OpenIoT & ELC Europe 2016

We had a few changes in -RT to improve its performance. One example is the introduction of ,,lazy preempt'' which aims at avoiding ,,bad'' preemption spots which would lead to lock contention followed by another scheduling event. Lazy preempt is also an attempt to stay close to what mainline does in similar situations and the result is an improvement in performance especially in disk throughput. This talk is about the proposed changes in kernel's futex implementation which may reduce the latency in certain contended locking and unlocking path and benefits to mainline as well especially on big NUMA machines. It also covers some of the unexpected review comments on the mailinglist which led to changes in the design of addressing the problem.