Markdown Version | Recording 1 | Recording 2 | Recording 3 | Recording 4

Session Date/Time: 23 Jul 2024 16:30

anrw

Summary

The Apply Networking Research Workshop (ANRW) 2024 kicked off with a welcome message, followed by a keynote from Sharad Gargwal of Microsoft on leveraging AI/ML for 5G and 6G systems. The workshop then transitioned into a measurement session featuring two papers. The first paper discussed the challenges of HTTP/3's Extensible Prioritization Scheme (EPS) in real-world deployments, and the second paper introduced the concept of the observer effect in computer networks and its implications for network measurement.

Key Discussion Points

• Keynote (Sharad Gargwal):
  • Leveraging cloud scale to overcome limitations of deployed network protocols.
  • The importance of using off-the-shelf learning algorithms whenever possible.
  • Mitigation strategies for hallucinations in generative AI models for network management.
  • High bandwidth, low latency, and high reliability are key challenges in Cloudified Telecom infrastructure.
  • Traffic engineering and management using learning algorithms to shift traffic from overloaded peering links.
  • Using AI agents to automate troubleshooting guide execution, reducing downtime in 5G deployments.
• HTTP/3's EPS in the Wild:
  • Significant heterogeneity in browser and server implementations of EPS.
  • Lack of full support for EPS by major deployments, creating a "chicken and egg" problem.
  • Inconsistent application of heuristics and default scheduling behaviors.
  • Recommendations include full support for EPS, improved developer APIs, further research into loading heuristics, and realistic testing resources.
• The Observer Effect in Computer Networks:
  • Measurement overhead can impact network performance (e.g., packet loss).
  • The uncertainty relation quantifies the trade-off between measurement accuracy and network impact.
  • Introduction of the "observer factor" as a practical metric for evaluating the overhead of measurement protocols.
  • The observer factor is easy to compute and allows for apples-to-apples comparison of different measurement methods.

Decisions and Action Items

• Potential Action Item: Investigate the possibility of standardizing the observer factor calculation to aid in the design of future measurement protocols within the IETF.

Next Steps

• The workshop will continue with sessions on routing and congestion, security and energy, and lightning papers.

Session Date/Time: 23 Jul 2024 20:00

anrw

Summary

The Applied Networking Research Workshop (anrw) session featured three presentations covering diverse topics: location-aware advertising in ENICAS IP networks, empirical characterization of Anycast convergence time, and diffusing computation for loop-free short-path routing. A fourth presentation covered traffic management focusing on Quality of Experience. The first presentation investigated regional trends and selective announcements in ENICAS routing. The second presentation empirically measured Anycast convergence times from an end-host perspective. The third presentation introduced a new routing protocol, "Dare," designed to improve the efficiency of diffusing computations. The fourth presentation looked at better Quality of Experience traffic management. Each presentation was followed by a Q&A session.

Key Discussion Points

• Location-Aware Advertisements in ENICAS IP Networks:
  • Observed that 84% of Enicast networks selectively announce at least one Enicast prefix.
  • Demonstrated that networks often leverage regional providers to connect POPs and end users.
  • Discussed the need to consider geolocation for accurate routing simulations.
• Empirical Characterization of Anycast Convergence Time:
  • Found that 80% of prefixes converge within 10 seconds in the Anycast setting, with faster convergence observed when more vantage points are active.
  • Identified challenges in reverse probing measurements due to load balancing at K-da's arc nodes.
• Dare: Diffusing Compute More Efficient for Loop Free Short-Spat Routing:
  • Introduced a new routing protocol ("Dare") to improve the efficiency of diffusing computations compared to DUAL.
  • Discussed how Dare remembers which neighbors initiated diffusing computations, reducing signaling overhead.
  • Explained that Dare allows routers to change successors at any time, regardless of active or passive state.
  • A concern was raised regarding interoperability with existing DUAL implementations.
• Traffic Management Focusing on Quality of Experience:
  • Presented a novel approach to traffic management, focused on the Quality of Experience in real time media.
  • Showed the need for better collaboration among the actors, in particular the Application and the Network.
  • Presented concerns with respect to Net Neutrality regulations

Decisions and Action Items

• Location-Aware Advertisements in ENICAS IP Networks:
  • Future work: Investigate the underlying factors driving the need for selective routing towards regional neighbors.
• Empirical Characterization of Anycast Convergence Time:
  • Future work: Increase measurement resolution by scanning faster and investigate regional differences in convergence times, including looking at IPV6.
• Dare: Diffusing Compute More Efficient for Loop Free Short-Spat Routing:
  • Future work: Incorporate "Dare" into a routing protocol for larger-scale simulations and evaluations, and comparison with Link State protocols.
• Traffic Management Focusing on Quality of Experience:
  • Future work: Explore how to efficiently calculate these spatial complexity curves all the time.

Next Steps

• The presenters will continue their research based on the feedback received.
• Further discussions and collaborations are encouraged on the topics presented.

Session Date/Time: 23 Jul 2024 22:30

anrw

Summary

This meeting focused on several aspects of network resilience and security, including carbon-aware routing with BGP, integrity protection for in-band telemetry (IOAM), security assessment of internet paths for critical infrastructure, and fixes for KeyTrap vulnerabilities. The session covered both theoretical concepts and practical implementation challenges, sparking discussions about deployment considerations, performance trade-offs, and potential improvements to existing protocols.

