Efficient Deployment of Partial Parallelized Service Function Chains in CPU+DPU-Based Heterogeneous NFV Platforms

The introduction of network function virtualization (NFV) leads to service function chain (SFC) deployment problems, promoting the idea of composing network services as virtualized network functions (VNFs). Meanwhile, the rapid development of edge computing, artificial intelligence and big data has led to a surge in data volume and explosive growth in computing and forwarding demands. As such, a traditional central processing unit (CPU)-based data forwarding mode in the NFV network appears to be a bottleneck, and a CPU-only computing framework can no longer meet the forwarding needs of diverse business scenarios and services. The data processing unit (DPU)-based architecture allows better forwarding performance to be achieved more cost-effectively, largely alleviating the computing pressure of the CPU and reducing the node forwarding delay. Therefore, in this paper, a heterogeneous CPU+DPU architecture is investigated to solve the SFC deployment problem. To handle diverse service needs, we establish a multi-objective SFC deployment scheme to optimize the service latency, deployment cost and service acceptance rate. Because extreme services require better real-time performance, DPUs are adopted for fast processing according to the requirement of service requests. To address the unacceptable delay in sequential mode, a parallel strategy is proposed to process SFCs. To solve the multi-objective SFC deployment problem, a deep reinforcement learning (DRL)-based heterogeneous algorithm that includes multiple subalgorithms is designed, named parallelizable, shared and horizontally scaled service function chain deployment (PSHD), which uses diverse processing algorithms to deploy SFCs and break the delay bottleneck in NFV-based networks.The performance of PSHD is evaluated through extensive experiments. PSHD is found to be time-efficient, and it achieves a higher request acceptance rate and 37.73% and 34.26% lower latencies than state-of-the-art methods.