Re: [Qemu-devel] Abnormal observation during migration: too many "write-not-dirty" pages

[Chunguang Li]

> -----Original Messages-----
> From: "Dr. David Alan Gilbert" <address@hidden>
> Sent Time: 2017-11-15 22:23:52 (Wednesday)
> To: "Chunguang Li" <address@hidden>
> Cc: address@hidden, address@hidden, address@hidden, address@hidden, 
> address@hidden
> Subject: Re: [Qemu-devel] Abnormal observation during migration: too many 
> "write-not-dirty" pages
> 
> * Chunguang Li (address@hidden) wrote:
> > 
> > 
> > 
> > > -----原始邮件-----
> > > 发件人: "Dr. David Alan Gilbert" <address@hidden>
> > > 发送时间: 2017-11-15 18:11:37 (星期三)
> > > 收件人: "Chunguang Li" <address@hidden>
> > > 抄送: address@hidden, address@hidden, address@hidden, address@hidden, 
> > > address@hidden
> > > 主题: Re: [Qemu-devel] Abnormal observation during migration: too many 
> > > "write-not-dirty" pages
> > > 
> > > * Chunguang Li (address@hidden) wrote:
> > > > Hi all!
> > > > 
> > > > I got a very abnormal observation for the VM migration. I found that 
> > > > many pages marked as dirty during migration are "not really dirty", 
> > > > which is, their content are the same as the old version.
> > > > 
> > > > 
> > > > 
> > > > 
> > > > I did the migration experiment like this:
> > > > 
> > > > During the setup phase of migration, first I suspended the VM. Then I 
> > > > copied all the pages within the guest physical address space to a 
> > > > memory buffer as large as the guest memory size. After that, the dirty 
> > > > tracking began and I resumed the VM. Besides, at the end
> > > > of each iteration, I also suspended the VM temporarily. During the 
> > > > suspension, I compared the content of all the pages marked as dirty in 
> > > > this iteration byte-by-byte with their former copies inside the buffer. 
> > > > If the content of one page was the same as its former copy, I recorded 
> > > > it as a "write-not-dirty" page (the page is written exactly with the 
> > > > same content as the old version). Otherwise, I replaced this page in 
> > > > the buffer with the new content, for the possible comparison in the 
> > > > future. After the reset of the dirty bitmap, I resumed the VM. Thus, I 
> > > > obtain the proportion of the write-not-dirty pages within all the pages 
> > > > marked as dirty for each pre-copy iteration.
> > > > 
> > > > I repeated this experiment with 15 workloads, which are 11 CPU2006 
> > > > benchmarks, Memcached server, kernel compilation, playing a video, and 
> > > > an idle VM. The CPU2006 benchmarks and Memcached are write-intensive 
> > > > workloads. So almost all of them did not converge to stop-copy.
> > > > 
> > > > 
> > > > 
> > > > 
> > > > Startlingly, the proportions of the write-not-dirty pages are quite 
> > > > high. Memcached and three CPU2006 benchmarks(zeusmp, mcf and bzip2) 
> > > > have the most high proportions. Their proportions of the 
> > > > write-not-dirty pages within all the dirty pages are as high as 
> > > > 45%-80%. The proportions of the other workloads are about 5%-20%, which 
> > > > are also abnormal. According to my intuition, the proportion of 
> > > > write-not-dirty pages should be far less than these numbers. I think it 
> > > > should be quite a particular case that one page is written with exactly 
> > > > the same content as the former data.
> > > > 
> > > > Besides, the zero pages are not counted for all the results. Because I 
> > > > think codes like memset() may write large area of pages to zero pages, 
> > > > which are already zero pages before.
> > > > 
> > > > 
> > > > 
> > > > 
> > > > I excluded some possible unknown reasons with the machine hardware, 
> > > > because I repeated the experiments with two sets of different machines. 
> > > > Then I guessed it might be related with the huge page feature. However, 
> > > > the result was the same when I turned the huge page feature off in the 
> > > > OS.
> > > > 
> > > > 
> > > > 
> > > > 
> > > > Now there are only two possible reasons in my opinion. 
> > > > 
> > > > First, there is some bugs in the KVM kernel dirty tracking mechanism. 
> > > > It may mark some pages that do not receive write request as dirty.
> > > > 
> > > > Second, there is some bugs in the OS running inside the VM. It may 
> > > > issue some unnecessary write requests.
> > > > 
> > > > 
> > > > What do you think about this abnormal phenomenon? Any advice or 
> > > > possible reasons or even guesses? I appreciate any responses, because 
> > > > it has confused me for a long time. Thank you.
> > > 
> > > Wasn't it you who pointed out last year the other possibility? - The
> > > problem of false positives due to sync'ing the whole of memory and then
> > > writing the data out, but some of the dirty pages were already written?
> > > 
> > > Dave
> > 
> > Yes, you remember that!
> 
> Yes, I remember that, and my TODO list told me it was you :-)
> 
> > It was me. After that, I did more analysis and experiments. I found that, 
> > in fact, both reasons contribute to the "fake dirty" pages (dirty pages 
> > that do not need to be resent, because their contents are the same as that 
> > in the target node). One is what I pointed out last year, which you have 
> > mentioned. The other reason is what I am talking about now, the 
> > "write-not-dirty" phenomenon.
> > In fact, according to my experiments results, the "wirte-not-dirty" is the 
> > main reason resulting to the "fake dirty" pages, while sync'ing the whole 
> > of memory contributes less.
> 
> How do you differentiate between "fake dirty' and the syncing?

I record the following two kinds of proportions:
(1): The proportion of "write-not-dirty" pages. This is done by the method I 
have described. I copied the whole VM memory in migration setup, and at the end 
of the first iteration, compared the pages marked as dirty to their copies in 
the buffer. Thus I got the proportion of "write-not-dirty" pages for the first 
iteration.

(2): Besides, I also got the following proportion, which is the "fake dirty". 
In the first iteration, when each page was sent to the target, I copied it to a 
buffer. Note that, the contents of these pages in the buffer may be different 
from that in (1). Buffer in (1) records the pages content before migration, and 
buffer in (2) records the content sent to target. Then during the second 
iteration, when each dirty pages was sent to target, I compared it with its 
copy in buffer. Thus I got the proportion of "fake dirty" pages sent in the 
second iteration.

The number of (2) is bigger than (1). I attribute this part to the syncing.

> 
> The cases where values change back to what they used to be seem
> the most likely to me (e.g. locks/counts that decrement back) - but
> that seems a high %.
> I wonder if there's any difference between page write protection
> based dirtying and PML (that I think can be used on some newer chips).

I think PML may only reduce the overhead for tracing dirty, but which pages are 
marked as dirty should be the same as page write protection. However, I am not 
sure for this.

Chunguang

> 
> One way to debug it I guess would be to keep the write protection
> and watch the progression of data values within a page - do they
> actually change and then change back or do the values never
> really change.
> 
> Dave
> 
> 
> > Chunguang
> > > 
> > > > 
> > > > --
> > > > Chunguang Li, Ph.D. Candidate
> > > > Wuhan National Laboratory for Optoelectronics (WNLO)
> > > > Huazhong University of Science & Technology (HUST)
> > > > Wuhan, Hubei Prov., China
> > > > 
> > > --
> > > Dr. David Alan Gilbert / address@hidden / Manchester, UK
> > 
> > 
> > 
> --
> Dr. David Alan Gilbert / address@hidden / Manchester, UK


--
Chunguang Li, Ph.D. Candidate
Wuhan National Laboratory for Optoelectronics (WNLO)
Huazhong University of Science & Technology (HUST)
Wuhan, Hubei Prov., China