Yorc Supported infrastructures

This section describes the state of our integration with supported infrastructures and their specificities. Yorc deals with the concept of locations as described in Alien4Cloud documentation

A location allows Yorc to connect to an infrastructure and to handle multiple locations of the same infrastructure, identified uniquely with its name.

Hosts Pool

prod

The Hosts Pool is a very special kind of infrastructure. It consists in registering existing Compute nodes into a pool managed by Yorc. Those compute nodes could be physical or virtual machines, containers or whatever as long as Yorc can SSH into it. Yorc will be responsible to allocate and release hosts for deployments. This is safe to use it concurrently in a Yorc cluster, Yorc instances will synchronize amongst themselves to ensure consistency of the pool.

To sum up this infrastructure type is really great when you want to use an infrastructure that is not yet supported by Yorc. Just take care you’re responsible for handling the compatibility or conflicts of what is already installed and what will be by Yorc on your hosts pool. The best practice is using container isolation. This is especially true if a host can be shared by several apps by specifying in Tosca with the Compute shareable property.

Hosts management

Yorc comes with a REST API that allows to manage hosts in the pool and to easily integrate it with other systems. The Yorc CLI leverage this REST API to make it user friendly, please refer to CLI Commands related to hosts pool for more information

Hosts Pool labels & filters

It is strongly recommended to associate labels to your hosts. Labels allow to filter hosts based on criteria. Labels are just a couple of key/value pair

Filter on label existence

These filters are used to check whether a label has been defined or not for the host, regardless of its value

  • label_identifier will match if a label with the given identifier is defined. Example : gpu
  • !label_identifier will match if no label with the given identifier has been defined for the host. Example : !gpu

Filter on string value

These filters are used to check whether a label value is matching a string. String value has to be between simple or double quotes : "" or ''.

  • label_identifier = "wanted_value" and label_identifier =='wanted_value' will match if the label with the given name has wanted_value as a value. Example : somename = "somevalue"
  • label_identifier != "wanted_value" will match if the label with the given name has not wanted_value as a value. Example : somename != "somevalue"

Please note that when used through Yorc CLI interface, the filter has to be between double quotes "", and the filter value has to be between simple quotes ' : yorc hp list -f "somename='someval'" is a valid command, while yorc hp list -f somename="someval" and yorc hp list -f 'somename="someval"' are not.

Filter on numeric value

These filters are used to check how a label value compares to a numeric value. Numeric value is a number written without quotes and an optional unit. Currently supported units are golang durations (“ns”, “us” , “ms”, “s”, “m” or “h”), bytes units (“B”, “KiB”, “KB”, “MiB”, “MB”, “GiB”, “GB”, “TiB”, “TB”, “PiB”, “PB”, “EiB”, “EB”) and International System of Units (SI). The case of the unit does not matter.

  • label_identifier == wanted_value and label_identifier == wanted_value will match if the label with the given name has a value equal to wanted_value. Example : somename = 100
  • label_identifier != wanted_value will match if the label with the given name has a value different from wanted_value. Example : somename != 100
  • label_identifier > wanted_value will match if the label with the given name has a value strictly superior to wanted_value. Example : somename > 100
  • label_identifier < wanted_value will match if the label with the given name has a value strictly inferior to wanted_value. Example : somename < 100
  • label_identifier >= wanted_value will match if the label with the given name has a value superior or equal to wanted_value. Example : somename >= 100 ms
  • label_identifier <= wanted_value will match if the label with the given name has a value inferior or equal to wanted_value. Example : somename <= 100

Filter on regex value

These filters are used to check if a label value contains or excludes a regex. Regex value has to be between simple or double quotes : "" or ''. “Contains” means that the value (string) of the label contains at least one substring matching the regex. “Excludes” means that the value (string) of the label contains no substring matching the regex.

  • label_identifier ~= "wanted_value" will match if the label with the given name has a value containing wanted_value. Example : somename ~= "(a|bc)+"
  • label_identifier !~ "wanted_value" will match if the label with the given name has a value excluding wanted_value . Example : somename !~ "(a|bc)+"

Filter on set appartenance

These filters are used to check is a label value is matching with one of the value of a set.

  • label_identifier in (firstval, "secondval") will match if the label with the given name has for value firstval or secondval. Example : somename in (gpu, cpu, none)
  • label_identifier not in ("firstval", "secondval") and label_identifier notin (firstval, secondval) will match if the label with the given name has not for value firstval or secondval. Example : somename notin (gpu, cpu, none)

Please note that quote around the values are optional, and that the values will always be considered as strings here. Therefore, label_identifier in (100) will not match if the string value of the label is 100.0.

