Now we have a great load on Nova API because ceilometer-polling with “compute” namespace will polling on Nova API periodically. This proposal aims to reduce the load by abandoning the current way of metrics collection with periodically polling.
NOTE: The ceilometer-polling service with “compute” namespace is previously named as ceilometer-agent-compute.
Consider the following example. There is an environment with 500 compute nodes and every compute node deployed with ceilometer-polling with “compute” namespace (this is a real deployment scale of our company). the polling interval is set to 60s. For every polling period, ceilometer-polling will invoke “instance_get_all_by_host()” method (and maybe need querying image and flavor also) to query the instances of this host from Nova API. That means the Nova API will receive 500 requests meanwhile every minutes. It would massively increase the workload on Nova API.
In the metadata caching proposal, the instance info from Nova is be cached in ceilometer-polling by using ‘change-since’ query filter of Nova server-list API. That has significantly decreased the amount of items of every Nova query. But that cannot help to reduce the query frequency.
To illustrate this, We assume that we can deploy 100 instances on one host at most. with proposal, that may reduce 100 instances to 20 instances returned in every Nova list request, it’s not a significant improvement. But with this proposed change it may will reduce the times of Nova API request from 500 requests/min to 1 requests/min, it would be a significant improvement and that is the goal of this proposal.
This proposal try to solve the problem described above with changes:
The ceilometer-polling agent with “compute” namespace runs on every compute nodes. Currently, the agent need to get the instances list of this host by calling “instance_get_all_by_host” method of novaclient, then traverse this instance list to get metrics of instance(e.g. cpu, disk.read.bytes, network.incoming.bytes) by inspector, which will invoke the interfaces of virt layer to get this metric data. Then ceilometer-polling will assemble the metric data from virt layer and instance info from Nova API as sample objects.
Differently, this change want to add a global cache mechanism to cache the resource metadata(instance info from Nova db). ceilometer-polling will use “instance_get_all” to query all the instance from Nova, and then cache them to the global cache. Also, ceilometer-polling will record the resource caching time to the cache system, and add an option named “interval_to_update_resource” to indicate the max interval of resource cache updating. If the time from last updating is larger than this option value, the resource cache need to be updated again.
The “change-since” will also used as a query filter of “instance_get_all” to massively reduce the amount of every query and only update the changed instance to the resource cache.
In every ceilometer-polling of every compute node, unlike previous, agent will try to query instances list of this host from the global resource cache if the cached resources are not out of date, otherwise, query all the instances from Nova API to update the cached instance resources and get the instances of this host. The rest of the process will be same as before. The above “out of date” means the time from last cached resource updating is larger than the “interval_to_update_resource” option value.
- It allows users to enable/disable this global resource cache mechanism, when disable this mechanism, the process is same as before, ceilometer-polling won’t use the global cache but use a local cache implemented by proposal.
Another suggested solution for this goal is using instance events by notifications instead of instance metrics by polling to collect instance resource metadata.
Using events to update resource metadata in Ceilometer or Gnocchi will be more efficient because the resource metadata may only need to be update when something about resource happened. but I have some concerns about that approach:
Instead of the global cache directly used by compute pollsters of ceilometer-polling service, another choice is using the global cache in notification-agent service. Instances info will be polled and cached to the global cache, metric data will be polled by ceilometer-polling and sent to notification-agent, in notification-agent, the instances info will be queried from the global cache, and then assemble the instance info and metric data to sample objects, then push them to pipeline.
With this proposal, the resource metadata won’t be taken from ceilometer-polling to notification-agent, only the lightweight metric data will be pushed to notification-agent. In a large scale deployment, this will also significantly reduce the load of message bus.
Need to deploy an external global cache system, more details please see.
Documentation should be updated to add description for this.