TL;DR
This paper introduces a novel router design that bridges the gap between flexible but slow software routers and high-throughput but rigid hardware routers. By offloading packet forwarding to programmable switches while retaining a PC for control plane functions, they achieve a balance of performance and adaptability, suitable for modern network aggregation.
What Problem Does This Paper Address?
Traditional IP routers offer high performance but are often proprietary and difficult to customize. On the other hand, open-source software routers running on commodity PCs provide flexibility but lack the raw throughput and port density required for large-scale networks like those found in carrier aggregation or data centers. This creates a performance-flexibility gap, leaving a need for a solution that combines the best of both worlds.
Methods
The core of their approach is a "hardware accelerated software router" that decouples the router's control plane (routing decisions) from its data plane (packet forwarding).
- Control Plane: A standard PC running an open-source routing suite (Quagga) handles routing protocols (BGP, OSPF, etc.) and maintains the full routing table.
- Data Plane: A programmable Ethernet switch, specifically an OpenFlow-enabled one, is used for high-speed packet forwarding.
- RouteVisor Controller: This custom software acts as the "glue" between the PC and the switch. It:
- Exposes switch ports to the PC's routing software via VLAN trunks.
- Monitors routing table updates from Quagga.
- Tracks traffic statistics on the switch.
- Implements a FIBpredict strategy to intelligently select the most frequently used destination prefixes and install them into the switch's limited FIB (Forwarding Information Base) cache.
- Fast Path/Slow Path: Packets matching an entry in the switch's FIB are forwarded directly by the switch (fast path). Packets that don't match are sent to the PC for processing (slow path) and then forwarded back to the switch.
Datasets
The authors evaluated their approach using an anonymized 48-hour packet-level trace from 20,000 residential DSL lines collected at an aggregation point within a large European ISP. This trace represents real-world carrier aggregation traffic.
Key Insights and Ideas
- Decoupling is Key: Separating control and data planes is a powerful architectural pattern for achieving both flexibility and performance.
- Programmable Switches as Accelerators: Leveraging modern programmable switches for the data plane is crucial for high-speed forwarding, even with limited FIB capacity.
- Intelligent FIB Caching is Essential: The limited FIB size on switches necessitates a smart strategy for populating it. Their FIBpredict strategy, which considers both long-term trends and recent popularity, significantly reduces FIB churn compared to simpler methods.
- Bridging the Gap: This design effectively fills the niche between high-end commercial routers and basic software routers, making advanced routing capabilities more accessible for specific network segments.
Key Results
- Reduced Slow Path Traffic: Their FIBpredict strategy, when compared to optimal and simple prefix selection methods, significantly reduces the amount of traffic that needs to be handled by the slower PC-based control plane.
- Minimized FIB Churn: FIBpredict dramatically outperforms other strategies (optimal, simple last bin, LRU, LFU) in terms of FIB modifications per second, which is critical for maintaining forwarding performance and stability. For a reasonably sized FIB cache, their strategy achieves very low slow path rates.
Limitations and Future Work
- FIB Cache Size: The performance is still constrained by the FIB cache size of the programmable switch. While their strategy optimizes its usage, a larger cache would further improve performance.
- Prototype Implementation: The prototype uses Quagga and an OpenFlow switch. Future work will explore applicability to different deployment scenarios, such as data centers.
- Security Enhancements: The authors plan to enhance RouteVisor to actively detect malicious activity and implement countermeasures against attacks.
Summary updated