Here are some example:

  • gpu
  • os.distribution != windows
  • os.architecture == x86_64
  • environment = "Q&A"
  • environment in ( "Q&A", dev, edge)
  • gpu.type not in (k20, m60)
  • gpu_nb > 1
  • os.mem_size >= 4 GB
  • os.disk_size < 1tb
  • max_allocation_time <= 120h

Implicit filters & labels

TOSCA allows to specify requirements on Compute hardware and Compute operating system . These are capabilities named host and os in the TOSCA node Compute . If those are specified in the topology, Yorc will automatically add a filter host.<property_name> >= <property_value> <property_unit> or os.<property_name> = <property_value> This will allow to select hosts matching the required criteria.

This means that it is strongly recommended to add the following labels to your hosts:
  • host.num_cpus (ie. host.num_cpus=4)
  • host.cpu_frequency (ie. host.cpu_frequency=3 GHz)
  • host.disk_size (ie. host.disk_size=50 GB)
  • host.mem_size (ie. host.mem_size=4GB)
  • os.architecture (ie. os.architecture=x86_64)
  • os.type (ie. os.type=linux)
  • os.distribution (ie. os.distribution=ubuntu)
  • os.version (ie. os.version=17.10)

Some labels are also automatically exposed as TOSCA Compute instance attributes:

  • if present a label named private_address will be used as attribute private_address and ip_address of the Compute. If not set the connection host will be used instead this allows to specify a network different for the applicative communication and for the orchestrator communication
  • if present a label named public_address will be used as attribute public_address of the Compute.
  • if present, following labels will fill the networks attribute of the Compute node:
    • networks.<idx>.network_name (ie. networks.0.network_name)
    • networks.<idx>.network_id (ie. networks.0.network_id)
    • networks.<idx>.addresses as a coma separated list of addresses (ie. networks.0.addresses)

The resources host pool labels (host.num_cpus, host.disk_size, host.mem_size) are automatically decreased and increased respectively when a host pool is allocated and released only if you specify any of these Tosca host resources capabilities Compute in its Alien4Cloud applications. If you apply a new configuration on allocated hosts with new host resources labels, they will be recalculated depending on existing allocations resources.

Hosts Pool Generic Resources

If you want to require specific resources as GPU by instance for your application’s computes, you can declare in the hosts pool configuration the available list of GPUs for each host. To do that, you will use a host.resource.gpu label, a comma-separated list of GPU resources as in the example below:

hosts:
- name: host1
  connection:
    user: centos
    host: 1.2.3.4
    private_key: /home/user/.ssh/yorc.pem
    port: 22
  labels:
    os.architecture: x86_64
    os.distribution: centos
    os.type: linux
    host.resource.gpu: "gpu2"
- name: hostspool-ci-1
  connection:
    user: centos
    host: 6.7.8.9
    private_key: /home/user/.ssh/yorc.pem
    port: 22
  labels:
    os.architecture: x86_64
    os.distribution: centos
    os.type: linux
    host.resource.gpu: "gpu0,gpu1"
In this example:
  • host1 provides a list of GPUs with a single GPU ID: gpu2.
  • host2 provides a list of GPUs with 2 ids: gpu0 and gpu1.

To require these specific resources in your application, an implicit matching will be done between the host.resource.<name> labels and the Tosca host capability.

The host capability has been extended with yorc.capabilities.hostspool.Container to provide a resources property.

The resource property is a list of yorc.datatypes.hostspool.GenericResource

A Generic Resource is defined with the following properties:

  • name: The name of the generic resource. Can be “gpu” by instance and must be bound to host labels as: host.resource.<name>.
  • ids: List of required generic resource ID’s by node instance. Each list entry corresponds to a comma-separated list of required generic resource ID’s for each node instance. An ID must only contains the following characters: -_0-9a-zA-Z_:./-
  • number: The number of generic resource required. Either ids or number must be filled to define the resource need.

Here is an example of an application which requires some GPUs:

topology_template:
  node_templates:
    ComputeA:
      type: yorc.nodes.hostspool.Compute
      properties:
        shareable: true
      capabilities:
        host:
          properties:
            resources:
            - name: gpu
              ids:
                - gpu2
    ComputeB:
      type: yorc.nodes.hostspool.Compute
      properties:
        shareable: true
      capabilities:
        host:
          properties:
            resources:
            - name: gpu
              number: 2

The ComputeA node requires a specific GPU’s ID: gpu2.

The ComputeB node requires 2 GPUs without specifying any ID’s requirement.

