With even small devices having multi-core processors and NUMA configuration being the norm in the server space, reducing cross-CPU traffic is one of the major optimization points, and per-cpu data structures are very effective and well-established way to achieve it. The Linux kernel has extensive per-cpu facilities which are still evolving. This presentation looks briefly at the evolution of per-cpu facilities in the kernel and goes through several examples of per-cpu constructs and their usages.
Tejun has been working on various aspects of Linux kernel since 2005 and is currently maintaining percpu memory allocator, control groups, and workqueue. He currently works as a software engineer for Red Hat.
Wednesday September 18, 2013 11:50am - 12:40pm