On Fri, Jan 24, 2020 at 09:42:42AM -0500, Waiman Long wrote: > On 1/24/20 2:52 AM, Peter Zijlstra wrote: > > On Thu, Jan 23, 2020 at 04:33:54PM -0500, Alex Kogan wrote: > >> Let me put this question to you. What do you think the number should be? > > I think it would be very good to keep the inter-node latency below 1ms. > It is hard to guarantee that given that lock hold times can vary quite a > lot depending on the workload. What we can control is just how many > later lock waiters can jump ahead before a given waiter. We're not into this for easy. And exactly because it depends on a lot we need a lot of data. Worst case lock acquisition times directly translate into worst case IRQ-off latencies, and even the most die hard throughput oriented workloads don't like to experience multiple ticks worth of irq-off latencies. > > But to realize that we need data on the lock hold times. Specifically > > for the heavily contended locks that make CNA worth it in the first > > place. > > > > I don't see that data, so I don't see how we can argue about this let > > alone call something reasonable. > > > In essence, CNA lock is for improving throughput on NUMA machines at the > expense of increasing worst case latency. If low latency is important, Latency is _always_ important. Otherwise we'd never have put so much time and effort into fair locks to begin with. Unbounded latency sucks unconditionally. > it should be disabled. If CONFIG_PREEMPT_RT is on, > CONFIG_NUMA_AWARE_SPINLOCKS should be off. You're spouting nonsense. You cannot claim any random number is reasonable without argument.
