Session Date/Time: 06 Feb 2024 13:00

DETNET

Summary

This interim DETNET working group session focused on discussing the framework and context for selecting enhanced DETNET data plane mechanisms, primarily through the lens of a proposed taxonomy. David Black provided an overview of the current state of mechanism evaluation against requirements, emphasizing the need for a taxonomy to characterize and differentiate proposed scheduling and queuing solutions. Gino provided a detailed presentation of his draft on DETNET data plane solutions taxonomy, leading to extensive technical discussion on classification criteria, applicability, and the strengths and limitations of various approaches. While no formal decisions were made regarding specific mechanisms, the discussion provided valuable input for refining the taxonomy and highlighted areas for future work, including the need for common use cases for evaluation.

Key Discussion Points

Introduction and Agenda: The co-chairs (Lou Berger, János Farkas) welcomed participants, reminded them of IETF rules and policies (Notwell), and emphasized the importance of discussion and collaborative note-taking. The agenda included a framing discussion by David Black and a detailed presentation of the taxonomy draft by Gino.
Framework for Mechanism Selection (David Black):
- Initial evaluation of proposed enhanced DETNET data plane mechanisms against requirements indicates that they generally address requirements partially or completely, with no major gaps. This has also improved the requirements draft.
- The next step is to decide which mechanisms to standardize. It was noted that different mechanisms will suit different network types, and similar to TSN, multiple DETNET mechanisms could be standardized if distinct applicability differences can be identified.
- The role of the taxonomy is to characterize proposed scheduling and queuing mechanisms, identify key differences, and clarify their applicability implications. The working group might select one mechanism if multiple are highly similar, but ultimately, consensus drives the decision.
- Goals for the Taxonomy Discussion: Agreement on classification criteria, a top-down taxonomy of TSN and DETNET mechanisms, and agreement on the taxonomy's role in mechanism selection.
DETNET Data Plane Solutions Taxonomy (Gino):
- Purpose: To facilitate understanding of current and future data plane solutions, providing criteria, examples, strengths, and limitations. The current draft omits detailed descriptions of candidate solutions and use cases/applicability, which are planned for future inclusion.
- Candidate Solutions: The draft implicitly covers CQF variants (CS-CQF, TC-CQF, ECQF), FSQ variants (C-Core, WCSCF), Time-Slot Shaper variants (TSQF), and EDF variants.
- Taxonomy 1: Per-Hop Dominant Factor for Latency Bound: Classifies mechanisms based on the largest sum term in the worst-case latency bound.
  - Category 1: Max packet length / service rate (e.g., FSQ, C-Core).
  - Category 2: Sum of max packet lengths / link capacity (e.g., DRR).
  - Category 3: Sum of maximum burst sizes / capacity (e.g., ATS, CQF variants).
  - Discussion: Norman Finn clarified that the credit-based shaper is no longer part of ATS and that in TSN, nothing is typically prioritized ahead of ATS in the final selection stage. There was also discussion regarding the accuracy of ATS classification based on the referenced paper.
- Taxonomy 2: Periodicity: Divides solutions into periodic (e.g., CQF variants, ECQF) and non-periodic (e.g., ATS, C-Core, EDF). Periodic solutions offer less jitter but may have higher latency; non-periodic are more flexible.
  - Discussion: Norman Finn noted that the per-hop delay for CQF does not necessarily have to be a full cycle time, and bursts are typically buffered at the network edge before entering a CQF segment.
- Taxonomy 3: Network Synchronization: Categorizes solutions by synchronization requirements: phase synchronous, frequency synchronous, or asynchronous. Phase synchronous is complex but offers precise delay control; asynchronous is less complex but needs jitter control. All solutions require some level of synchronization, though precision varies.
- Taxonomy 4: Traffic Granularity: Classifies by the specificity of traffic entities handled: flow-level, flow-aggregate level, or class-level.
  - Discussion: Significant debate arose on how to classify solutions composed of multiple functional entities with different granularities (e.g., ATS, ECQF). The draft's initial proposal that the coarsest granularity component dominates was challenged by Norman Finn and Youi Li. They argued that "dominance" should clarify whether it refers to complexity (state to maintain) or performance (latency), and that a strict priority scheduler, even if class-based, might not always define the entire solution's granularity in a way that implies shared fate among flows.
- Taxonomy 5: Work-Conserving: Solutions are classified as work-conserving or non-work-conserving. Work-conserving offers lower average latency and statistical multiplexing gain but may require larger buffers to prevent congestion loss. Non-work-conserving avoids burst accumulation and offers better jitter control.
- Taxonomy 6: Target Transmission Time: Distinguishes between solutions that define an "on-time" (strict target) or "in-time" (before a target) for packet transmission.
- Taxonomy 7: Service Order: Classifies based on the primary decision factor for packet service order (e.g., rate-based, time/delay-based, arrival-based).
- General Discussion on Taxonomy Criteria:
  - Buffer Space Requirements: Norman Finn suggested this as an important criterion, noting its relation to work conservation, and potentially a new taxonomy or an aspect of strengths/weaknesses. The group agreed it's a critical aspect to consider.
  - Computational/State Requirements for Flow Establishment: Norman Finn highlighted the importance of "zero-touch" flow addition for scaling. David Black suggested this is more an engineering aspect for trade-off analysis than a primary taxonomy classification. Gino acknowledged that complexity has many facets, some of which are difficult to capture in a simple taxonomy.
  - Overall Role of Taxonomy: The group affirmed that the taxonomy helps identify distinct classes of algorithms but does not limit the engineering considerations for selecting suitable algorithms. The goal is not to create a point-based scoring system for selection.
Discussion on Next Steps: János Farkas reiterated that the group is contribution-driven and seeks to down-select solutions to standardize DETNET queuing mechanisms, aiming to address scaling requirements.

Decisions and Action Items

Decision: The working group leadership (Chairs and AD) will have an offline discussion regarding the immediate next steps for the working group.
Action Item: Gino and co-authors will revise and update the taxonomy draft, taking into account the discussions, particularly concerning:
- The classification of ATS and ECQF, clarifying aspects of "dominance" and "granularity" for multi-component solutions.
- Incorporating considerations for buffer space requirements and jitter as part of the solution characteristics.
- Checking recent ATS standards (802.1Qcr) for accuracy in the classification of its components and behavior.
Action Item: The working group encourages contributions of common use cases and simple topologies that can be used to evaluate and compare proposed DETNET queuing solutions.
Action Item: Further contributions and updates to the taxonomy or new drafts presenting alternative classification approaches are welcome on the mailing list.

Next Steps

Continue detailed technical discussions on the DETNET mailing list.
Target IETF 119 (Brisbane) for further discussions on the taxonomy and proposed solutions, with two slots requested for the DETNET WG. Gino plans to present an updated draft at IETF 119.
The working group leadership will discuss the possibility of scheduling additional virtual interim meetings before the July IETF.