Abstract
The efficiency of dispatching incoming network flow packets to destination virtual network functions (VNFs) is a critical bottleneck in Network Function Virtualization (NFV) platforms. While specialized hardware like SmartNICs can mitigate this, they come with high costs. We present FlowLever, a software-hardware co-designed system that repurposes the standard Flow Director (FDir) feature in commodity NICs. By implementing a hybrid dispatch model with intelligent flow activeness detection and optimized driver extensions, FlowLever achieves 1.6x to 2.8x throughput improvements at a significantly lower total cost of ownership.
Introduction: The Dispatch Bottleneck
In modern NFV environments, a single server hosts multiple VNFs. A central challenge is the "dispatcher" problem: efficiently routing incoming packets from the physical Network Interface Card (NIC) to the correct VNF. Traditional software-based dispatchers can consume up to 50% of a CPU core's processing power just for table lookups. While high-end solutions like SmartNICs or programmable SDN switches exist, their high cost and complexity often outweigh their benefits for many deployments.
FlowLever: A Hybrid Approach
FlowLever addresses this by leveraging the Flow Director (FDir) feature found in common off-the-shelf NICs (e.g., Intel XL710). FDir allows the NIC to steer packets to specific hardware queues based on flow rules. However, FDir is typically limited by small rule capacities (~8K entries) and slow reconfiguration times.
FlowLever introduces a hybrid architecture consisting of a hardware "fast path" and a software "slow path." Active flows are cached in the NIC's hardware FDir table, allowing packets to bypass software dispatch and go directly to the VNF core. New or infrequent flows are handled by a software manager (FManager), which performs lookups and monitors flow activity to decide which flows to promote to hardware.
Technical Innovations
To overcome the limitations of standard FDir, FlowLever implements several key mechanisms:
Hardware-Software Co-Design
The system uses a "Delivery Module" within the NIC to match incoming packets against a flow table. If a match (hit) occurs, the packet is sent to a specific hardware queue. If a miss occurs, it defaults to the software dispatcher.
Activeness Detection and Cache Management
Since hardware memory is scarce, FlowLever must intelligently manage the FDir cache. It employs a Least Recently Used (LRU) policy for hardware entries and a heuristic "Activity Level" (AL) calculation for software-managed flows. This ensures that only the most "valuable" flows occupy the limited hardware slots.
Asynchronous Driver Extensions
Updating hardware rules over the PCIe bus is traditionally slow. FlowLever optimizes this by modifying DPDK drivers to support asynchronous operations. A novel replace API combines rule deletion and addition into a single PCIe round-trip, reducing reconfiguration overhead significantly.
Evaluation and Results
The system was tested using real-world CAIDA UCSD internet traces. Key performance metrics include:
- Throughput: A 1.6x to 2.8x improvement over software-only platforms like OpenNetVM.
- Latency: The hardware fast path achieved ~15 μs latency, nearly half that of software dispatch (26–32 μs).
- Resource Efficiency: FlowLever can save 37.5% to 65% of CPU cores compared to software systems while maintaining equivalent throughput.
- Scaling Agility: VNF scaling and load balancing incur a performance dip of only 0.2 ms to 0.6 ms, thanks to a "Flow Group" (FGroup) abstraction that minimizes hardware updates.
Conclusion
FlowLever demonstrates that high-performance NFV dispatch does not require expensive, specialized hardware. By intelligently wrapping commodity NIC features with a sophisticated software management layer, FlowLever provides a cost-effective path to scalable, high-throughput network virtualization.
Summary updated