• (10min) Drafts Update Summary arch usecase gap_openstack
• (20min) g-API guidelines discussion
• (20min) Nomad SB API discussion

Lingli, Keith, Yannan, Denis, Rajesh, Howard, Barat, Ola, Mario, Paul, Michele, Trinath, Jesper

• discussion slides
• work-in-progress

Instead of mandating AC, the proposed updated doc mandates SAL as required. The following sentences are added to clarify that we need SAL, which includes AC, in the aggregated view.

SAL is the software to hardware abstraction layer, which makes possible additional services which can be controlled by the orchestration layer. SAL can be deployed to either guest or host. SAL is required in all DPACC usages identified in this document.

Agreed.

Changes are proposed to both the figures and the text in Section 2.2.1 and 2.2.2, in order to clarify the correspondence between these two parts. The proposal is to use Section 2.2.1 to identify two usage types, each of which is with options and can be further instantiated with several usage examples. Some of the examples of the two usage types are illustrated in the following Section 2.2.2. But the intention is not to change what we have already agreed on for the usage examples.

Need more review and feedback offline.

Due to the time limit, Lingli suggested authors and contributors to review these changes after the meeting and leave comment online. Keith acked he would review and confirm the changes.

• discussion slides
• work-in-progress

Yannan refined the figure of the LTE stack, according to Vineet' comment to differentiate the control plane and data plane towards S1 interface. Keith suggested adding terminology section and define the wireless terms in this figure. Ola suggested to change "S1 User Plane" into "raw traffic" in the figure, and consider the integration of another application of IPSec tunneling between vRAN and vEPC.

Offline discussion need for figure revision. Yannan could help with adding definitions of the wireless terms.

Updated figures for virtio look-aside acceleration in Section 3 are confirmed. Started update discussion for subsequent figures virtio-ipsec-la (Denis) and virtio-pdcp-la in Section 5.

Lingli suggested to define NFVI packet processing case in Section 4, and depicted with a single box, like crypto, in other figures of Section 5. Discussion on how to reflect the cases of g-API, SAL and legacy kernel API usages in the figures.

1. Figure update: Denis would update ipsec-la and Rajesh would update pdcp-la figures in Section 5. Lingli, Yannan and Ola would discuss offline on LTE figure revision.
2. RAN Terminology: Yannan would help add definitions for RAN terms in Terminology Section.

• discussion slides
• work-in-progress

Fault Detection for accelerators (MUST) Reclaim of accelerators allocated to faulty VMs (MUST) Automatic detection and recovery of faulty agent (MAY) Fault Analysis or Problem Diagnosis for acceleration resources (MAY) Gaps are to be added next week after the consensus on the requirements statement.

Agreed.

The metedata example table is suggested to be removed from the doc, and serve as the basis for a separate discussion thread dedicated to DPACC metadata to be started next week.

Agreed.

• For Current draft

1. Finalize text for Section 5.
2. Address remaining issues/comments.

• New Threads

1. S-API for dpacc mgmt. discussion
2. Metadata for acceleration resource

• discussion slides
• latest proposal draft
• earlier proposal draft

Denis and Rajesh presented the proposal of specifying common guidelines for various DPACC usecases, including suggestions on basic types, api/variable naming convention, function structure, and return values.

Keith raised the question on the closed box approach would not work for open source, Lingli clarified that the position of this work is not to regulate the practice but serve as experience sharing from existing implementations to be shared with subsequent endeavors.

Agreed to start the work, Lingli would compose another Google doc for later collaboration.

Discussion went on to the proposal details (some of the suggestions from Keith)

• for naming convention, the acceleration model can be used as a parameter rather than part of the API name
• for common operations, one may use SET for MOD, so no need for MOD. Do not mandate that all implementations would support any common operations, like OPEN or CLOSE.
• for errcode definitions,
  • Keith suggested to use the errcode to identify the responsive module (assigned when it gets integrated with the framework) as well as the actual err conditions (might be further defined within the responsive module). Take 32-bit errcode for example, the first 16 bits can be used as the ID for the accelerator functionality, and the second 16 bits can be used as the actual errcodes for that specific accelerator. Ola was concerned about how to implement management for allocating the IDs for different modules/accelerators from different parties. Lingli suggested to specify the rules/keep a registry in this intended DPACC document.
  • Ola suggested that it does not make much sense to define +ve codes for success. Instead, consider use +ve to convey other useful information, e.g. the number of successfully processed data units.
  • Rajesh is concerned if it makes sense to define common errcodes, if one is expecting each module/accelerator to define its local errcodes. Lingli suggested common errcodes can be helpful either way we take.
• for call back functions, the suggestion is to define callback function types and then further define its arguments.
• for type term, use the Linux standard.

Rajesh suggested in also collaborate with other teams working on virtio-cryto at OASIS. Lingli would try to involve them into further discussion.

Examples are to be added for la-ipsec and la-pdcp usecases.

1. post the initial proposal of guidelines to online doc
2. address the comments
3. add examples

discussion slides

Skipped due to time limit, postponed to next week's call.