On 14/12/2017 13:06, Quan Xu wrote: > > > On 2017/12/14 19:56, Paolo Bonzini wrote: >> On 13/12/2017 17:28, Konrad Rzeszutek Wilk wrote: >>> 1) VM idle path and network req/resp services: >>> >>> Does this go away if you don't hit the idle path? Meaning if you >>> loop without hitting HLT/MWAIT? I am assuming the issue you are facing >>> is the latency - that is first time the guest comes from HLT and >>> responds to the packet the latency is much higher than without? >>> >>> And the arming of the timer? >>> 2) process context switches. >>> >>> Is that related to the 1)? That is the 'schedule' call and the process >>> going to sleep waiting for an interrupt or timer? >>> >>> This all sounds like issues with low-CPU usage workloads where you >>> need low latency responses? >> Even high-CPU usage, as long as there is a small idle time. The cost of >> setting the TSC deadline timer twice is about 3000 cycles. >> >> However, I think Amazon's approach of not intercepting HLT/MWAIT/PAUSE >> can recover most of the performance and it's way less intrusive. > > Paolo, could you share the Amazon's patch or the LML link? thanks. Here: https://www.spinics.net/lists/kvm/msg159356.html Paolo
