1.2: VPP Architecture

Summary

The transcript discusses the architecture and features of VPP (Vector Packet Processing), focusing on its flexible and efficient packet processing capabilities. It covers the integration of various input drivers, such as DPDK and TUN/TAP, and explains the use of native buffer structures for seamless data handling. The plugin infrastructure allows for dynamic feature addition, while VPP's Binary API provides fast programmatic control and debugging. Additionally, the system supports multi-core scalability and NUMA awareness, optimizing performance for high-speed networking environments. The talk emphasizes VPP's adaptability, speed, and efficiency in packet processing.

Takeaways

• 😀 VPP supports multiple packet input drivers, such as DPDK, TAP, Vhost-user, and Netmap, allowing flexible packet handling.
• 😀 VPP dynamically switches from interrupt mode to polling mode when packet traffic increases above a threshold to optimize performance.
• 😀 VPP uses its own native buffer structure, the VPP buffer (VB), which is integrated into DPDK buffers when using DPDK.
• 😀 Plugins in VPP are implemented as shared libraries, allowing users to add custom functionality without affecting core performance.
• 😀 VPP's plugin infrastructure enables subscribing to specific traffic types (e.g., UDP packets) and processing them via the VPP graph system.
• 😀 Modifying existing VPP graph nodes or adding new custom FIB types requires changes to the core VPP code, which isn't possible via plugins alone.
• 😀 VPP uses the Binary API for fast programmatic interactions and supports language bindings for C, Java, and Python.
• 😀 VPP can handle over a million routes per second via its Binary API, which is designed for speed and efficiency.
• 😀 VPP supports CLI for debugging but encourages using the Binary API for real-world interactions due to its superior performance.
• 😀 VPP's multi-threading architecture relies on embarrassing parallelism, where packets remain on the same core throughout their processing path.
• 😀 VPP is NUMA-aware, ensuring that buffers are allocated on the same CPU socket as the processing threads to optimize memory access and performance.

Q & A

What is VPP's approach to packet input processing?

VPP supports multiple input methods, including DPDK, Virtio, Netmap, and more. Each method handles packet input either in polling mode or interrupt mode, allowing dynamic switching between these modes based on traffic levels to optimize CPU usage.

How does VPP manage packet buffers when using DPDK?

When VPP is used with DPDK, it utilizes DPDK's native memory buffer structure (mbuf) but adds its own metadata, effectively integrating the VPP buffer structure (V buffer) within the DPDK mbuf. This integration avoids conversion costs between different buffer structures.

Can plugins in VPP modify existing graph nodes?

No, plugins cannot modify existing graph nodes in VPP. However, they can create new graph nodes, adding custom functionality such as packet handling or encapsulation, without modifying the core nodes in the system.

How does VPP ensure performance when using plugins?

VPP's plugin system is designed to maintain performance by treating plugins like native code. There is no performance penalty when using plugins as long as they do not alter existing graph node behaviors. Plugins are loaded dynamically and can register to handle specific packet types.

What is the role of the Binary API in VPP?

The Binary API in VPP allows external applications to interact with VPP programmatically using a fast and efficient shared memory interface. It supports operations like route addition or removal and is crucial for handling high throughput in data plane applications.

Why is VPP's Binary API considered fast?

The Binary API in VPP is fast because it is designed to be lightweight, running over shared memory. It can process over a million routes per second, making it suitable for high-performance networking applications.

How does VPP handle multi-core and parallelism?

VPP uses an approach known as 'embarrassing parallelism' where packet processing remains on the same CPU core throughout its lifecycle. While VPP processes packets on different cores, it avoids moving packets between cores during processing, ensuring better performance and minimizing latency.

What is NUMA awareness in VPP, and why is it important?

NUMA (Non-Uniform Memory Access) awareness in VPP ensures that memory allocation is optimized for the CPU socket where VPP is running. This minimizes cross-socket memory accesses, reducing latency and improving overall performance in systems with multiple CPU sockets.

Can VPP be used in commercial environments, and what licensing does it use?

Yes, VPP can be used in commercial environments. The licensing allows for the creation of binary plugins that can be distributed and sold as value-added features, enabling businesses to monetize their custom VPP functionalities.

How does VPP manage traffic distribution across multiple cores?

VPP uses RSS (Receive Side Scaling) to distribute network traffic across multiple cores. By hashing incoming packet data, it ensures an even distribution of traffic to different cores, enhancing scalability and performance in multi-core systems.