Collaboration Summit 2014

Database, high-performance computing (HPC), and real-time developers have often asked: "Can't you get the kernel out of the way?". Recent adaptive-idle work permits just that: Linux is there when you need it, but if you follow a few simple rules, it is out of your way otherwise. This approach will provide bare-metal multicore performance and scalability to databases as well as to HPC and real-time applications. However, timekeeping requires that at least one CPU continue in high-power mode if any non-idle execution is in flight. Unfortunately, simple code to determine if all CPUs are idle is not scalable. This talk will give an overview of adaptive idle, outline some of the work to scalably determine whether the timekeeping CPU can go into low-power mode while avoiding any embarrassing time-skew incidents, and discuss some timer-related issues.

This presentation will help developers of the kernel and of HPC or real-time applications understand, better use, or improve the kernel's recent NO_HZ_FULL functionality.

The audience is anyone interested in getting bare-metal performance from the Linux kernel, for example, those running high-performance computing (HPC) or real-time workloads.