Re: [PATCH v12 0/6] support dirtyrate at the granualrity of vcpu

[Hyman Huang]

在 2021/10/27 14:31, Zheng Chuan 写道:
> Hi.
> I have no objection for the implement code itself.
> But we should know or let the user know the performance penalty and conflicted 
> with migration compared to the hash method, especially for the performance of 
> vm with hugepages.

i dirty guest memory with 1G and do the measurement with two method.

the copy rate is almost 1,665 MB/s in vm

the following output is guest memory performance when do measurement 
with dirty ring method:
/init (00001): INFO: 1635392998977ms copied 1 GB in 00616ms
/init (00001): INFO: 1635392999593ms copied 1 GB in 00615ms
/init (00001): INFO: 1635393000211ms copied 1 GB in 00616ms
----------------- start measurement -----------------------
/init (00001): INFO: 1635393000884ms copied 1 GB in 00672ms             
/init (00001): INFO: 1635393001849ms copied 1 GB in 00963ms     
/init (00001): INFO: 1635393002578ms copied 1 GB in 00727ms
----------------- end measurement -----------------------
/init (00001): INFO: 1635393003195ms copied 1 GB in 00615ms
/init (00001): INFO: 1635393003811ms copied 1 GB in 00614ms
/init (00001): INFO: 1635393004427ms copied 1 GB in 00615ms

guest memory performance do not trigger any changes almostly with hash 
method:

the following is test results (measurment interval=1s):

method          measurement result      copy rate during measurement
hash            44 MB/s                 1,665MB/s       
dirty ring      1167 MB/s               1,523MB/s、1,063MB/s、1,408MB/s

the max penalty is 36% during test interval(1s), the average penalty is 
20%。

if we trade off accurance, the dirty ring method may be a availiabe 
method for user. users can select a appropriate method as they need.

> On 2021/10/15 10:07, Hyman Huang wrote:
> > 在 2021/10/15 9:32, Peter Xu 写道:
> > > On Wed, Jun 30, 2021 at 12:01:17AM +0800, huangy81@chinatelecom.cn wrote:
> > > > From: Hyman Huang(黄勇) <huangy81@chinatelecom.cn>
> > > >
> > > > v12
> > > > - adjust the order of calculating dirty rate
> > > >     let memory_global_dirty_log_sync before calculating as
> > > >     v11 version description.
> > >
> > > Ping for Yong. >
> > > Dave/Juan, any plan to review/merge this series (along with the other series of
> > > dirty logging)?
> > >
> > > I found it useful when I wanted to modify the program I used to generate
> > > constant dirty workload - this series can help me to verify the change.
> > >
> > > I still keep thinking this series is something good to have.  Thanks,
> > the dirtyrate calculation has already been used to estimate time of live migration in 
> > "e cloud" production of chinatelecom, it also predict the migration success 
> > ratio, which provide valuable information for the cloud management plane when selecting 
> > which vm should be migrated.
> > >

--
Best regard

Hyman Huang(黄勇)