Network Performance in the Linux Kernel, Getting the most out of the Hardware
HTML-код
- Опубликовано: 29 июн 2024
- Video from Bootlin engineer Maxime Chevallier's talk "Network Performance in the Linux Kernel, Getting the most out of the Hardware" given at FOSDEM 2021.
Slides: bootlin.com/pub/conferences/2...
Check out Bootlin training courses at bootlin.com/training/ and Bootlin embedded Linux engineering services at bootlin.com/engineering/.
Abstract of the talk:
The networking stack is one of the most complex and optimized subsystems in the Linux kernel, and for a good reason. Between the wild range of applications, the complexity and variety of the networking hardware, getting good performance while keeping the stack easily usable from userspace has been a long-standing challenge.
Nowadays, complex Network Interface Controllers (NICs) can be found even on small embedded systems, bringing powerful features that were previously found only in the server world closest to day to day users.
This is a good opportunity to dive into the Linux Networking stack, to discover what is used to make networking as fast as possible, both by using all the features of the hardware and by implementing some clever software tricks.
In this talk, we cover these various techniques, ranging from simple batch processing with NAPI, queue management with RSS, RPS, XPS and so on, flow steering and filtering with ethool and TC, to finish with the newest big change that is XDP.
We dive into these various techniques and see how to configure them to squeeze the most out of your hardware, and discover that what was previously in the realm of datacenters and huge computers can now also be applied to embedded linux development.
Speaker biography:
bootlin.com/company/staff/max... 
Наука
Excellent. Thanks !
Excellent information into Kernel Network Architecture
Great presentation. Really usefull information. Thanks you.
Really excellent talk
Thanks. Very useful.
Awesome video. 👌
Was super helpful to discover the feature and test it on Mellanox cards. By the end the aRFS code for Mellanox cards is very dangerous and leads to a massive ksoftirq storm if too many connections occurs. The race between creating and deleted rules takes too much time and prevents processing packets. Systems dies of it.
Excellent video! Thank you very much for posting this. Can you recommend any books that also discuss this topic? Thanks.
Thanks for the in depth explanation. A few questions.
What would be the preferred Preemption Model kernel config used in order to achieve maximum performance in a communication server?
Also, will the PREEMPT_RT patch be useful in achieving better performance in the case of a communication server?
Thanks!