Lei huang presented <link>

• HPA mentioned that they want to support GPUs and FPGAs but it was postponed beyond dublin due to it not being in the openstack S release.

In dublin timeframe; agree on the architecture requirements to prepare for Release E for implementation. This means to aim to have the architecture implications understood around April for the potential Architecture F2F meeting.

Alex brought up that HPA is also based on standards and the modelling of FPGA.

It was commented that with the HPA; multi-vim has a way to discover the accelerator and monitor the accelarator. The presentation process that a new function (AM) contacts multicould. It was commented that m-cloud does this today for other accelators. What is not there is to do the health checks.

• It was suggested that AM microsevice lies in multi-cloud.

• It is suggested that this process needs to be revisited in light of this information.

Note: It was acknodeged that the requirements were correct.

Alex is pointing out that FPGAs may need special considerations as they may not be directly attatched.

Instantiate flow:

-

The proposal identifies improvements in the HPA approach.

-> the proposal is that use should specific "what" is required instead of "how to achieve" in the HPA modelling. It was commented that HPA was designed to do the "how". It was indicated that there could be a layer above HPA (sepcification wise) to specify "how".

• Ramki raised that we shoul look at K8s capabilities as well, and have a look at a gap analysis.

• it was commented that there dependicies on openstack capabiltiies when it comes to reporting the VIM level performance of VNFs.

On the slide of the relationship between AM & HPA.

• AM is used to both described a capability; as well as discribe an implementation approach. It would be useful to seperate these.

There was the question about what role ONAP should play in managing the acceleration resources. There was alignment in that discovery, monitoring, consumption, remediation; There was more to discuss in HPA.

Have the HPA competence join the AM calls.

Architectural impact diagram would be helpful.

It was commented that it should be done agnostic to the orchestration both (VFC, or SO).

Conclusion

• Recongition for the requirements

• Come back to Arch after further discussion.

• discussion on the methods to be used.

• An architecture view would be appreciated.

• A new model is proposed to described the accelration managment

• Consider beyond openstack, and try to be infrastructure agnostic.