DPDK Technical Overview, Architecture, and Performance Optimization Guide

Based on the China Telecom DPDK Technical Whitepaper v1.0, the DPDK framework consists of basic DPDK technology (standard data‑plane development kit and I/O forwarding) and optimization techniques aimed at improving forwarding performance of user applications.

Technical Principles and Architecture – Software forwarding and switching make the internal forwarding capacity of a single server the main performance bottleneck in NFV systems. To overcome this, techniques such as interrupt elimination, kernel‑bypass, reduced memory copies, multi‑core task distribution, and Intel VT are employed, with DPDK serving as a representative acceleration solution.

DPDK is an open‑source user‑space data‑plane library that bypasses the kernel protocol stack, uses poll‑mode packet I/O, optimizes memory/buffer/queue management, and supports multi‑queue NICs and flow‑based load balancing, enabling high‑performance packet forwarding on x86 platforms.

Software Architecture

The lowest layer in kernel space includes the KNI and IGB_UIO modules; KNI provides access to the Linux kernel protocol stack and tools, while IGB_UIO maps NIC registers to user space via UIO. Above this, user‑space libraries are organized into core libraries, platform modules, PMD (poll‑mode driver) modules, QoS libraries, and classification algorithms.

Core Libraries – Initialized by the Environment Abstraction Layer (EAL), they handle huge‑page allocation, memory/buffer/queue management, lock‑free operations, CPU affinity, and provide APIs for I/O bypass, memory pools, and ring buffers.

Platform Modules – Include KNI, power management, and IVSHMEM for zero‑copy sharing between VMs and the host.

Poll‑Mode Driver (PMD) Modules – Implement interrupt‑free packet I/O, supporting both physical and virtual NICs from vendors such as Intel, Cisco, Broadcom, Mellanox, and Chelsio, and work with KVM, VMware, and XEN.

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

Huge‑Page Technology – Uses 2 MiB or 1 GiB pages to reduce TLB misses; DPDK allocates all memory from hugepages, creating mempools and mbufs for packet buffers.

Polling Technique – Eliminates interrupt handling by continuously polling packet arrival flags, optionally using DDIO to store packets directly in CPU cache.

CPU Affinity – Binds DPDK threads to specific CPU cores using Linux pthreads to avoid context‑switch overhead and improve cache locality.

OS Tuning for DPDK Applications

Key configuration steps (example for CentOS 7) include:

isolCPUs=16-23,40-47 – isolate CPUs for DPDK.

nohz_full=16-23,40-47 – reduce periodic timer interrupts.

nmi_watchdog=0 – disable NMI monitoring.

selinux=0 – disable SELinux.

intel_pstate=disable nosoftlockup – lock CPU frequency.

systemctl set-default multi-user.target – disable GUI.

systemctl disable irqbalance.service – disable IRQ balancing.

systemctl disable auditd.service – disable audit.

systemctl disable bluetooth.service – disable Bluetooth.

systemctl disable ksm.service and systemctl disable ksmtuned.service – disable KSM.

Bind vCPU threads to physical CPUs via QEMU monitor and taskset .

OVS Integration – OVS 2.4+ supports DPDK acceleration, creating a PMD thread per NUMA node that processes DPDK interfaces; CPU affinity should be applied to these threads for optimal performance.

Memory Management – DPDK relies on Linux HugePages and NUMA‑aware allocation. The mempool library distributes objects across memory channels to avoid bottlenecks, while APIs such as rte_memcpy() , rte_malloc() , and NUMA‑specific memzone creation minimize latency.

NUMA Considerations – Local memory access is faster than remote; DPDK provides APIs to allocate memory on specific NUMA nodes, reducing cross‑node traffic and improving cache utilization.

Inter‑Core Lock‑Free Communication – DPDK ring API offers a lock‑free ring for message‑passing between cores, supporting batch and burst operations with minimal atomic instructions.

CPU Type Optimization – DPDK can be compiled with CONFIG_RTE_MACHINE to target specific micro‑architectures; using the latest compiler ensures support for AVX, AVX2, etc., otherwise performance may degrade.

For deeper technical details and solutions, refer to the cited DPDK whitepaper and related resources.