AKS Installation
For AKS Installation, move to Step 5. Run the AI Fabric Application Installer
Run the AI Fabric infrastructure installer. Completing this installer will produce the Kots admin console where you are able to manage application updates, application configuration, resource usage (CPU/mem pressure), and download support bundles to troubleshoot any issues.
Caution
Do not kill the process or shutdown the machine will this step is running. This step will complete in 15-25 minutes. If you accidentally cause the process to terminate in the middle, the machine needs to be re-provisioned and brand new disks need to be attached.
Online Installation
First step is to download installer zip file here and move it to AI Fabric server. Alternatively, you can download it directly from the machine using following command
Make sure you have enough space
The script will download some files locally as part of installation process, please make sure you have 4GB available on the directory where you are executing script.
By default Azure RHEL VMs have only 1 GB available on home directory which is default directory.
wget https://download.uipath.com/aifabric/online-installer/v2020.10.2/aifabric-installer-v20.10.2.tar.gz
Then untarthe file and go inside main folder using following command:
tar -xvf aifabric-installer-v20.10.2.tar.gz
cd ./aifabric-installer-v20.10.2
You can then run AI Fabric installer by running:
./setup.sh
The first step is to accept license agreement by pressing Y. The script will then ask you what type of platform you want to install, enter onebox and press enter as on image below:
You will then be asked if a GPU is available for your setup and Y or N depending of your hardware. Make sure that drivers are already installed.
Using GPU
Only NVIDIA GPU's are supported and drivers need to be installed prior AI Fabric installation.
Depending of your system you might be asked to press Y few times for the installation to complete.
This step will take between 15-25 minutes to complete. Upon completion, you will see on the terminal output a message Installation Complete.
Airgapped Installation
On on local machine with access to a browser (e.g. a Windows server) download bundle install using link provided by your account manager.
Extract the contents of the downloaded file using 7zip from a windows file explorer or tar -zxvf aifabric-installer-v2020.10.2.tar.gz from a machine that supports tar.
This will create two folders:
aif_infra_20.10.2.tar.gzcontaining infrastructure components (about 3.6 GB)ai-fabric-v2020.10.2.airgapcontaining application components (about 8.7 GB). This will be uploaded to the UI in step 5. Run the AI Fabric Application Installer.
Copy the folder aif_infra_20.10.2.tar.gzto the airgapped AI Fabric machine.
Then run following command to start infrastructure installer:
tar -zxvf aif_infra_20.10.2.tar.gz
cd aif_infra_20.10.2
sudo ./setup.sh
Admin Console Access
In both case, successfull installation will ouput address and password of KotsAdmin Ui
...
Install Successful:
configmap/kurl-config created
Installation
Complete ✔
Kotsadm: http://13.59.108.17:8800
Login with password (will not be shown again): NNqKCY82S
The UIs of Prometheus, Grafana and Alertmanager have been exposed on NodePorts 30900,
30902 and 30903 respectively.
To access Grafana use the generated user:password of admin:msDX5VZ9m .
To access the cluster with kubectl, reload your shell:
bash -l
...
Note of the address of the kotsadm UI, this is on <machine-ip>:8800. In some case it ma display internal IP instead of public IP, make sure you are using public IP if you are accessing it from outside.
Note on the line below you will see the password login. Make a note of this password. You can regenerate this password if it is lost or if you would like to reset it:
bash -l
kubectl kots reset-password -n default
Adding GPU after install
If GPU wasn't available during installation but is later added to the machine you need to complete the following steps to make it accessible to AI Fabric.
First, validate that GPU's drivers are correctly installed (prerequisite) running the command:
nvidia-smi
You should see GPU information, if you see an error it means that GPU is not accessible or drivers are not correctly installed, please fix that before proceeding.
Then you need to run a script for adding the GPU to the cluster so that Pipelines and ML Skills can consume it.
