Session Date/Time: 09 Nov 2021 16:00

grow

Summary

The grow session included two technical presentations. Paolo Lucente provided an update on the draft to add TLV support to BMP Route Monitoring and Peer Down messages, proposing a specific mechanism for indexed TLVs. Edward R. P. B. M. L. K. M. L. E. presented research on using eBPF for on-host load balancing of telemetry data, focusing on ensuring correlated streams from a single device are directed to the same processing daemon.

Key Discussion Points

TLVs for BMP Route Monitoring and Peer Down Messages (Paolo Lucente)
- Problem: Existing BMP messages, specifically Route Monitoring and Peer Down, lack support for TLVs.
- Proposed Solution: Introduce TLV support to these messages, requiring a bump in the BMP version to v4.
- TLV Types: The draft defines both generic and indexed TLVs.
- Indexed TLVs: These are crucial for Route Monitoring messages to specify whether a TLV applies to the entire BGP message or a specific Network Layer Reachability Information (NLRI).
- Revision 6 Update: The key change in revision 6 clarifies the use of the index field:
  - An index value of 0 indicates the TLV applies to the entire BGP message.
  - A positive index value indicates the TLV applies to a specific NLRI within the message.
- Status: This change is considered the final outstanding piece of work for the draft.
On-Host Load Balancing for Telemetry using eBPF (Edward R. P. B. M. L. K. M. L. E.)
- Context: The need to combine diverse network telemetry streams (e.g., BGP, IPFIX, Yang Push) from a single network device on a target host for correlation and analysis.
- Challenge: Ensuring that all telemetry data streams originating from the same device, potentially using different transport protocols and ports, are consistently directed to the same processing daemon on the collection host.
- Baseline (Proxy) Approach: Using traditional proxies (e.g., HAProxy for TCP, nfd for UDP) in front of telemetry daemons to route traffic.
  - Drawbacks: Introduces additional components, overhead, latency, reliability issues (e.g., update cycles, configuration complexity), and administrative burden due to disparate configuration syntaxes.
- eBPF-based Solution: Leverage the SO_REUSEPORT socket option in conjunction with a custom eBPF program.
  - SO_REUSEPORT provides stateless load balancing but defaults to hashing the entire flow, potentially splitting streams from a single device if they use different protocols.
  - eBPF Enhancement: An eBPF program is attached to the SO_REUSEPORT group to customize the hash function. It specifically hashes only the IP source address, ensuring all traffic from a given device (assuming a single source IP) is directed to the same daemon.
  - Daemon Management: Collectors register their intent (e.g., "I am 1 of N collectors"). The eBPF program balances across the intended number of daemons. If a daemon is temporarily down (e.g., during a rolling update), the eBPF program will respond with a TCP reset or drop UDP datagrams, preventing cascading failures and ensuring stability.
- Benefits of eBPF Approach:
  - Stateless load balancing with minimal configuration (only the intended number of daemons).
  - Ensures correlation by device identity (source IP).
  - Improved stability across restarts and during updates; prevents cascading failures.
  - Eliminates the need for additional proxy daemons, reducing system complexity, vulnerabilities, and configuration overhead.
  - Portable due to eBPF's "compile once, run everywhere" capability.
  - Demonstrated performance improvements, including approximately 20% CPU time savings and better ramp-up performance compared to proxy-based solutions.

Decisions and Action Items

Decision: The draft "BMP TLVs for Route Monitoring and Peer Down Messages" appears to be complete with the proposed changes in revision 6.
Action Item: Paolo Lucente will initiate a Work Group Last Call for the "BMP TLVs for Route Monitoring and Peer Down Messages" draft on the mailing list.

Next Steps

The grow working group will review and provide feedback on the Work Group Last Call for Paolo Lucente's BMP TLVs draft.
No immediate next steps were identified for the eBPF telemetry load balancing research, as it was presented as academic work and a master's thesis.