2017-06-23 12:08 GMT+08:00 Yang Zhang <yang.zhang.wz@xxxxxxxxx>: > On 2017/6/22 19:50, Wanpeng Li wrote: >> >> 2017-06-22 19:22 GMT+08:00 root <yang.zhang.wz@xxxxxxxxx>: >>> >>> From: Yang Zhang <yang.zhang.wz@xxxxxxxxx> >>> >>> Some latency-intensive workload will see obviously performance >>> drop when running inside VM. The main reason is that the overhead >>> is amplified when running inside VM. The most cost i have seen is >>> inside idle path. >>> This patch introduces a new mechanism to poll for a while before >>> entering idle state. If schedule is needed during poll, then we >>> don't need to goes through the heavy overhead path. >>> >>> Here is the data i get when running benchmark contextswitch >>> (https://github.com/tsuna/contextswitch) >>> before patch: >>> 2000000 process context switches in 4822613801ns (2411.3ns/ctxsw) >>> after patch: >>> 2000000 process context switches in 3584098241ns (1792.0ns/ctxsw) >> >> >> If you test this after disabling the adaptive halt-polling in kvm? >> What's the performance data of w/ this patchset and w/o the adaptive >> halt-polling in kvm, and w/o this patchset and w/ the adaptive >> halt-polling in kvm? In addition, both linux and windows guests can >> get benefit as we have already done this in kvm. > > > I will provide more data in next version. But it doesn't conflict with Another case I can think of is w/ both this patchset and the adaptive halt-polling in kvm. > current halt polling inside kvm. This is just another enhancement. I didn't look close to the patchset, however, maybe there is another poll in the kvm part again sometimes if you fails the poll in the guest. In addition, the adaptive halt-polling in kvm has performance penalty when the pCPU is heavily overcommitted though there is a single_task_running() in my testing, it is hard to accurately aware whether there are other tasks waiting on the pCPU in the guest which will make it worser. Depending on vcpu_is_preempted() or steal time maybe not accurately or directly. So I'm not sure how much sense it makes by adaptive halt-polling in both guest and kvm. I prefer to just keep adaptive halt-polling in kvm(then both linux/windows or other guests can get benefit) and avoid to churn the core x86 path. Regards, Wanpeng Li
