Traditionally apps have two choices: 1) trade increased CPU usage to get lower latency or 2) lower CPU usage at increased latency. The available choices were onerous because these mechanisms were only available interface-wide - that is until this work from Joe.
In this talk, Joe Damato will discuss the IRQ Suspension feature which was recently merged in 6.13. The IRQ Suspension feature adds per-NAPI storage instances which enables more fine-grained tuning to the provided choices. Joe will describe the challenges overcame and explain when an application may decide to use the feature and how, exactly, to use it. Further discussion on current on-going work and possibly world production measurements, if available at time of the talk, will be shared.
More details: https://netdevconf.info/0x19/4
cheers, jamal