Re: [PATCH 0/15] KVM: optimize for MMIO handled

[Xiao Guangrong <xiaoguangrong@cn.fujitsu.com>]

On 06/08/2011 11:25 AM, Xiao Guangrong wrote:
> On 06/08/2011 11:11 AM, Takuya Yoshikawa wrote:
>> On Tue, 07 Jun 2011 20:58:06 +0800
>> Xiao Guangrong <xiaoguangrong@cn.fujitsu.com> wrote:
>>
>>> The performance test result:
>>>
>>> Netperf (TCP_RR):
>>> ===========================
>>> ept is enabled:
>>>
>>>       Before         After
>>> 1st   709.58         734.60
>>> 2nd   715.40         723.75
>>> 3rd   713.45         724.22
>>>
>>> ept=0 bypass_guest_pf=0:
>>>
>>>       Before         After
>>> 1st   706.10         709.63
>>> 2nd   709.38         715.80
>>> 3rd   695.90         710.70
>>>
>>
>> In what condition, does TCP_RR perform so bad?
>>
>> On 1Gbps network, directly connecting two Intel servers,
>> I got 20 times better result before.
>>
>> Even when I used a KVM guest as the netperf client,
>> I got more than 10 times better result.
>>
> 
> Um, which case did you test? ept = 1 or ept=0 bypass_guest_pf=0 or both?
> 
>> Could you tell me a bit more details of your test?
>>
> 
> Sure, KVM guest is the client, and it uses e1000 NIC, and uses NAT
> network connect to the netperf server, the bandwidth of our network
> is 100M.
> 

And this is my test script:

#!/bin/sh

echo 3 > /proc/sys/vm/drop_caches
./netperf -H $HOST_NAME -p $PORT -t TCP_RR -l 60