Dell-Omnia-Clusters-HPC-AI/docs

Omnia (Latin: all or everything) is a deployment tool to configure Dell EMC PowerEdge servers running standard RPM-based Linux OS images into clusters capable of supporting HPC, AI, and data analytics workloads. It uses Slurm, Kubernetes, and other packages to manage jobs and run diverse workloads on the same converged solution. It is a collection of Ansible playbooks, is open source, and is constantly being extended to enable comprehensive workloads.

What Omnia does

Omnia can build clusters which use Slurm or Kubernetes (or both!) for workload management. Omnia will install software from a variety of sources, including:

• Standard CentOS and ELRepo repositories
• Helm repositories
• Source code compilation
• OpenHPC repositories (coming soon!)
• OperatorHub (coming soon!)

Whenever possible, Omnia will leverage existing projects rather than reinvent the wheel.

Omnia stacks

Omnia can install Kubernetes or Slurm (or both), along with additional drivers, services, libraries, and user applications.

Installing Omnia

Omnia requires that servers already have an RPM-based Linux OS running on them, and are all connected to the Internet. Currently all Omnia testing is done on CentOS. Please see Preparation to install Omnia for instructions on network setup.

Once servers have functioning OS and networking, you can use Omnia to install and start Slurm and/or Kubernetes. Please see Install Omnia using CLI for detailed instructions.

Installing the Omnia appliance

Ensure all the prerequisites listed in the PREINSTALL_OMNIA_APPLIANCE are met before installing the Omnia appliance.

For detailed instructions on installing the Omnia appliance, see INSTALL_OMNIA_APPLIANCE.

System requirements

Ensure the supported version of all the software are installed as per the following table and other versions than those listed are not supported by Omnia. This is to ensure that there is no impact to the functionality of Omnia.

Software and hardware requirements	Version
OS installed on the management node	CentOS 7.9 2009
OS deployed by Omnia on bare-metal servers	CentOS 7.9 2009 Minimal Edition
Cobbler	2.8.5
Ansible AWX	15.0.0
Slurm Workload Manager	20.11.2
Kubernetes Controllers	1.16.7
Kubeflow	1
Prometheus	2.23.0
Supported PowerEdge servers	R640, R740, R7525, C4140, DSS8440, and C6420

Software managed by Omnia

Ensure the supported version of all the software are installed as per the following table and other versions than those listed are not supported by Omnia. This is to ensure that there is no impact to the functionality of Omnia.

Software	Licence	Compatible Version	Description
MariaDB	GPL 2.0	5.5.68	Relational database used by Slurm
Slurm	GNU General Public	20.11.2	HPC Workload Manager
Docker CE	Apache-2.0	20.10.2	Docker Service
NVIDIA container runtime	Apache-2.0	3.4.2	Nvidia container runtime library
Python PIP	MIT Licence	3.2.1	Python Package
Python2	-	2.7.5	-
Kubelet	Apache-2.0	1.16.7	Provides external, versioned ComponentConfig API types for configuring the kubelet
Kubeadm	Apache-2.0	1.16.7	"fast paths" for creating Kubernetes clusters
Kubectl	Apache-2.0	1.16.7	Command line tool for Kubernetes
JupyterHub	Modified BSD Licence	1.1.0	Multi-user hub
Kfctl	Apache-2.0	1.0.2	CLI for deploying and managing Kubeflow
Kubeflow	Apache-2.0	1	Cloud Native platform for machine learning
Helm	Apache-2.0	3.5.0	Kubernetes Package Manager
Helm Chart	-	0.9.0	-
TensorFlow	Apache-2.0	2.1.0	Machine Learning framework
Horovod	Apache-2.0	0.21.1	Distributed deep learning training framework for Tensorflow
MPI	Copyright (c) 2018-2019 Triad National Security,LLC. All rights reserved.	0.2.3	HPC library
CoreDNS	Apache-2.0	1.6.2	DNS server that chains plugins
CNI	Apache-2.0	0.3.1	Networking for Linux containers
AWX	Apache-2.0	15.0.0	Web-based User Interface
PostgreSQL	Copyright (c) 1996-2020, PostgreSQL Global Development Group	10.15	Database Management System
Redis	BSD-3-Clause Licence	6.0.10	In-memory database
NGINX	BSD-2-Clause Licence	1.14	-

Known issue

Issue: Hosts do not display on the AWX UI.

Resolution:

• Verify if provisioned_hosts.yml is present in the omnia/appliance/roles/inventory/files folder.
• Verify if hosts are not listed in the provisioned_hosts.yml file. If hosts are not listed, then servers are not PXE booted yet.
• If hosts are listed in the provisioned_hosts.yml file, then an IP address has been assigned to them by DHCP. However, hosts are not displayed on the AWX UI as the PXE boot is still in process or is not initiated.
• Check for the reachable and unreachable hosts using the provisioned_report.yml tool present in the omnia/appliance/tools folder. To run provisioned_report.yml, in the omnia/appliance directory, run playbook -i roles/inventory/files/provisioned_hosts.yml tools/provisioned_report.yml.

Frequently asked questions

• Why is the error "Wait for AWX UI to be up" displayed when appliance.yaml fails?
  Cause:

  1. When AWX is not accessible even after five minutes of wait time.
  2. When isMigrating or isInstalling is seen in the failure message.
     

  Resolution:
  Wait for AWX UI to be accessible at http://<management-station-IP>:8081, and then run the appliance.yaml file again, where management-station-IP is the ip address of the management node.

