Re: [openstack-dev] [ironic][nova] Indivisible Resource Providers

[Jay Pipes]

On 07/27/2016 10:48 AM, Sam Betts (sambetts) wrote:

While discussing the proposal to add resource_class’ to Ironic nodes for
interacting with the resource provider system in Nova with Jim on IRC, I
voiced my concern about having a resource_class per node. My thoughts
were that we could achieve the behaviour we require by every Ironic node
resource provider having a "baremetal" resource class of which they can
own a maximum of 1. Flavor’s that are required to land on a baremetal
node would then define that they require at least 1 baremetal resource,
along with any other resources they require.  For example:

Resource Provider 1 Resources:
Baremetal: 1
RAM: 256
CPUs: 4

Resource Provider 2 Resources:
Baremetal: 1
RAM: 512
CPUs: 4

Resource Provider 3 Resources:
Baremetal: 0
RAM: 0
CPUs: 0

(Resource Provider 3 has been used, so it has zero resources left)

 Given the thought experiment it seems like this would work great with
one exception, if you define 2 flavors:

Flavor 1 Required Resources:
Baremetal: 1
RAM: 256

Flavor 2 Required Resources:
   Baremetal: 1
RAM: 512

Flavor 2 will only schedule onto Resource Provider 2 because it is the
only resource provider that can provide the amount of resources
required. However Flavor 1 could potentially end up landing on Resource
Provider 2 even though it provides more RAM than is actually required.
The Baremetal resource class would prevent a second node from ever being
scheduled onto that resource provider, so scheduling more nodes doesn’t
result on 2 instance on the same node, but it is an inefficient use of
resources.

To combat this inefficient use of resources, I wondered if it was
possible to add a flag to a resource provider to define that it is an
indivisible resource provider, which would prevent flavors that don’t
use up all the resources a provider provides from landing on that provider.


Hi Sam,

As Ed said, this isn't the direction we are going (in fact, it's 
essentially the situation we are trying to get ourselves *out of*). The 
new placement API has a resource provider record for each baremetal 
resource node that Ironic exposes to tenants. Each of those resource 
providers has an inventory record containing a total value of 1 for a 
resource class that identifies the type of baremetal hardware (the 
Ironic node class that is being currently introduced).


There are no inventory records for the VCPU or MEMORY_MB resource 
classes for any resource provider that is an Ironic baremetal resource 
node. The inventory is only a single unit of a dynamic resource class 
that matches the Ironic node class -- thus representing the indivisible 
nature of the baremetal resources.


Best,
-jay

__
OpenStack Development Mailing List (not for usage questions)
Unsubscribe: openstack-dev-requ...@lists.openstack.org?subject:unsubscribe
http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev

Re: [openstack-dev] [ironic][nova] Indivisible Resource Providers

[Ed Leafe]

On Jul 27, 2016, at 9:48 AM, Sam Betts (sambetts)  wrote:

> While discussing the proposal to add resource_class’ to Ironic nodes for 
> interacting with the resource provider system in Nova with Jim on IRC, I 
> voiced my concern about having a resource_class per node. My thoughts were 
> that we could achieve the behaviour we require by every Ironic node resource 
> provider having a "baremetal" resource class of which they can own a maximum 
> of 1. Flavor’s that are required to land on a baremetal node would then 
> define that they require at least 1 baremetal resource, along with any other 
> resources they require.

I was going to respond pointing out the issues with that approach, but then the 
rest of your email did just that. :)

I strongly preferred the approach that each particular hardware configuration 
would be a class, so that if you had 50 nodes with configuration A, and 20 
nodes with configuration B, that that would be reflected in two resource 
classes, with corresponding inventories to match the nodes. When a node is 
provisioned, that inventory is decremented. This would be much more consistent 
with the rest of the resource provider design, as having many, many classes all 
of which represent identical hardware seems backwards.

-- Ed Leafe






__
OpenStack Development Mailing List (not for usage questions)
Unsubscribe: openstack-dev-requ...@lists.openstack.org?subject:unsubscribe
http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev

[openstack-dev] [ironic][nova] Indivisible Resource Providers

[Sam Betts (sambetts)]

While discussing the proposal to add resource_class' to Ironic nodes for 
interacting with the resource provider system in Nova with Jim on IRC, I voiced 
my concern about having a resource_class per node. My thoughts were that we 
could achieve the behaviour we require by every Ironic node resource provider 
having a "baremetal" resource class of which they can own a maximum of 1. 
Flavor's that are required to land on a baremetal node would then define that 
they require at least 1 baremetal resource, along with any other resources they 
require.  For example:

Resource Provider 1 Resources:
Baremetal: 1
RAM: 256
CPUs: 4

Resource Provider 2 Resources:
Baremetal: 1
RAM: 512
CPUs: 4

Resource Provider 3 Resources:
Baremetal: 0
RAM: 0
CPUs: 0

(Resource Provider 3 has been used, so it has zero resources left)

 Given the thought experiment it seems like this would work great with one 
exception, if you define 2 flavors:

Flavor 1 Required Resources:
Baremetal: 1
RAM: 256

Flavor 2 Required Resources:
Baremetal: 1
RAM: 512

Flavor 2 will only schedule onto Resource Provider 2 because it is the only 
resource provider that can provide the amount of resources required. However 
Flavor 1 could potentially end up landing on Resource Provider 2 even though it 
provides more RAM than is actually required. The Baremetal resource class would 
prevent a second node from ever being scheduled onto that resource provider, so 
scheduling more nodes doesn't result on 2 instance on the same node, but it is 
an inefficient use of resources.

To combat this inefficient use of resources, I wondered if it was possible to 
add a flag to a resource provider to define that it is an indivisible resource 
provider, which would prevent flavors that don't use up all the resources a 
provider provides from landing on that provider.

Sam


__
OpenStack Development Mailing List (not for usage questions)
Unsubscribe: openstack-dev-requ...@lists.openstack.org?subject:unsubscribe
http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev