> From: Paolo Bonzini <pbonzini@xxxxxxxxxx> > Sent: Friday, August 26, 2022 6:14 PM > To: Mi, Dapeng1 <dapeng1.mi@xxxxxxxxx>; Christopherson,, Sean > <seanjc@xxxxxxxxxx> > Cc: rafael@xxxxxxxxxx; daniel.lezcano@xxxxxxxxxx; linux-pm@xxxxxxxxxxxxxxx; > linux-kernel@xxxxxxxxxxxxxxx; kvm@xxxxxxxxxxxxxxx; > zhenyuw@xxxxxxxxxxxxxxx > Subject: Re: [PATCH] KVM: x86: use TPAUSE to replace PAUSE in halt polling > > On 8/25/22 13:31, Mi, Dapeng1 wrote: > >> I say "if", because I think this needs to come with performance > >> numbers to show the impact on guest latency so that KVM and its users > >> can make an informed decision. > >> And if it's unlikely that anyone will ever want to enable TPAUSE for > >> halt polling, then it's not worth the extra complexity in KVM. > > I ever run two scheduling related benchmarks, hackbench and schbench, I > didn't see there are obvious performance impact. > > > > Here are the hackbench and schbench data on Intel ADL platform. > > Can you confirm (using debugfs for example) that halt polling is used while > hackbench is running, and not used while it is not running? Sorry, I may not describe the test case clearly. The hackbench and schbench are run on Host rather than a VM. When the hackbench or schbench is run on Host, there is a FIO workload running in a VM in the background and the FIO would trigger a large number of HLT VM-exits and eventually invoke halt polling. In this test, I want to check whether potential polling time extending would increase the scheduling latency on host. But it looks the impact for scheduling latency is quite minimal. > > In particular, I think you need to run the server and client on different VMs, > for example using netperf's UDP_RR test. With hackbench the ping-pong is > simply between two tasks on the same CPU, and the hypervisor is not > exercised at all. > Here are the netperf's UDP_RR test result between two VMs locate on two different physical machines. Netperf Vanilla (Avg.) TPAUSE (Avg.) %Delta UDP_RR (Trans. Rate/s) 503.8 503.9 0.02% It looks there is no obvious difference with TPAUSE change on UDP RR test. > Paolo
