ping... Thanks, Chao Fan
On Tue, Dec 06, 2016 at 04:52:11PM +0800, Chao Fan wrote: >Hi all, > >Here is an issue in auto-converge feature of migration. > >When migrating a guest which consumes too much CPU & memory, dirty >pages amount will increase significantly, so does the migration >time, migration can not even complete, at worst. > >I did some simple tests on this feature. Set the two parameters >the same as 10,20,30,40,50,60,70,80,99 and run the same task in the >same guest. The result roughly is, with the increment of the >two parameters, the total_time and the dirty_sync_count will decrease. >Result shows larger the value of the two parameters is, faster the >migration is, but much more slowly the guest runs. > >So I think there should be a appropriate throttle value according to >the workload of guest. But users do not know how to determine the >appropriate value. > >So I want to do a job that qemu can set the throttle value according >to the workload of guest. I think qemu could calculate the instant >dirty pages rate, and then determine a appropriate throttle value. >The instant dirty pages rate means in a short fixed time, how >many dirty pages born. But I have two questions: >1. Where to add this feature. I have two options: > a. Now qemu detects the rest migration time and decides whether > to execute the CPU throttle. It can be changed to that qemu > executes the CPU throttle when instant dirty pages rate increases > to a certain threshold and sets the throttle value according to > the instant dirty pages rate. > b. Using the current way as it is, when the rest migration time > is too long and begin to execute the CPU throttle, assign > appropriate throttle value according to the workload. Codes > will be changed fewer in this method. >2. How to determine the CPU throttle value according to the dirty pages. > My preliminary idea is, the CPU throttle should be related to > the instant dirty pages rate and the total memory. > But I am not sure how to do the map from instant dirty pages rate > and total memory to CPU throttle value is best. > >Any comments will be welcome, and I want to know whether more people >think this feature is needed. >If anyone has good ideas, please tell me. > >Thanks, >Chao Fan