DPDK Summit 2026

In this presentation, we will explore the application of header-data split in zero-copy mechanisms within a disaggregated userspace protocol stack to improve performance and multi-tenant memory safety. We focus on two key points:
1. Performance Optimization
a. Zero-Copy: Achieve zero-copy between the application and the NIC, reducing latency and saving CPU resources.
b. Cache Efficiency: By applying HDS, the userspace stack only loads the protocol headers, reducing cache pollution and making protocol parsing faster.
c. Cross-Socket Access: Headers and payloads can reside on different sockets, eliminating the need for remote data access by applications.
2. Memory Permission Isolation
a. Memory Sharing and Safety: Each application creates memory pool and shares it with the userspace stack. Only read permission is retained in userspace stack after registration on device.
b. Fault Containment and Protection: In the event of memory exceptions in either the application or the userspace stack, their operational stability remain unaffected.
Through detailed design, we have enhanced the performance and safety. Attendees will get an in-depth look at the practical application.

Overview of GFNI based Toeplitz hash implementation

Physical AI systems demand ultra‑low‑latency transport of rich sensor data from the edge to centralized AI compute. This talk showcases a DPDK‑based sensor bridge using NXP i.MX processors as NVIDIA Holoscan Sensor Bridge endpoints, delivering deterministic, zero‑copy perception pipelines. It highlights how DPDK unlocks a new class of real‑time workloads for robotics, industrial vision, and Physical AI.
 

DPDK library supports cryptodev which mainly deals with static keys. It does not have support for any dynamic internet key exchange. In this talk, we present an approach to integrate DPDK enabled cloud native router with opensource strongswan to support IKE for IPSec. A custom packet processing pipeline will also be presented where dedicated encryption/decryption cores are reserved and packets are steered to it depending on the routing policy.

Post Quantum Cryptography (PQC) requires seamless integration into modern packet frameworks. This work introduces post-quantum cryptography (PQC) support in the Data Plane Development Kit (DPDK) through the OpenSSL Poll Mode Driver (PMD), providing developers with a practical entry point to experiment with PQC primitives using the familiar cryptodev API. The OpenSSL PMD implementation serves as a reference design, demonstrating how PQC algorithms such as key encapsulation and digital signatures can be exposed consistently to applications with minimal changes. For developers, this enables immediate validation, benchmarking, and integration of PQC workloads in software, while establishing a clear API model for future extensions. The long-term vision is for hardware crypto accelerators—including Octeon CNXK and VirtIO crypto drivers—to adopt this reference approach, ensuring interoperability and contributing back to the community. By aligning software and hardware PMDs around a unified PQC API, DPDK fosters a collaborative ecosystem where developers, vendors, and operators can build scalable, quantum resilient applications such as VPNs, TLS termination, IPsec gateways, and 5G cores.

DPDK is used as the packet-ingest layer for a traffic monitoring system feeding an internal DPI engine.
Today, throughput is maximized by tuning for a specific server/NIC/OS combination, but this reduces portability and increases operational effort.

This lightning talk discusses the throughput impact observed when moving toward a more “generic” deployment.
Results are presented for: (1) optimized (“native”) vs portable (“generic”) builds, (2) CPU isolation, (3) hugepage configuration and (4) PMD/driver configuration.

In the field of high-performance networking, traditional packet capture tools like tcpdump are facing a massive blind spot in observability. tcpdump relies heavily on the kernel protocol stack, whereas modern high-performance applications widely adopt DPDK technology. In DPDK, packets flow directly between userspace drivers and applications via mbuf, completely bypassing the kernel.

To address these pain points, we have introduced an internal open-source network observability tool — netcap.

Netcap not only inherits the ability to trace kernel skbs but also achieves breakthrough support for DPDK mbufs:
- Deep Userspace Introspection: It can delve directly into userspace memory, non-intrusively capturing and parsing mbuf data flowing through critical DPDK processing functions.
- Zero Learning Curve: It perfectly supports tcpdump filter syntax (BPF filters), allowing developers to precisely capture the mbuf traffic they care about, just like using tcpdump.

The emergence of netcap effectively fills the gap in "mbuf observability" within high-performance network development, making black-box debugging of DPDK applications a thing of the past.

Doing packet capture in a DPDK application leads to several technical issues. Packet capture needs to be fast, transparent and be easy to use. This talk will cover the different solutions and how they can be improved in the future.

Events (jobs) in Eventdev are non-preemptive, and thus must be relatively short run time.

One way to deal with this issue is to split large jobs into multiple events.

For legacy code, originally design for POSIX threads, such an exercise may be costly. Also, for certain domains, being able to retain the stack between invocations (events) may results in a simpler program.

This talk briefly introduce this topic and discuss an Ericsson prototype coroutine library running on top of Eventdev, allowing for a code written for a thread-like, cooperative multi-tasking programming model to coexist with native event-driven programs running on the same DPDK lcores, making the coroutine library in combination the event device a task/thread scheduler.

Note: I have not reviewed the two-hour "Inclusive Speaker Orientation" training. If this is mandatory, please ignore this submission.