Hi, guys.

> jay wrote:
> > So you are saying that the Synaps server is storing 14,400,000 samples
> > in memory (2 days of 5000 samples per minute)? Or are you saying that
> > Synaps is storing just the 5000 alarm records in memory and then
> > processing (determining if the alarm condition was met) the samples as
> > they pass through to a backend data store? I think it is the latter but
> > I just want to make sure :)
> 
> Swann wrote:
> > @jay : the first case seems to be impossible, no scalable .. I bet for 
> > the last :)

Jay and Swann, your guess is right.

Synaps holds samples in memory rolled up by 1 minute resolution in its 
sliding windows per stream. The size of sliding window is 5 minutes 
by default. It helps rolling samples up without DB read operation.

So, if there was no alarm, Synaps would hold 25,000 samples (5 
minutes of 5,000 samples per minute) in memory.

When a stream has alarms, its sliding window grows according to the 
longgest 'periods * evaluation periods + default window size' of its
alarms.

In the load test case, Synaps held 5,000 alarms and 70,000 samples 
(the recent 14 minutes of 5,000 samples) in memory as they pass 
through to a backend data store. Because the alarms had 3 minutes 
periods and 3 times of evaluation periods and default window size is 
5 minutes. (3 * 3 + 5 = 14)

Swann wrote:
> The Ceilo team will work on the improvements IIUC.
> I found two relevant links [1] [2]
> [1] https://wiki.openstack.org/wiki/Ceilometer/AlarmImprovements
> [2] 
> https://etherpad.openstack.org/p/icehouse-summit-ceilometer-future-of-alarming

Thank you for the useful links. But I just want to point out that Synaps 
has already implemented some important things in the blueprint.

Swann wrote:
> @June Yi
> I am curious to know how have you generate load to Ceilometer with 
> Ganglia ?
> 
> what was the system usage of your servers during the 2 tests  ? cpu,
> mem, io..

Ganglia was just for collecting performance data. I used my own load 
generator script. Here I attach performance data collected by ganglia. 
Please keep in mind that evaluation throughput of Ceilometer was lower 
than Synaps.

> what are response time for alarm evaluations for Ceilometer, 50 seconds 
> in mean  ?

Mean(or average) is important. But in the aspect of real-time constraint, 
I think predictability is also important. I think that there are too many 
variable 
factors in alarm evaluation in current Ceilometer to adapt it as a solution of 
'monitoring as a service'.

Best regards,
June Yi

Attachment: loadtest_result.png
Description: Binary data

_______________________________________________
OpenStack-dev mailing list
OpenStack-dev@lists.openstack.org
http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev

Reply via email to