.google.com/forum/#!topic/osv-dev/BAdpXmjtC24
    supports virtio-net and virtio-blk via mmio - needs virtio mmio support
    • eval firecracker && see what happens - eval firecracker #483
  • AWS C5:
    • nvme storage driver
    • ena network driver
  • AWS T2
    • xen-platform block driver
    • IDE storage driver - ATA PCI driver #871
    • xen-platform network driver

• boot to digital ocean

• boot to linode

• boot to alicloud

• vsphere rollup tkt #1118 - vsphere

  • vmxnet network driver
  • pvscsi driver
    It's also worth investigating deploying to these as well:

• openstack
  openstack is KVM-based
  openstack rollup #1225
  supports virtio-net and virtio-blk - should just work

• virtualbox

  • IDE or SCSI storage driver (does not support virtio-blk)

• hyper-v support #484 - azure (Hyper-V)

  • VMBus support
  • VMBus storage driver
  • VMBus network driver

so to clarify - whatever is made should be extensible enough to satisfy these - we can have a default whatever that is but also want/need the capability to switch out deploy targets as needed

https://devel.unikernel.org/t/unikernels-on-clouds-just-a-mirage/227

[eyberg]

@mkhon thanks for enumerating these -

I'm open to how this should be dealt - with but I think qemu/kvm should be our 'base case' which we build by default and everything else is an 'export' of some kind - if firecracker is truly that easy to get going that might be worth looking at cause it's new/relevant (although I think there's a lot more to do there)

i'm thinking at least from the end-user perspective one would do something like 'ops export gcloud my.img' - whether that means we re-build images from scratch or convert them that's up to you

[eyberg]

also - it'd be good to look at all the boot options available for these first too - if everything demands/wants grub then it's pretty easy to just rip out the current bootloader and throw that in instead but if they are going to have differences from provider to provider it should be taken into account

[mkhon]

I think we don't need grub for GCE (our stage0 should be enough), and I suspect that also for everything else our stage0 is enough (unless something will require EFI).