Open Stack

On 09/05/2017 11:02 AM, Andrey Volkov wrote:
> For example, I have SR-IOV PF with four ports (P_i), two of them are
> connected to one switch (SW_1) and other two to another (SW_2). I
> would like to get VFs from distinct ports connected to distinct
> switches (more details can be found in spec [1]), how it can be
> modeled with nested resource providers?
> 
> Several possible solutions I see:
> 
> 1)
>                            --- compute node -----
>                       ----/     /          \     \------
>                 -----/         /            \            \-------
>                /              /              \                   \
>           SR-IOV PF         SR-IOV PF         SR-IOV PF         SR-IOV PF
>         (traits:P1,SW1)  (traits:P2,SW1)   (traits:P3,SW2)   
>   (traits:P4,SW2)
>              :                 :                 :                 :
>             / \               / \               / \               / \
>            /   \             /   \             /   \             /   \
>         VF1    VF2        VF3    VF4        VF5    VF6        VF7    VF8
> 
> 
> 2)
>                          compute node
>                        /              \
>                      /                  \
>             SR-IOV PF                   SR-IOV PF
>            (traits:SW1)                (traits:SW2)
>              /      \                    /      \
>             /        \                  /        \
>        SR-IOV PF     SR-IOV PF     SR-IOV PF     SR-IOV PF
>       (traits:P1)   (traits:P2)   (traits:P3)   (traits:P4)
>           :             :             :             :
>          / \           / \           / \           / \
>         /   \         /   \         /   \         /   \
>      VF1    VF2    VF3    VF4    VF5    VF6    VF7    VF8
> 
> 
> 3) Use tags for inventories, so the problem can be solved without 
> complex structures.
> 
> Are the described options applicable or there are other to solve the issue?

Hey Andrey, sorry for the long delay in getting back to you on this.

A few points to consider.

First, you would only need to have VFs represented as separate nodes in 
the tree (i.e. separate resource providers) if the VFs had different 
traits associated with them (for instance, if some hardware offloads 
were programmable on some of the VFs but not others).

Secondly, the port should not be a trait. I think perhaps you're 
referring to a physical network as a port, though. If this is the case, 
then yes, you'd want that physical network to be a custom trait (e.g. 
CUSTOM_PHYSNET_INTRANET or something like that).

So, with that said, you're looking instead at a structure like this:

CN
|--> PF1
|--> PF2
|--> PF3
|--> PF4

with the following records in the inventories table:

RP            RC              Total
------------------------------------------
CN            VCPU            4
CN            MEMORY_MB       8192
RP1           SRIOV_NET_VF    2
RP2           SRIOV_NET_VF    2
RP3           SRIOV_NET_VF    2
RP4           SRIOV_NET_VF    2

and the following in the resource_provider_traits table:

RP            TRAIT
-------------------------------------------
RP1           CUSTOM_PHYSNET_A
RP1           CUSTOM_SWITCH_1
RP2           CUSTOM_PHYSNET_B
RP2           CUSTOM_SWITCH_1
RP3           CUSTOM_PHYSNET_A
RP3           CUSTOM_SWITCH_2
RP4           CUSTOM_PHYSNET_B
RP4           CUSTOM_SWITCH_2

We discussed yesterday that the Neutron ML2 agent would be responsible 
for inserting trait associations for the resource providers representing 
the SRIOV physical functions.

We still need a way to represent a request to placement to find 
allocation candidates for like resources, though. As you pointed out, 
I've thought about using multiple requests to placement from the 
conductor or scheduler. We could also do something like this:

  GET 
/allocation_candidates?resources=VCPU:1,MEMORY_MB:1024&resources1=SRIOV_NET_VF:1&required1=CUSTOM_PHYSNET_A,CUSTOM_SWITCH_1&resources2=SRIOV_NET_VF:1&required2=CUSTOM_PHYSNET_A,CUSTOM_SWITCH_2

Best,
-jay