For online install do the following:
# navigate to where you untar installer (or redo it if you have removed it)
cd ./aicenter-installer-v21.4.0/infra/common/scripts
./attach_gpu_drivers.sh
For airgapped this file won't be present so you need to create it first, navigate to aif_infra directory and create it there (it is important that nvidia-device-plugin.yaml is located on same folder):
#!/bin/bash
function edit_daemon_json(){
echo "################## Updating docker configuration ######################"
sudo bash -c '
echo \
"{
\"default-runtime\": \"nvidia\",
\"exec-opts\": [\"native.cgroupdriver=systemd\"],
\"runtimes\": {
\"nvidia\": {
\"path\": \"/usr/bin/nvidia-container-runtime\",
\"runtimeArgs\": []
}
}
}" > /etc/docker/daemon.json'
}
function kubernetes_cluster_up() {
count=0
swapoff -a
while [ $count -lt 50 ]; do
sudo chmod +r /etc/kubernetes/admin.conf
export KUBECONFIG=/etc/kubernetes/admin.conf
result=$(kubectl get nodes| grep master)
if [[ "$result" == *"master"* ]]; then
echo "Kubernetes up after " $((count * 5)) "seconds"
break
else
echo "Kubernetes not up, retry : " $count
count=$(( $count + 1 ))
sleep 5
fi
done
if [ $count == 50 ]; then
echo "Kubernetes Failed to come up"
exit
fi
}
function validate_gpu_updated() {
count=0
while [ $count -lt 50 ]; do
result=$(kubectl describe nodes| grep nvidia.com/gpu)
if [[ "$result" == *"nvidia.com/gpu"* ]]; then
echo $result
echo "Node gpu info updated after " $((count * 5)) "seconds"
echo "##################### Successfully installed GPU #########################"
break
else
echo "kubectl gpu info not updated, retry : " $count
count=$(( $count + 1 ))
sleep 5
fi
done
if [ $count == 50 ]; then
echo "################## Failed to install gpu ####################"
swapoff -a
exit
fi
}
function restart_docker() {
sudo pkill -SIGHUP dockerd
sudo systemctl restart docker
count=0
while [ $count -lt 50 ]; do
result=$(sudo systemctl status docker| grep running)
if [[ "$result" == *"running"* ]]; then
echo "docker is up " $((count * 5)) "seconds"
break
else
echo "docker is not up, retry : " $count
count=$(( $count + 1 ))
sleep 5
fi
done
if [ $count == 50 ]; then
echo "Docker Failed to come up"
swapoff -a
exit
fi
}
echo "################################# Attach Gpu Driver #####################################"
edit_daemon_json
echo "################################# restarting docker #####################################"
restart_docker
sleep 2
# This is required because when kubeadm init start kubelet, its check docker cgroup driver and uses cgroupfs
# in case of discrepancy. So we have to change this driver and restart kubelet again
echo "################################# restarting kubelet #####################################"
sudo sed -i 's/cgroup-driver=cgroupfs/cgroup-driver=systemd/' /var/lib/kubelet/kubeadm-flags.env
sudo systemctl restart kubelet
sleep 10
kubernetes_cluster_up
kubectl apply -f nvidia-device-plugin.yaml
validate_gpu_updated
then just execute that script:
./attach_gpu_drivers.sh
At the end you should see installation as successful
故障排除
The infrastructure installer is not idempotent. This means that running the installer again (after you have already run it once) will not work. If this installer fails, you will need to reprovision a new machine with fresh disks.
The most common sources of error are that the bootdisk becomes full during the install or that the external data disks are mounted/formatted. Remember to only attach the disks, not format them.
If the installation fails with unformatted disks and a sufficiently large boot risk, contact our support team and include in your email a support bundle. A support bundle can be generated by running this command:
curl https://krew.sh/support-bundle | bash
kubectl support-bundle https://kots.io
Alternatively if you don't have access to the internet you can create file support-bundle.yaml with following text:
apiVersion: troubleshoot.replicated.com/v1beta1
kind: Collector
metadata:
name: collector-sample
spec:
collectors:
- clusterInfo: {}
- clusterResources: {}
- exec:
args:
- "-U"
- kotsadm
collectorName: kotsadm-postgres-db
command:
- pg_dump
containerName: kotsadm-postgres
name: kots/admin_console
selector:
- app=kotsadm-postgres
timeout: 10s
- logs:
collectorName: kotsadm-postgres-db
name: kots/admin_console
selector:
- app=kotsadm-postgres
- logs:
collectorName: kotsadm-api
name: kots/admin_console
selector:
- app=kotsadm-api
- logs:
collectorName: kotsadm-operator
name: kots/admin_console
selector:
- app=kotsadm-operator
- logs:
collectorName: kotsadm
name: kots/admin_console
selector:
- app=kotsadm
- logs:
collectorName: kurl-proxy-kotsadm
name: kots/admin_console
selector:
- app=kurl-proxy-kotsadm
- secret:
collectorName: kotsadm-replicated-registry
includeValue: false
key: .dockerconfigjson
name: kotsadm-replicated-registry
- logs:
collectorName: rook-ceph-agent
selector:
- app=rook-ceph-agent
namespace: rook-ceph
name: kots/rook
- logs:
collectorName: rook-ceph-mgr
selector:
- app=rook-ceph-mgr
namespace: rook-ceph
name: kots/rook
- logs:
collectorName: rook-ceph-mon
selector:
- app=rook-ceph-mon
namespace: rook-ceph
name: kots/rook
- logs:
collectorName: rook-ceph-operator
selector:
- app=rook-ceph-operator
namespace: rook-ceph
name: kots/rook
- logs:
collectorName: rook-ceph-osd
selector:
- app=rook-ceph-osd
namespace: rook-ceph
name: kots/rook
- logs:
collectorName: rook-ceph-osd-prepare
selector:
- app=rook-ceph-osd-prepare
namespace: rook-ceph
name: kots/rook
- logs:
collectorName: rook-ceph-rgw
selector:
- app=rook-ceph-rgw
namespace: rook-ceph
name: kots/rook
- logs:
collectorName: rook-discover
selector:
- app=rook-discover
namespace: rook-ceph
name: kots/rook
And then create support-bundle file using following command:
kubectl support-bundle support-bundle.yaml
This will create a file called supportbundle.tar.gz which you can upload when raising a support ticket.
Updated 13 days ago