[Qemu-devel] 答复: 答复: 答复: 答复: quest ion about performance of dataplane

[Zhangleiqiang]

> -----邮件原件-----
> 发件人: Abel Gordon [mailto:address@hidden
> 发送时间: 2013年4月8日 0:40
> 收件人: 张磊强
> 抄送: Luohao (brian); Haofeng; address@hidden; Stefan Hajnoczi;
> Stefan Hajnoczi; Zhangleiqiang
> 主题: Re: [Qemu-devel] 答复: 答复: 答复: question about performance of
> dataplane
> 
> 
> 
> 张磊强 <address@hidden> wrote on 07/04/2013 07:10:24 PM:
> 
> >
> > HI, Abel & Stefan:
> >
> >      After thinking twice about the benchmarks and the idea of
> > dataplane, I am still confused.
> 
> Please note while I am familiar with the documentation and architecture
> of dataplane, I didn't contribute to the dataplane code. So Stefan is
> actually the right person to answer your questions.
> 
> >      It's my understanding that the advantage of dataplane mainly
> > includes two parts. The first is that splitting the IO thread from
> > vcpu thread will avoid the global mutex competition, and the second
> > is that the individual IO thread will not be blocked when "vm exit"
> > occurs in vcpu thread.  Am I right?
> 
> As far as I understand you are right. But wait for Stefan's confirmation.

OK, I will also wait for Stefan's confirmation, ：）

> 
> >      These two advantages will always be effective whether the vcpus
> > is more than host's cores or not.
> 
> I would say that the advantage is effective as far as at least one vcpu
> thread executing guest I/O runs simultaneously with the back-end
> I/O thread (each thread on a different core).
> 
> 
> > But why the advantage of dataplane
> > is only so obvious when vcpus is more than host's cores?
> 
> Maybe the issue is related to the fact that no-dataplane experiences
> more bottlenecks when your run more and more VCPU threads. So it not
> that dataplane performs better with more VCPUs, the issue is that
> no-dataplane performs worse with more VCPUs.

I think maybe Anthony is right. In previous benchmarks, maybe the non-dataplane 
already reached the physical disk's IOPS upper limit. 

So I did another benchmark which ensures the vcpus is less than the host's 
cores, but also make continuous IO pressure by one VM when testing in the other 
VM. The result showed that dataplane did have some advantage over non-dataplane.

1. IO Pressure Mode:  8 worker, 16K IO size, 25% Read, 100% Random,  and 50 
outstanding IO
2. Benchmark Mode:  8 worker, 16K IO size, 0% Read,  100% Random,  and 50 
outstanding IO
2. Testing Results:
        a). IOPS:       178.324867 (non-dataplane)  vs  230.956328 (dataplane)
        b). MBPS:       2.786326 (non-dataplane)  vs  3.608693 (dataplane)

----------
Leiqzhang

Best Regards