Transparent Interconnection of Lots of Links (TRILL): Centralized Replication for Active-Active Broadcast, Unknown Unicast, and Multicast (BUM) Traffic

Weiguo Hao, Yizhou Li, Muhammad Durrani, Sujay Gupta, Andrew QuRFC 2018 · DOI

TL;DR

This paper addresses an issue in TRILL networks where active-active access can lead to incorrect packet drops due to Reverse Path Forwarding (RPF) checks failing. The proposed solution is "centralized replication," where Broadcast, Unknown Unicast, and Multicast (BUM) traffic is sent to a designated central node for replication and distribution, effectively resolving the RPF check problem and ensuring reliable traffic flow.

What Problem Does This Paper Address?

In modern data centers and enterprise networks, achieving high availability and load balancing often involves using "active-active" configurations. This means that multiple network paths or devices are simultaneously active, providing redundancy and distributing traffic. The Transparent Interconnection of Lots of Links (TRILL) protocol is designed to simplify network design by creating a Layer 2 overlay that behaves like a Layer 3 network, enabling multipath forwarding.

However, when TRILL is used in an active-active setup, particularly with a mechanism called "pseudo-nickname" (introduced in RFC 7781 to prevent MAC address flapping), a specific problem arises. This pseudo-nickname mechanism, while solving one issue, inadvertently causes another: Reverse Path Forwarding (RPF) checks can fail. RPF checks are a security and loop-prevention mechanism that ensures a packet is arriving on the expected interface. When the pseudo-nickname is used, intermediate network devices (RBridges in TRILL terminology) might see traffic arriving from an unexpected "ingress" point, leading to legitimate BUM traffic being dropped. This paper tackles this specific RPF check failure issue in active-active TRILL environments.

Methods

The core of the proposed solution is centralized replication. Instead of each RBridge handling BUM traffic independently and potentially causing RPF issues, all BUM traffic from an active-active edge group is sent to a single, designated centralized node. This centralized node is typically a distribution tree root, chosen for its high capacity and connectivity.

Here's how it works:

  1. Unicast TRILL Encapsulation: When an ingress RBridge receives BUM traffic from a Customer Equipment (CE) device in an active-active setup, it encapsulates this traffic using unicast TRILL, rather than the traditional multicast encapsulation. The ingress nickname in this encapsulation is set to the pseudo-nickname of the edge group.
  2. Centralized Node Processing: The unicast TRILL packet is sent to the designated centralized node. This node uses a special nickname, termed an "R-nickname", to identify itself for this specific replication purpose.
  3. Decapsulation and Replication: The centralized node decapsulates the packet and then replicates it.
  4. Distribution Tree Forwarding: The replicated packets are then forwarded to their destination RBridges using a standard TRILL distribution tree, established according to the base TRILL protocol (RFC 6325).
  5. Modified RPF Check: Crucially, to prevent the RPF failure, the RPF check on intermediate RBridges is modified. Instead of checking against the actual ingress RBridge (which is part of the edge group), the RPF check is performed as if the centralized node were the ingress RBridge. This ensures that traffic arriving at the centralized node's facing port is considered valid.

The document also introduces two new flags, "R" and "C", in the Nickname Flags APPsub-TLV to identify R-nicknames (for centralized replication nodes) and C-nicknames (for edge groups using centralized replication).

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • Pseudo-nickname's Double-Edged Sword: The pseudo-nickname mechanism, essential for preventing MAC flapping in active-active TRILL, creates a new problem with RPF checks. This highlights the complex interplay of features in network protocols.
  • Centralized Control for Distributed Problems: The solution elegantly shifts the burden of handling BUM traffic from potentially problematic distributed edge RBridges to a single, powerful centralized node. This simplifies the logic at the edge.
  • RPF Check Adaptation: The key to making centralized replication work is not just replicating traffic but also intelligently modifying the RPF check logic. Treating the centralized node as the "ingress" for RPF purposes is a clever workaround.
  • R-nickname for Specialization: The introduction of the "R-nickname" provides a clear way to distinguish traffic destined for the centralized replication process from regular unicast TRILL traffic.
  • Coexistence with Other Solutions: The paper emphasizes that this centralized replication solution can coexist with other TRILL features, such as Coordinated Multicast Trees (CMT), allowing network operators to choose the best approach for different edge groups.

Key Results

The primary result is the resolution of the RPF check failure issue for BUM traffic in active-active TRILL access scenarios. This leads to:

  • Prevention of incorrect BUM traffic drops: Legitimate BUM traffic is no longer dropped due to RPF failures.
  • Improved reliability and performance: Ensures that BUM traffic reaches its intended destinations in active-active environments.
  • Simplified edge RBridge logic: Offloads complex BUM handling to a central point.
  • Network Upgrade Path: The paper outlines that centralized nodes require software/hardware upgrades, edge RBridges need upgrades for unicast TRILL encapsulation, and transit nodes primarily need software updates for the RPF algorithm change.

Limitations and Future Work

  • Centralized Node Bottleneck: While the paper describes a solution, a single centralized node could become a bottleneck. The paper does address this by suggesting load-balancing across multiple centralized nodes using data labels (like VLANs) and multiple R-nicknames.
  • Complexity of RPF Modification: Implementing the modified RPF check requires careful configuration and understanding of the TRILL protocol extensions.
  • Future Work: The paper implicitly suggests future work in optimizing the distribution of centralized nodes, enhancing load-balancing mechanisms, and ensuring seamless interoperability with evolving TRILL features. The discussion on network upgrade analysis also points towards the practical implementation challenges and necessary steps for deployment.

Summary updated

BibTeX
@article{hao2018transparent,
  author = {Weiguo Hao and Yizhou Li and Muhammad Durrani and Sujay Gupta and Andrew Qu},
  title = {{Transparent Interconnection of Lots of Links (TRILL): Centralized Replication for Active-Active Broadcast, Unknown Unicast, and Multicast (BUM) Traffic}},
  journal = {RFC},
  year = {2018},
  doi = {10.17487/rfc8361},
}

Loading related content...

Figure