Session

IRQ suspension was merged to kernel 6.13 and provides a powerful mechanism to balance inherent trade-offs in kernel network processing.

In order to understand IRQ suspension, how it works, and when to consider using it, this talk will begin by briefly examining and discussing key prerequisite knowledge.

We’ll examine and discuss the inherent trade-offs that exist in packet processing in the kernel: increased CPU usage to get lower latency vs reduced CPU usage for higher latency, the difficulty in choosing the “right values” for an application and how, until recently, these mechanisms were only available interface-wide.

From there, per-NAPI storage will be introduced which shows how these settings, which were once interface-wide, can be set on a per-NAPI instance basis.

With this background information provided, the talk will describe the recently merged IRQ suspension mechanism, how and when an application may decide to use it, and how, exactly, to use it.

There is current on-going work to add support for IRQ suspension to an open source HTTP server. If this work is completed by the conference, the talk will conclude by mentioning the implementation, providing a link to the source as a model for others to follow, and mention any measurements taken from real world production systems running IRQ suspension.