Matthias,

Thanks a lot for your answer.
Yes, you win the bet :) I've used swift and currently struggling to disable collectd to make my cloud usable again! :))

I've seen this STF (Service Telemetry Framework) but it seems a little bit too complicated. I should implement an OKD cluster to monitor my openstack, isn't it too much work?
Have you tried it yourself?

If I understand correctly, with your first and main opinion you mean adding this files to my overcloud deploy command:

/usr/share/openstack-tripleo-heat-templates/environments/enable-legacy-telemetry.yaml
/usr/share/openstack-tripleo-heat-templates/environments/services/collectd.yaml

and for performance tuning I've checked this page:

Is that what you mean?
If so I should make my cloud usable again and just change GnocchiBackend to a path to a file on a shared file system (i.e. NFS) because I have 4 controller nodes, because the rest is exactly what I've done up to now.

Thanks a lot,
Khodayar

On Fri, Oct 23, 2020 at 10:01 AM Matthias Runge <mrunge@redhat.com> wrote:
On 22/10/2020 17:46, Khodayar Doustar wrote:
> Hi everybody,
>
> I am searching for a good and useful method to monitor my 40 nodes cloud.
>
> I have tried
>
> - Prometheus + Grafana (with
> https://github.com/openstack-exporter/openstack-exporter
> <https://github.com/openstack-exporter/openstack-exporter>) but it
> cannot monitor nodes load and cpu usage etc.
> and 
> - Gnocchi +Collectd + Grafana but it enforces unbelievable load on nodes
> and make the whole cloud completely unusable!
>
> I've tried to use Graphite + Grafana but I failed.
>
> Do you have any suggestions?


Hi,

yes, I have some opinions here.

My proposal here is:

- use collectd to collect low level metrics from your baremetal machines
- use ceilometer to collect OpenStack related info, like project usage,
etc. That is nothing you'd get by using node-exporter
- hook them both together and send metrics over to something called
Service Telemetry Framework. The configuration *is* included in tripleo.
The website has documentation available
https://infrawatch.github.io/documentation
- graphite + grafana (plus collectd) is also a single node setup and
won't provide you reliability.
- collectd also provides the ability to send events, which can be acted
on. That is not included if you use node-exporter, openstack-exporter
etc. Prometheus monitoring creates events from metrics, but will be slow
to detect failed components.

Since prometheus is meant to be single server, there is no HA per se in
prometheus. That makes handling prometheus on standalone machines a bit
awkward, or you'd have a infrastructure taking care of that.

In your tests with gnocchi, collectd and grafana, I bet you used swift
as backend for gnocchi storage. That is not a good idea and may lead to
bad performance.

Matthias

--
Matthias Runge <mrunge@redhat.com>

Red Hat GmbH, http://www.de.redhat.com/, Registered seat: Grasbrunn,
Commercial register: Amtsgericht Muenchen, HRB 153243,
Man.Directors: Charles Cachera, Brian Klemm, Laurie Krebs, Michael O'Neil

_______________________________________________
users mailing list
users@lists.rdoproject.org
http://lists.rdoproject.org/mailman/listinfo/users

To unsubscribe: users-unsubscribe@lists.rdoproject.org