From mboxrd@z Thu Jan  1 00:00:00 1970
From: "Longpeng (Mike)" <longpeng2@huawei.com>
Subject: Re: [PATCH 3/3] KVM: Add capability to not exit on PAUSE
Date: Wed, 29 Nov 2017 12:34:43 +0800
Message-ID: <5A1E38E3.2050802@huawei.com>
References: <1511615373-32615-1-git-send-email-jschoenh@amazon.de> <1511615373-32615-4-git-send-email-jschoenh@amazon.de> <5A1CD9DC.7050804@huawei.com> <b767cd12-9ea0-f969-5b88-d03066158403@amazon.de>
Mime-Version: 1.0
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: 8bit
Cc: Paolo Bonzini <pbonzini@redhat.com>,
        =?UTF-8?B?UmFkaW0gS3LEjW3DocWZ?= <rkrcmar@redhat.com>,
        Joerg Roedel <joro@8bytes.org>,
        "Michael S. Tsirkin" <mst@redhat.com>,
        KarimAllah Ahmed <karahmed@amazon.de>, <kvm@vger.kernel.org>
To: =?UTF-8?B?IkphbiBILiBTY2jDtm5oZXJyIg==?= <jschoenh@amazon.de>
Return-path: <kvm-owner@vger.kernel.org>
Received: from szxga06-in.huawei.com ([45.249.212.32]:38318 "EHLO huawei.com"
        rhost-flags-OK-FAIL-OK-FAIL) by vger.kernel.org with ESMTP
        id S1751908AbdK2EfQ (ORCPT <rfc822;kvm@vger.kernel.org>);
        Tue, 28 Nov 2017 23:35:16 -0500
In-Reply-To: <b767cd12-9ea0-f969-5b88-d03066158403@amazon.de>
Sender: kvm-owner@vger.kernel.org
List-ID: <kvm.vger.kernel.org>



On 2017/11/29 8:09, Jan H. Schönherr wrote:

> On 11/28/2017 04:37 AM, Longpeng (Mike) wrote:
>>
>> On 2017/11/25 21:09, Jan H. Schönherr wrote:
>>
>>> Allow to disable pause loop exit/pause filtering on a per VM basis.
>>>
>>> If some VMs have dedicated host CPUs, they won't be negatively affected
>>> due to needlessly intercepted PAUSE instructions.
>>>
>> Hi Jan,
>>
>> Is there any difference between 'disable PLE in vmcs' and 'make ple_gap per
>> VM/VCPU and set ple_gap=0 for vcpus which is dedicated' ?
> 
> "Just" disabling PLE in vmcs would still call into some of the PLE window
> adjustment paths and potentially do some VMCS writes at times. My patch should
> have eliminated these cases as well.
> 

Ah, I see, thanks. :)

We used a proprietary test suite to benchmark performance with ple_gap=0 and
ple_gap=1 (very little VMexits due to PLE in kvm trace), we found that ple_gap=1
is always better than ple_gap=0, we don't know why, maybe impacts the hardware
logical. Do you have any idea about this ?

> However, making all the PLE configuration knobs per VM has the difficulty
> that you'd need to be able to specify them in some way. That would not only
> be a x86 specific interface, but a VMX specific one as well. VMX-PLE and
> SVM-PF don't look compatible enough for a shared configuration.
> 
> That's why I only went for the binary on/off interface.
> 

Yes, I think your patch is pretty good.

> Regards
> Jan
> 
> 
> 
> .
> 


-- 
Regards,
Longpeng(Mike)