On Wed, 1 May 2019 at 03:31, Sean Christopherson <sean.j.christopherson@xxxxxxxxx> wrote: > > On Sun, Apr 28, 2019 at 08:54:30AM +0800, Wanpeng Li wrote: > > Hi Sean, > > On Thu, 18 Apr 2019 at 01:18, Sean Christopherson > > <sean.j.christopherson@xxxxxxxxx> wrote: > > > > > > KVM's recently introduced adaptive tuning of lapic_timer_advance_ns has > > > several critical flaws: > > [.../...] > > > > > > - TSC scaling is done on a per-vCPU basis, while the advancement value > > > is global. This issue is also present without adaptive tuning, but > > > is now more pronounced. > > > > Did you test this against overcommit scenario? Your per-vCPU variable > > can be a large number(yeah, below your 5000ns) when neighbour VMs on > > the same host consume cpu heavily, however, kvm will wast a lot of > > time to wait when the neighbour VMs are idle. My original patch > > evaluate the conservative hypervisor overhead when the first VM is > > deployed on the host. It doesn't matter whether or not the VMs on this > > host alter their workload behaviors later. Unless you tune the > > per-vCPU variable always, however, I think it will introduce more > > overhead. So Liran's patch "Consider LAPIC TSC-Deadline Timer expired > > if deadline too short" also can't depend on this. > > I didn't test it in overcommit scenarios. I wasn't aware of how the I think it should be considered. > automatic adjustments were being used in real deployments. > > The best option I can think of is to expose a vCPU's advance time to > userspace (not sure what mechanism would be best). This would allow > userspace to run a single vCPU VM with auto-tuning enabled, snapshot > the final adjusted advancment, and then update KVM's parameter to set > an explicit advancement and effectively disable auto-tuning. This step is too complex to deploy in real environment, the same as w/o auto-tuning. My auto-tuning patch evaluates the conservative hypervisor overhead when the first VM is deployed on the host, and auto-tuning it only once for the whole machine. Regards, Wanpeng Li
