Session Date/Time: 27 Jul 2023 22:30

rift

Summary

The rift meeting in San Francisco covered the progress of the base specification, key-value registry, auto-flood reflection, and YANG model. There was also discussion on RIFT extensions for SRv6, and DC routing for high-performance computing (HPC) specifically dragon fly topologies. The meeting discussed implementation efforts, potential interoperability testing, and future directions for the RIFT protocol.

Key Discussion Points

Base Spec Progress: Readability improvements in version 18. Addressing AD comments related to applicability draft considerations such as misscabling, TTL values, and IPv4 broadcast usage. Focus on addressing all existing comments and reviews.
Key Value Registry: Minor maintenance and editorial fixes. Dependent on the base spec being finalized.
Auto-Flood Reflection: Fixed errors in ASCII topologies. Added matching SVGs. Completed the YANG model. Discussed working group adoption after the base spec is finalized.
RIFT YANG Model: Nearing completion with minor modifications. Further review encouraged and has been submitted to the IESG. Will wait for base spec stability before further IESG processing.
RIFT Extensions for SRv6 (SRv6 Locator KW TIE): Initial attempt to support SRv6 in data centers. Discussion on scalability challenges associated with flooding large key-value entries. Need to avoid broadcast approach and instead target configuration distribution. Suggestion to evolve the draft into a more generic mechanism for configuration distribution.
DC Routing for High-Performance Compute (Dragonfly Topologies): Explored the applicability of RIFT to dragonfly topologies for HPC and potentially AI/ML clusters. Concerns about the complexity of BGP solutions. Agreement that RIFT could support dragonfly topologies with summarization. Discussion of horizontal links between spines. Concerns were raised about supporting non-equal cost multipath and minimizing latency for best performance in sparse dragonfly implementations. Congestion notification and awareness. Also discussed HPC versus ML use cases with Dragonfly.

Decisions and Action Items

Jordan Head (Juniper): Double-check and address all outstanding AD comments on the base spec and reach out to Drew.
Sanghi (ZTE): Update the working group on progress towards interoperability testing, and investigate the possibility of a RIFT hackathon in Prague.
Tony P. (Juniper): Investigate how to address concerns on scaling the size of the locator key value tie in the SRv6 draft and how to generalize the draft into something more generic.
Authors of RIFT extensions for SRv6 (SRv6 Locator KW TIE): Evolve the draft to be more generic to handle various use cases.

Next Steps

Finalize the RIFT base specification.
Progress towards working group adoption of auto-flood reflection.
Continued discussion on applying RIFT to dragonfly topologies and high-performance computing environments in particular AI/ML and synchronization methodologies.
Potential interim meeting to discuss primary/secondary approaches for state synchronization.