• What are the next steps after the nodes in a Kubernetes cluster reboots?
  Resolution: Wait for upto 15 minutes after the Kubernetes cluster reboots. Next, verify status of the cluster using the following services:

  • kubectl get nodes on the manager node provides correct k8s cluster status.
    
  • kubectl get pods --all-namespaces on the manager node displays all the pods in the Running state.
  • kubectl cluster-info on the manager node displays both k8s master and kubeDNS are in the Running state.

• What to do when the Kubernetes services are not in the Running state?
  Resolution:

  1. Run kubectl get pods --all-namespaces to verify the pods are in the Running state.
  2. If the pods are not in the Running state, delete the pods using the command:kubectl delete pods <name of pod>
  3. Run the corresponding playbook that was used to install Kubernetes: omnia.yml, jupyterhub.yml, or kubeflow.yml.

• What to do when the JupyterHub or Prometheus UI are not accessible?
  Resolution: Run the command kubectl get pods --namespace default to ensure nfs-client pod and all prometheus server pods are in the Running state.

• While configuring the Cobbler, why does the appliance.yml fail with an error during the Run import command?
  Cause:

  • When the mounted .iso file is corrupt.
    

  Resolution:

  1. Go to var->log->cobbler->cobbler.log to view the error.
  2. If the error message is repo verification failed then it signifies that the .iso file is not mounted properly.
  3. Verify if the downloaded .iso file is valid and correct.
  4. Delete the Cobbler container using docker rm -f cobbler and rerun appliance.yml.

• Why does the PXE boot fail with tftp timeout or service timeout errors?
  Cause:

  • When RAID is configured on the server.
  • When more than two servers in the same network have Cobbler services running.
    
    

  Resolution:

  1. Create a Non-RAID or virtual disk in the server.
     
  2. Check if other systems except for the management node has cobblerd running. If yes, then stop the Cobbler container using the following commands: docker rm -f cobbler and docker image rm -f cobbler.

• What to do when the Slurm services do not start automatically after the cluster reboots?
  Resolution:

  • Manually restart the slurmd services on the manager node by running the following commands:
    • systemctl restart slurmdbd
    • systemctl restart slurmctld
    • systemctl restart prometheus-slurm-exporter
  • Run systemctl status slurmd to manually restart the following service on all the compute nodes.

• What to do when the Slurm services fail? Cause: The slurm.conf is not configured properly.
  Resolution:

  1. Run the following commands:
     • slurmdbd -Dvvv
     • slurmctld -Dvvv
  2. Verify /var/lib/log/slurmctld.log file.

• What to do when when the error "ports are unavailable" is displayed? Cause: Slurm database connection fails.
  Resolution:

  1. Run the following commands:
     • slurmdbd -Dvvv *slurmctld -Dvvv
  2. Verify the /var/lib/log/slurmctld.log file.
  3. Verify: netstat -antp | grep LISTEN
  4. If PIDs are in the Listening state, kill the processes of that specific port.
  5. Restart all Slurm services:
     • slurmctl restart slurmctld on manager node
     • systemctl restart slurmdbd on manager node
     • systemctl restart slurmd on compute node
       

• What to do if Kubernetes Pods are unable to communicate with the servers when the DNS servers are not responding?
  Cause: With the host network which is DNS issue.
  Resolution:

  1. In your Kubernetes cluster, run kubeadm reset -f on the nodes.
  2. In the management node, edit the omnia_config.yml file to change the Kubernetes Pod Network CIDR. Suggested IP range is 192.168.0.0/16 and ensure you provide an IP which is not in use in your host network.
  3. Execute omnia.yml and skip slurm using _skip tag slurm.

• What to do if time taken to pull the images to create the Kubeflow containers exceeds the limit and the Apply Kubeflow configurations task fails?
  Cause: Unstable or slow Internet connectivity.
  Resolution:

  1. Complete the PXE booting/ format the OS on manager and compute nodes.
  2. In the omnia_config.yml file, change the k8s_cni variable value from calico to flannel.
  3. Run the Kubernetes and Kubeflow playbooks.

Limitations

1. Removal of Slurm and Kubernetes component roles are not supported. However, skip tags can be provided at the start of installation to select the component roles.​
2. After the installation of the Omnia appliance, changing the manager node is not supported. If you need to change the manager node, you must redeploy the entire cluster.
   
3. Dell Technologies provides support to the Dell developed modules of Omnia. All the other third-party tools deployed by Omnia are outside the support scope.​

4. To change the Kubernetes single node cluster to a multi-node cluster or to change a multi-node cluster to a single node cluster, you must either redeploy the entire cluster or run kubeadm reset -f on all the nodes of the cluster. You then need to run omnia.yml file and skip the installation of Slurm using the skip tags.

   Contributing to Omnia

   The Omnia project was started to give members of the Dell Technologies HPC Community a way to easily setup clusters of Dell EMC servers, and to contribute useful tools, fixes, and functionality back to the HPC Community.

Open to All

While we started Omnia within the Dell Technologies HPC Community, that doesn't mean that it's limited to Dell EMC servers, networking, and storage. This is an open project, and we want to encourage everyone to use and contribute to Omnia!

Anyone can contribute!

It's not just new features and bug fixes that can be contributed to the Omnia project! Anyone should feel comfortable contributing. We are asking for all types of contributions:

• New feature code
• Bug fixes
• Documentation updates
• Feature suggestions
• Feedback
• Validation that it works for your particular configuration

If you would like to contribute, see [CONTRIBUTORS](https://github.com/dellhpc/omnia/b