[openstack-dev] 回复: [ceilometer] time based auto scaling

Luo Gangyi lgy181 at foxmail.com
Wed Apr 29 04:53:10 UTC 2015


Agree with ZhiQiang. 


Maybe we could achieve this by heat itself or other project like Mistral, 
but it seems more natural to achieve this through ceilometer alarm system.


------------------
Luo Gangyiluogangyi at chinamobile.com



 




------------------ 原始邮件 ------------------
发件人: "ZhiQiang Fan";<aji.zqfan at gmail.com>;
发送时间: 2015年4月29日(星期三) 中午12:23
收件人: "OpenStack Development Mailing List"<openstack-dev at lists.openstack.org>; 

主题: [openstack-dev] [ceilometer] time based auto scaling



Hi devs,

I'm thinking to add new type of alarm for time based auto scaling, but not sure if there is a better way to achieve it outside ceilometer scope


Currently we can auto scaling based on vm load, but it will take several minutes to do it. For the worst case, when the vm load is heavy, ceilometer needs 10 minutes (by default) to collect the performance data, and alarm need 1 minutes to evaluate it, maybe there is evaluation_periods which set to higher that 1 to avoid performance peak. 


So even though we can collect data by 1 minutes, but evaluation periods may be set to 3, so the worst case is after vm load perfomance in high level for 4 minutes, then the alarm is triggered, then heat will expand vm count, nova will take dozens seconds or more to create, finally the service on the in the heat server group will performance bad or unavailable for several minutes, which is not acceptable for some critical applications.


But if we can know some high load which related with time, for example, 08:00am will be a peak, and after 22:00pm will definitely drop to idel level, then heat can increase some vms at 07:30am, then decrease some vms at 22:00 (or decrease by load as normal routine)


However, current ceilometer only provide time constraint alarm, which will only evaluate but not guarantee it will be triggered. And heat cannot provide something like timer, but just can wait for the signal.



So I propose to add a new type of alarm, which will definitely send a signal to action when it is evaluated (with or without repeat, it will work well with time constraint)


Any suggestion?
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.openstack.org/pipermail/openstack-dev/attachments/20150429/c61ceb54/attachment.html>


More information about the OpenStack-dev mailing list