DPDK Technical Overview: Architecture, Core Libraries, and Performance Optimization Techniques

Based on the China Telecom DPDK Whitepaper v1.0, DPDK is divided into basic technology (the standard data‑plane development kit and I/O forwarding implementation) and optimization technology, which aims to further improve forwarding performance of user applications.

The article explains that software forwarding and switching make the internal forwarding capability of a single server the main performance bottleneck in NFV systems, describing the multiple CPU‑intensive steps such as interrupt handling, virtual I/O, kernel‑space switching, memory copying, and more.

To overcome these bottlenecks, the industry adopts techniques like massive interrupt elimination, kernel‑stack bypass, memory‑copy reduction, multi‑core CPU task distribution, and Intel VT, with DPDK serving as a typical comprehensive performance‑optimization solution.

DPDK is an open‑source user‑space library that provides high‑performance packet processing through environment abstraction layer (EAL) bypass, poll‑mode packet I/O, optimized memory/buffer/queue management, multi‑queue NIC support, and load‑balancing based on flow identification, enabling fast packet forwarding on x86 platforms.

The software architecture consists of kernel‑mode modules (KNI and IGB_UIO) and a rich set of user‑space libraries, including core libraries, platform modules, poll‑mode driver (PMD) modules, QoS libraries, and classification algorithms.

Core libraries, built on Linux via EAL, handle huge‑page memory allocation, lock‑free buffer/queue management, CPU affinity binding, and provide APIs that hide kernel and NIC I/O operations, allowing applications to avoid kernel‑stack overhead.

Platform modules include KNI for kernel‑stack protocol handling, power‑management APIs for dynamic CPU frequency scaling, and IVSHMEM for zero‑copy shared memory between VMs and the host.

PMD modules implement poll‑mode NIC drivers, eliminating interrupt‑driven latency and supporting both physical and virtual NICs from vendors such as Intel, Cisco, Broadcom, Mellanox, and Chelsio, as well as virtualization platforms like KVM, VMware, and Xen.

DPDK also defines extensive APIs for ACL, QoS, flow classification, load balancing, and hardware‑accelerated encryption/decryption extensions.

Huge‑page technology is used to allocate memory from HugePages, creating mempools and fixed‑size mbufs for each packet, thereby reducing TLB misses and improving memory access efficiency.

Polling techniques replace interrupt handling by continuously checking packet‑arrival flags, allowing packets to be stored directly in CPU cache (with DDIO) or memory, which dramatically increases processing throughput.

CPU‑affinity techniques bind threads or processes to specific cores, eliminating context‑switch overhead and cache invalidation, and DPDK leverages Linux pthreads to achieve this binding for optimal performance.

The article concludes with references to the original whitepaper and promotional information for additional architecture‑related technical resources.