Gap Analysis on OpenStack for DPACC
Attendants: Lingli, Keith, Bob, Paul, Howard, Mario, Ola, Micheal, Rajesh
-
-
Update Summary
Minor editing on Section 1-3
Merge Section 4 "Requirements" with Section 5 "Gap Analysis". Requirements are grouped with the specific lifecycle management. And gaps are intended to be added accordingly
-
On Page 7 for "Nova Extension vs NOMAD"
The current table groups CPU instruction set extensions into the category that should be exposed the NOVA way.
Some of them, for instance Intel AES-NI and AVX2, can also be exposed as software implementation for abstract accelerator and be managed by NOMAD. (Keith)
Need another review and grouping accordingly. (Lingli)
Add a column for abastracted accelerator types to be supported by NOMAD. (Micheal)
In the table, "VMs direct access to accelerators" is stated as one of the pros for NOMAD way
For the Cons of NOMAD way, there is potential scalability issues for NOMAD, as the number of compute nodes under its management increases. (Paul)
On Page 8 for "Requirements for Resource Discovery"
A new requirement is proposed
Requirement VIM (or AC?) SHOULD maintain the mapping (change to mapping dependency?) between abstract acceleration resource to physical acceleration resource.
Remaining issues to be addressed/confirmed
Who should maintain the mapping, VIM or AC? Howard prefers VIM.
Should we change "mapping" into "mapping dependency"? Micheal suggests the latter.
Lingli suggests to post these questions to the online doc for feedback.
Keith feels the need to include also notification scheme when an accelerator break down. Lingli explains that there are corresponding requirements on "fault management" for "resource release" in following sections.
On Page 9 for "Acceleration resources features", which is aligned with ETSI NFV IFA004
Bob will bring more feedback after team review later.
Micheal is confused about the relationship between IFA004 and this document. Howard clarifies that IFA004 is high level considerations, while this doc is more focused on implementation gaps.
On Page 10 for "Gaps for Resource Discovery"
Page 10-15 are left for further discussion due to time limit
Follow-up Actions
Update the doc according to the comments during the call (Lingli)
Refine the table for "EPA v.s NOMAD" and post it to the online doc, for further feedback (Michele)
Refine work plans in Section 5, and bring Nomad Architecture and workflows to the next discussion (Lingli&Howard)
DPACC Use-cases
-
Lingli goes through the table of contents and calls for further review and feedback online
Rajesh adds a subsection for "DPDP offload acceleration usecase"
Lingli adds a subsection for "vRAN" as another "targeting applications"
Follow-up Actions
Summarize the current documents and call for review in the list (Lingli)
About Next Discussion
DPACC calls are canceled for next week, for the approaching Chinese New Year.
Lingli will continue work on Requirements and Gaps for the other four management phases (Resource Selection, Allocation, Update and Release) bring them to our further discussion.
Rajesh will continue work on common g-
API for virtio-XXX acceleration usecases and bring it to our next discussion.