Introduction
In modern data centers, Network Function Virtualization (NFV) has replaced traditional hardware middleboxes with software-based Virtual Network Functions (vNFs). To handle high-speed traffic, these vNFs are often offloaded from general-purpose CPUs to specialized Smart Network Interface Cards (SmartNICs). However, unpredictable traffic patterns can lead to "hot spots" where a specific vNF on a SmartNIC becomes overloaded.
The Challenge: The PCIe Bottleneck
The intuitive solution to an overloaded vNF is to migrate that specific bottlenecked function back to the CPU. However, service chains require packets to traverse vNFs in a strict sequence. Moving a middle-of-the-chain vNF to the CPU forces packets to cross the PCIe bus multiple times (SmartNIC → CPU → SmartNIC), as shown in Figure 3(b). Since each PCIe crossing adds significant latency, this "naive" approach often degrades performance for latency-sensitive applications.
The PAM Solution: Push Aside Migration
The researchers propose PAM (Push Aside Migration), a strategy that shifts focus from the bottlenecked vNF to its neighbors. Instead of moving the overloaded function, PAM identifies "border" vNFs—those already at the transition point between the SmartNIC and the CPU.
The PAM algorithm follows three key steps:
- Border vNF Identification: It targets vNFs at the hardware boundary. Migrating these does not introduce new PCIe crossings; it merely shifts the existing boundary.
- Lightweight Migration Selection: PAM selects the border vNF with the lowest capacity requirement to migrate first, ensuring minimal overhead.
- Overload Alleviation Check: The system mathematically verifies that the migration provides sufficient relief to the SmartNIC without overloading the CPU.
By "pushing" a neighbor aside, the bottlenecked vNF gains the necessary resources on the SmartNIC to maintain high performance.
Experimental Evaluation
The system was evaluated using a hardware testbed featuring:
- Hardware: A server with a Netronome Agilio CX 2×10GbE SmartNIC and dual Intel Xeon E5-2620 v2 CPUs.
- Traffic: DPDK-based generation with packet sizes from 64B to 1500B.
- Service Chain: A sequence consisting of a Firewall, Logger, Monitor, and Load Balancer.
Results
The evaluation compared PAM against the naive migration strategy.
- Latency: PAM achieved an 18% average reduction in latency compared to the naive solution. Crucially, PAM's latency remained nearly identical to the pre-overload state because it avoided redundant PCIe crossings.
- Throughput: PAM showed a slight improvement in throughput, effectively rebalancing the workload across the CPU and SmartNIC architectures.
Conclusion and Future Work
PAM demonstrates that in high-performance networking, the physical bus (PCIe) is often a more critical constraint than raw compute power. By maintaining holistic awareness of the service chain, PAM preserves low latency during traffic spikes. Future research will focus on extending PAM to FPGA-based SmartNICs and refining performance models for more complex, branching service chains.
Summary updated