In general Cloud 66 services will consume resources and spawn containers as needed, but you may want to explicitly limit a particularly resource-hungry service, or ensure that a service is only assigned (scheduled) to a node with sufficient resources to handle its needs. You can also set a minimum or maximum number of containers a service may use.
Cloud 66 supports four kinds of resource allocations:
- Number of containers
- Node allocation
All of these service configurations are defined in your application's
service.yml file. Read our guide to using service.yml for more help on customizing your service configuration.
Understanding resource management options
When setting resource parameters, it's vital to understand the differences between your options.
- Limits act like a ceiling on the resource usage for any service. They tell your cluster that the service may only use a maximum of the specified resource.
- Requests act like a floor on the resources for a service. They tell your cluster "this service needs X" in order to run. Cloud 66 uses these requests to decide how services should be distributed ("scheduled") on your physical nodes. For example a service which has a request of 4GB of RAM will never be scheduled to a node with only 1GB of RAM.
- Number of containers sets a hard limit on the number of containers your cluster will spawn for a service. This can be a minimum (e.g. "always spawn at least one container") or a maximum ("no more than 10") or a combination.
- Node allocation limits containers to run only on certain nodes based on the names and/or tags of those nodes.
For more detailed information on Limits and Requests, please read the in-depth Kubernetes documentation on the subject. For more detailed info on node allocation please read the Kubernetes documentation on node affinity.
Limiting CPU and RAM usage
By default, Docker services will use as much CPU and memory as they require. You may wish to set a hard limit on memory, or the relative CPU shares used by a service.
This can be accomplished in Cloud 66 by adding the
constraints/resources directive to your service's
CPU usage is measured in millicpu. One full CPU is 1000m, so 100m is equivalent to 10% of a CPU's capacity. In the context of limits, this is how much CPU capacity a service can consume, but the service may not need this capacity and will not reserve it.
If multiple services attempt to use a CPU and the aggregate demand is greater than 100%, Cloud 66 will share the CPU on a proportional basis (e.g. giving larger shares to the services with higher limits).
Memory limits apply to RAM usage and SWAP usage. A valid entry is a positive number, followed by one of, "K", "M", or "G", with the minimum being "4M". (You can also use the power-of-two equivalents like Gi, Mi, Ki if you prefer.)
Example of limits
Requesting CPU and RAM
To request resources for a service, you should add the
constraints/resources directive to your service's
service.yml. If the aggregate resource requests of all of the services in an application exceed the capacity of the destination cluster, some services will not be distributed. Please read the in-depth Kubernetes documentation to understand the implications of this.
CPU usage is measured in millicpu. One full CPU is 1000m, so 100m is equivalent to 10% of a CPU's capacity. In the context of requests, this is the minimum CPU capacity a service requires, but a service can use more resources than it has requested, as long as it is not also limited.
Memory requests set the minimum RAM required by a service. A valid entry is a positive number, followed by one of, "K", "M", or "G", with the minimum being "4M". (You can also use the power-of-two equivalents like Gi, Mi, Ki if you prefer.)
Example of requests
Combining limits and requests
You can set both requests and limits for a service. By definition request values must always be lower than limit values.
The example above will ensure this service always has access to at least 10% of a CPU and 100MB of RAM, and will allow this service to use up to a full CPU and 500MB of RAM.
Limiting the number of containers
You can limit the number of containers that a service may use. You can set both minimum and maximum limits for each service. For example:
This would ensure that the service would always have at least 2 containers, but that it would never have more than 7.
Allocating services to nodes
Node allocation is a mechanism for ensuring that services only spawn on nodes that meet certain criteria. You might want services with high volume of disk I/O to only spawn on the nodes that have access to SSD drives, for example. To allocate a service to nodes, you need to specify nodes using name(s) and/or tag(s).
Allocating nodes using names
This will limit services to only use nodes that match specified names.
This will ensure that the webapp service will only ever spawn on any nodes (servers) named alpha or zeta. These names are totally arbitrary and can be whatever you like.
Kubernetes will still handle the scheduling as usual, and other constraints may also affect the outcome. For example if all the nodes named Alpha are over capacity, the service may only spawn on Zeta nodes.
Allocating nodes using tags
This will limit services to only use nodes that match (arbitrary) tags. To use this method, you first need to tag the node(s) in question using the format "key=value" (for example
disk=large). Once you have tagged nodes according to your needs, you can allocate services to them like this:
This will ensure that the webapp service will only spawn on nodes that are tagged with
disk=large. Since the tags are arbitrary and do not reflect the underlying nature of a Node, it's critical to tag your nodes with great care. Setting everything to
disk=large, for example, will defeat the point.
Allocating nodes using a combination of names and tags
You can use a combination of names and tags allocate your services to nodes. If you include both these criteria, this is interpreted as a logical AND. For example this:
...tells your service that it must spawn on nodes that are named
zeta AND that are tagged
priority=medium. If, for example, a node is named Alpha but is tagged as low priority, the webapp service will not use that node.