Key Discussion Points

• Low-Car BGP:
  • Discussion around the motivation for green routing and how to incentivize ISPs to prioritize greener AS paths.
  • Metrics for measuring greenness were explored, including carbon intensity and carbon performance ratings.
  • Concerns were raised about the potential for increased congestion on green paths and the impact on overall network performance.
  • Question on how carbon intensity values can accurately represent the ingress and egress pairing of traffic through an AS.
• IOAM Integrity Protection:
  • Several options for integrity protection were reviewed, focusing on GMAC-based methods.
  • Trade-offs between check-on-transit nodes and zero-trust solutions were discussed.
  • Performance evaluations of different options were presented, highlighting the impact on encapsulation, transit, and decapsulation nodes.
  • The use of IPsec for inter-domain use cases was considered, but ultimately deemed less desirable due to potential path changes.
• Assessing Security of Internet Paths:
  • Research questions centered on the number of fully and partially ROV-protected paths and the impact of a CI's upstream provider implementing ROV.
  • Method for calculating the security status of a path by combining path finding with ROV scoring and RPKI invalid prefix filtering.
  • Concerns raised about detecting IP addresses of critical services to improve the effectiveness of the study.
  • Challenges collecting paths on the data plan.
• Protocol Fixes for KeyTrap Vulnerabilities:
  • Recap of KeyTrap attacks, focusing on denial-of-service attacks through DNSSEC validation.
  • Discussion on how short-term fixes address specific attack vectors but may introduce unmatched complexity and performance issues.
  • Suggestions for long-term fixes: validation budget management, unique key identifiers, and relaxation of absolute specification requirements.
  • Considerations of how DNS key tags should be generated and how they impact validation.
• Unified SAV Framework:
  • Discussion around challenges for wide deployment of Source Address Validation (SAV), including technical knowledge, limited open-source efforts and performance concerns.
  • Presentation of a unified architecture for implementing and emulating different SAV mechanisms.
  • Evaluation of different SAV mechanisms in terms of accuracy, control plane and data plane performance, and scalability.
  • Comparison of traditional and new implementations of SAV, such as Enhanced D-South (E-D-South) and traditional D-South.

Decisions and Action Items

• IOAM Integrity Protection:
  • The draft will specify option 2 (validation at the end with no header check) because it fulfills the requirement of zero-trust solution.
  • Explore the use of delegation to remove validator jobs out of forwarding path.
  • Consider defining option 3 (neighbor validation) in a separate document later.
• KeyTrap Vulnerabilities:
  • Discuss with protocol designers how to approach KeyTag issues generically.

Next Steps

• Further discussion on the anrw mailing list to explore the topics raised in the meeting.
• Continued work on refining the IOAM integrity protection draft.
• Investigate practical incentives for wider deployment of UNISAV.

Session Date/Time: 24 Jul 2024 00:30

anrw

Summary

This session covered a variety of topics related to networking research. Presentations included measuring content locality, investigating data center network protocols, defending against QUIC handshake flooding attacks using deep reinforcement learning, evaluating BBRv3 in the public internet, using machine learning for network modeling, and exploring the potential of large language models for protocol analysis and diagnostics.

Key Discussion Points

• Content Locality Measurement:
  • Definition of local vs. external traffic and the three Ps of motivation (policy, performance, persistence).
  • Methodology involving domain lists, censorship filtering, website hosting determination, and geolocation.
  • Limitations including website complexity, proxy probe availability, and the "fog of cloud".
  • Future work: recursive searching, trace route/latency measurements, reason analysis, greater visualization, and increased categorization.
• Data Center Network Protocols:
  • Investigating protocols for folded clause topologies, focusing on energy, security, and configuration concerns.
  • Comparison of BGP-based stacks with a new protocol (MRMTP).
  • Testing using the FABRIC testbed and automation tools.
  • Future work: a non-IP-based solution backward compatible with IPv4/IPv6.
• QUIC Handshake Flooding Defense:
  • Presentation of QuickPro, a deep reinforcement learning approach to defend against QUIC handshake flooding attacks.
  • Use of real-time rate limiting, connection prioritization, traffic shaping, and anomaly detection.
  • Integration with existing QUIC implementations and future testing against diverse attacks.
• BBRv3 Evaluation:
  • Evaluation of BBRv3's coexistence with cubic flows and the impact of ECN.
  • Results indicated unfairness of BBRv3 towards cubic flows.
  • Suggestion for community consensus on guidelines for evaluating CCAs.
• ML for Network Modeling:
  • Challenges in deploying ML models for networking due to generalizability issues.
  • Presentation of NetMosaic, a tool for scaling data collection using public GitHub repositories.
  • Demonstration of improved generalizability through training with realistic network conditions.
• LLMs and Protocols:
  • Exploring the use of large language models for protocol diagnostics, testing, simulation, and prototyping.
  • Challenges including complex fields, encryption, system configurations, security, hallucinations, and efficiency.
  • Potential use cases, such as feeding traces to an LLM for diagnostics of behavior.

Decisions and Action Items

• Action Item: Donesh (BBRv3 presentation) to review the congestion controller working group draft for guidelines on evaluating congestion controllers. (Reese Enghardt offered the reference)
• Action Item: Danesh Sein Ali to look into ECT zero vs ECT one regarding ECN

Next Steps

• Content Locality Measurement: Future work includes: recursive searching, trace route/latency measurements, reason analysis, greater visualization, and increased categorization.
• Data Center Network Protocols: Looking at non-IP-based solution backward compatible with IPv4/IPv6.
• QUIC Handshake Flooding Defense: Testing against diverse attacks and developing a comprehensive QUIC attack dataset.
• ML for Network Modeling: Develop closed-loop workflows to study how the network conditions affect the data quality and scale data collection for more wrappers.
• LLMs and Protocols: Continued research on diagnostics and continued exploration of different methodologies of leveraging the LLMs for network protocols.