DPDK Summit 2025

Writing multi-process DPDK applications using external queues" explores a novel approach to leveraging DPDK’s multi-process capabilities without requiring shared data or a designated main process. While DPDK supports offloading flows to hardware for performance gains, integrating offloading logic into existing applications can be challenging. By using APIs that enable one application to register queues with another, users can introduce a separate offloading application without modifying the original one. This approach is particularly valuable in scenarios like P4-based architectures, where the offloading application operates independently from the main application, enabling greater flexibility and modularity

With Increase of high performance data applications require data to written at very high speeds , where traditional Row based systems struggle with the scale and complexity of data. This session explores the usage of DPDK and kernel bypass techniques to massively improve the throughput of I/O operations when scaling in order of PetaBytes. Gives insights and key decisions and trade offs while building and implementing such systems. Using DPDK to enhance parallel writing of nested data in columnar formats at petabyte scale combines the framework’s high-performance ethos with the needs of modern data systems. By leveraging DPDK’s multi-threading, memory management, and hardware acceleration, this approach removes traditional bottlenecks, enabling scalable, efficient data storage for applications from scientific research to big data analytics. As data volumes grow, acceleration techniques are very key to improve efficiency and scaling the system.

In this tech talk, we benchmark connection tracking implementations by extending rte_hash based tables with advanced flow aging mechanisms. We compare aging algorithms like timer wheels, smart periodic scanning vs. LRU tables.
We define test cases based on telco traffic profiles using T-Rex traffic-gen and interpret the benchmark results.
Join us to gain insights into enhancing DPDK rte_hash with flow aging features.

There have been changes to ARM build system to ensure users get the full performance possible from their build platforms. At the same time, care has been taken to ensure good out of box experience. This talk will discuss ARM ISA revisions, how to create binaries for portability and for performance. It will also discuss typical issues users and developers might face and how to resolve them.

I have presented the project in Montreal last September and it received positive reactions from the DPDK community.

Since then, the project has received significant improvements and is now used internally at Red Hat for validating Telco infrastructure (for both OpenStack and OpenShift).

I would like to give a heads up to anybody who couldn't make it to Montreal. Also, I would like to challenge developers of core libraries about changes that would be beneficial to grout and DPDK.

The argparse library provides argument parsing functionality, it was integrated in DPDK 24.03, and makes it easy to write user-friendly command-line program. This topic will introduce this library usage.

DPDK's collaborative development model continues to deliver an open and inclusive community, but it isn't without its challenges. Developers, eager to deliver new features as quickly as they can, will submit frequent versions of patches, including large series and large patches. A CI failure occuring in one patch leads to a quick respin and submit cycle. The result of this turnaround is an unintended effect: lots of wasted CI cycles, incomplete reviews, and a slower merging process.

In this talk, we'll explore the recent efforts to introduce intentional delays into the process to help upstream become more efficient. We will explain how small changes such as smarter individual testing, delayed submits, and improved communication can lead to a more sustainable and productive open-source development model.

If you've ever been frustrated waiting for the next CI result, or feel like the mailing list has become a bit 'write-only', this talk may be for you. Let's patch smarter, not harder!