This is an interesting concept and a good addition to all the IaaS offerings around. Just one year ago, while doing my Master's thesis, I pondered the possiblity of creating an IaaS that didn't involve virtual machines but physical servers instead. This should be great especially for high performance and special hardware needs. Next step would be all sorts of mixed blends between virtual and physical servers: real CPUs and GPUs, but virtualized storage, networking or what have you.
OpenStack 'can', but in practice the baremetal nova driver[0] has many issues, and the newer OpenStack Ironic Project[1] is not yet ready for production workloads.
Additionally, both baremetal and OpenStack Ironic have a focus on 'triple-O', or OpenStack on OpenStack[2] use cases -- that is using these projects to bootstrap an 'undercloud', for the hardware to be used by a higher level virtualized cloud.
> Additionally, both baremetal and OpenStack Ironic have a focus on 'triple-O', or OpenStack on OpenStack[2] use cases
Well, I'm sure they have issues but what gave you the impression they are focused on TripleO use cases?
As a TripleO dev I just use them.
Compared to more 'standalone' softwares like MAAS, these tools are integrated already into OpenStack, allowing you to deploy images on either virtual instances or baremetal without adding more software layers in charge of the images hosting, authentication, dhcp management and what not.
I was totally convinced that Amazon had baremetal instances available, but now I cannot find anything, so that could probably be a figment of my imagination.
I believe at one point the GPU instances were being given direct hardware access to the GPU as virtualized GPUs weren't yet a thing. So the GPU part was sort of baremetal I guess.
SoftLayer has been doing this since 2005. One API to provision and deploy compute on demand physical or virtual. Over 90% of their customer bases uses bare metal servers.
:-) Of course it's not inventing anything, it's just providing a convenient way to commission those servers that can also be driven by other tools. If you bring a lot of server resource on line this is very handy. Juju, for example, can put things on MAAS servers, and one of the things it could put on there is a virtualised environment that could also be managed by MAAS and juju to deploy other workloads. MAAS knows some stuff about the hardware too, for example, how many network cards the machine has and on which networks, which means you can interrogate MAAS to find machines suitable for a given workload.
