qemu-devel
[Top][All Lists]
Advanced

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [PATCH] migration: Count new_dirty instead of real_dirty


From: zhukeqian
Subject: Re: [PATCH] migration: Count new_dirty instead of real_dirty
Date: Tue, 16 Jun 2020 09:22:29 +0800
User-agent: Mozilla/5.0 (Windows NT 10.0; WOW64; rv:45.0) Gecko/20100101 Thunderbird/45.7.1

Hi Jay Zhou,

On 2020/6/15 19:50, Zhoujian (jay) wrote:
> Hi Keqian,
> 
>> -----Original Message-----
>> From: zhukeqian
>> Sent: Monday, June 15, 2020 11:19 AM
>> To: qemu-devel@nongnu.org; qemu-arm@nongnu.org; Paolo Bonzini
>> <pbonzini@redhat.com>; Zhoujian (jay) <jianjay.zhou@huawei.com>
>> Cc: Juan Quintela <quintela@redhat.com>; Chao Fan <fanc.fnst@cn.fujitsu.com>;
>> Wanghaibin (D) <wanghaibin.wang@huawei.com>
>> Subject: Re: [PATCH] migration: Count new_dirty instead of real_dirty
>>
>> Hi Paolo and Jian Zhou,
>>
>> Do you have any suggestion on this patch?
>>
>> Thanks,
>> Keqian
>>
>> On 2020/6/1 12:02, Keqian Zhu wrote:
>>> DIRTY_LOG_INITIALLY_ALL_SET feature is on the queue. This fixs the
> 
> s/fixs/fixes
Thanks.
> 
>>> dirty rate calculation for this feature. After introducing this
>>> feature, real_dirty_pages is equal to total memory size at begining.
>>> This causing wrong dirty rate and false positive throttling.
> 
> I think it should be tested whether DIRTY_LOG_INITIALLY_ALL_SET is enabled
> in ram_init_bitmaps(maybe?) in order to be compatible with the old path.
Yeah, you are right ;-)

But after I tested old path yesterday, I found that the num_dirty_pages_period
becomes total ram size after log sync in ram_init_bitmaps. The reason is that
bitmap of ramblock is initialized to be all set, so old path counts them as 
dirty
by mistake.

This bug causes false positive throttling at the end of first ram save 
iteration,
even without our DIRTY_LOG_INITIALLY_ALL_SET feature.
> 
> Thanks,
> Jay Zhou
> 
>>>
>>> BTW, real dirty rate is not suitable and not very accurate.
>>>
>>> 1. For not suitable: We mainly concern on the relationship between
>>>    dirty rate and network bandwidth. Net increasement of dirty pages
>>>    makes more sense.
>>> 2. For not very accurate: With manual dirty log clear, some dirty pages
>>>    will be cleared during each peroid, our "real dirty rate" is less
>>>    than real "real dirty rate".
> 
I should correct these commit messages for reason above :-)
> 
> 
>>>
>>> Signed-off-by: Keqian Zhu <zhukeqian1@huawei.com>
[...]

Thanks,
Keqian



reply via email to

[Prev in Thread] Current Thread [Next in Thread]