BMC (Bare-Metal Container) is relased

Jeremy Eder jeder at redhat.com
Wed Dec 21 12:29:53 UTC 2016


On Tue, Dec 20, 2016 at 8:34 PM, Kuniyasu Suzaki <k.suzaki at aist.go.jp>
wrote:

>
> From: Jeremy Eder <jeder at redhat.com>
> Subject: Re: BMC (Bare-Metal Container) is relased
> Date: Tue, 20 Dec 2016 10:04:12 -0500
>
> >  Eric W. Biederman <ebiederm at xmission.com>
> > wrote:
> >
> >> In the past I have always seen this sort of problem handled as a
> >> property to the scheduler, and a declaration that a certain job needs
> >> certain properties.
> >
> > Indeed, we have discussed this as a property of the scheduler in
> kubernetes.
> >
> > - Nodes are given attributes (i.e. THP=never)
> > - Users declare their requirements (i.e. THP=never)
> > - User submits job, scheduler does matching to node
> > - Node executes job
> >
> > This allows for shared control plane, serving disparate workloads on
> hybrid
> > infra.
> >
> > ​THP=never is just another node attribute, as could be things like
> > - Does a DPDK-capable NIC or GPU or FPGA exist, and is unused
> > - What kernel version is it running
> > - Is the node tuned correctly
> >
> > ...myriad other possibilities, in a rapidly evolving space.
>
> It sounds interesting. I did not know that kubernetes has such a function.
> Is there HP or paper to described the detail? I want to compare the
> function with our BMC (Bare-Metal Container).
>

A lot of it is there now using taints and tolerations.  Arguably needs some
end-to-end polish/usability improvements.  Folks from Intel are working on
at least the hardware detection stuff to enable their CPUs/NICs in a
generic way.


> I guess the function requires nodes which turn on/off THP.
> On the other hand, BMC boots a node with the Linux kernel which is
> suitable for an application.
> BMC needs the overhead to boot a node and allows to select a node with
> consideration of machine architecture, power usage, etc.


Trying to avoid an orchestration debate :-)  Just figured I'd +1 Eric's
comment about how knowledge of node properties tends to eventually be
recognized at the scheduler level and use kube as an example.


More information about the Containers mailing list