If you deploy the application on the hosts pool location previously defined, you will get the following allocations:

$ yorc hp list -l hp
+----------------+--------------------------------------------+-----------+--------------------------------+---------+-----------------------------------+
| Name           | Connection                                 | Status    | Allocations                    | Message | Labels                            |
+----------------+--------------------------------------------+-----------+--------------------------------+---------+-----------------------------------+
| host1          | user: centos                               | allocated | deployment: testApp            |         | host.resource.gpu: ""             |
|                | private key: /home/user/.ssh/yorc.pem      |           | node-instance: ComputeA        |         | os.architecture: x86_64           |
|                | host: 1.2.3.4                              |           | shareable: true                |         | os.distribution: centos           |
|                | port: 22                                   |           | host.resource.gpu: "gpu2"      |         | os.type: linux                    |
|                |                                            |           |                                |         |                                   |
|                |                                            |           |                                |         |                                   |
|                |                                            |           |                                |         |                                   |
|                |                                            |           |                                |         |                                   |
| host2          | user: centos                               | allocated | deployment: testApp            |         | host.resource.gpu: ""             |
|                | private key: /home/user/.ssh/yorc.pem      |           | node-instance: ComputeB        |         | os.architecture: x86_64           |
|                | host: 6.7.8.9                              |           | shareable: true                |         | os.distribution: centos           |
|                | port: 22                                   |           | host.resource.gpu: "gpu0,gpu1" |         | os.type: linux                    |
+----------------+--------------------------------------------+-----------+--------------------------------+---------+-----------------------------------+

The ComputeA GPU requirement on a gpu2 ID has been done by host1.

The ComputeB GPU requirement of 2 GPUs ID has been done by host2.

Both host1 and host2 are no longer providing GPUs resources as these resources are defined as consumable.

By default, a generic resource is consumable. It means a resource can be only used by a single compute. If you want to share a generic resource among many computes, you have to specify the following label host.resource.gpu.no_consume: true as below in the host declaration:

hosts:
- name: hostspool-ci-1
  labels:
    host.resource.gpu: "gpu0,gpu1"
    host.resource.gpu.no_consume: true

A Tosca instance attribute “gpu” will be exposed with the allocated resources for each node instance once the application is deployed.

Note: If you apply a new configuration on allocated hosts with new host generic resources labels, they will be recalculated depending on existing allocations resources.

Slurm

prod

Slurm is an open source, fault-tolerant, and highly scalable cluster management and job scheduling system for large and small Linux clusters. It is wildly used in High Performance Computing and it is the default scheduler of the Bull Super Computer Suite .

Yorc interacts with Slurm to allocate nodes on its cluster but also to run jobs.

Slurm jobs have been modeled in Tosca and this allows Yorc to execute them, either as regular jobs or as Singularity jobs.

Singularity is a container system similar to Docker but designed to integrate well HPC environments. Singularity allows users execute a command inside a Singularity or a Docker container, as a job submission. See Working with jobs for more information.

Yorc supports the following resources on Slurm:

  • Node Allocations as Computes
  • Jobs
  • Singularity Jobs.

Resources based scheduling

TOSCA allows to specify requirements on Compute nodes if specified num_cpus and mem_size requirements are used to allocate only the required resoures on computes. This allows to share a Slurm managed compute across several deployments. If not specified a whole compute node will be allocated.

Yorc also support Slurm GRES based scheduling. This is generally used to request a host with a specific type of resource (consumable or not) such as GPUs.

Google Cloud Platform

prod

The Google Cloud Platform integration within Yorc is ready for production and we support the following resources:

  • Compute Instances
  • Persistent Disks
  • Virtual Private Clouds (VPC)
  • Static IP Addresses.

Future work

It is planned to support soon the following feature:

  • Cloud VPN

AWS

dev

The AWS integration within Yorc allows to provision:
  • EC2 Compute Instances.
  • Elastic IPs.

This part is ready for production but we plan to support soon the following features to make it production-ready:

  • Elastic Block Store provisioning
  • Networks provisioning with Virtual Private Cloud

Future work

  • We plan to work on modeling AWS Batch Jobs in TOSCA and execute them thanks to Yorc.
  • We plan to work on AWS ECS to deploy containers

OpenStack

prod

The OpenStack integration within Yorc is production-ready. Yorc is currently supporting:

  • Compute Instances
  • Block Storages
  • Virtual Networks
  • Floating IPs provisioning.

Future work

Kubernetes

prod

The Kubernetes integration within Yorc is now production-ready. Yorc is currently supporting the following K8s resources:

  • Deployments.
  • Jobs.
  • Services.
  • StatefulSets.
  • PersistentVolumeClaims.

The Google Kubernetes Engine is also supported as a Kubernetes cluster.

Future work

It is planned to support soon the following features:

  • ConfigMaps.
  • Secrets.