- Release Notes and Announcements
- Release Notes
- Announcements
- Security Vulnerability Fix Description
- Host Operation System Release for Super Node Pods (Mitigated NodeLost Issue)
- TKE Native Node Sub-product Name Change Notice
- Announcement on Authentication Upgrade of Some TKE APIs
- Discontinuing Update of NginxIngress Addon
- qGPU Service Adjustment
- Version Upgrade of Master Add-On of TKE Managed Cluster
- Upgrading tke-monitor-agent
- Instructions on Cluster Resource Quota Adjustment
- Decommissioning Kubernetes Version
- Deactivation of Scaling Group Feature
- Notice on TPS Discontinuation on May 16, 2022 at 10:00 (UTC +8)
- Basic Monitoring Architecture Upgrade
- Starting Charging on Managed Clusters
- Instructions on Stopping Delivering the Kubeconfig File to Nodes
- Release Notes
- Product Introduction
- Purchase Guide
- Quick Start
- TKE General Cluster Guide
- TKE General Cluster Overview
- Purchase a TKE General Cluster
- High-risk Operations of Container Service
- Deploying Containerized Applications in the Cloud
- Open Source Components
- Permission Management
- Cluster Management
- Cluster Overview
- Cluster Hosting Modes Introduction
- Cluster Lifecycle
- Creating a Cluster
- Creating a Cluster (New)
- Changing the Cluster Operating System
- Deleting a Cluster
- Cluster Scaling
- Connecting to a Cluster
- Upgrading a Cluster
- Enabling IPVS for a Cluster
- Custom Kubernetes Component Launch Parameters
- Using KMS for Kubernetes Data Source Encryption
- Images
- Worker node introduction
- Normal Node Management
- Native Node Management
- Overview
- Native Node Parameters
- Purchasing Native Nodes
- Lifecycle of a Native Node
- Creating Native Nodes
- Modifying Native Nodes
- Deleting Native Nodes
- Self-Heal Rules
- Declarative Operation Practice
- Native Node Scaling
- In-place Pod Configuration Adjustment
- Enabling Public Network Access for a Native Node
- Management Parameters
- Enabling SSH Key Login for a Native Node
- FAQs for Native Nodes
- Supernode management
- Registered Node Management
- Memory Compression Instructions
- GPU Share
- Kubernetes Object Management
- Overview
- Namespace
- Workload
- Deployment Management
- StatefulSet Management
- DaemonSet Management
- CronJob Management
- Job Management
- Setting the Resource Limit of Workload
- Setting the Scheduling Rule for a Workload
- Setting the Health Check for a Workload
- Setting the Run Command and Parameter for a Workload
- Using a Container Image in a TCR Enterprise Instance to Create a Workload
- Configuration
- Auto Scaling
- Service Management
- Ingress Management
- Storage Management
- Policy Management
- Application and Add-On Feature Management Description
- Add-On Management
- Add-on Overview
- Add-On Lifecycle Management
- Cluster Autoscaler
- OOMGuard
- NodeProblemDetectorPlus Add-on
- NodeLocalDNSCache
- DNSAutoscaler
- COS-CSI
- CFS-CSI
- CFSTURBO-CSI
- CBS-CSI Description
- UserGroupAccessControl
- TCR Introduction
- TCR Hosts Updater
- DynamicScheduler
- DeScheduler
- Network Policy
- Nginx-ingress
- HPC
- Description of tke-monitor-agent
- tke-log-agent
- GPU-Manager Add-on
- Helm Application
- Application Market
- Network Management
- Container Network Overview
- GlobalRouter Mode
- VPC-CNI Mode
- VPC-CNI Mode
- Multiple Pods with Shared ENI Mode
- Pods with Exclusive ENI Mode
- Static IP Address Mode Instructions
- Non-static IP Address Mode Instructions
- Interconnection Between VPC-CNI and Other Cloud Resources/IDC Resources
- Security Group of VPC-CNI Mode
- Instructions on Binding an EIP to a Pod
- VPC-CNI Component Description
- Limits on the Number of Pods in VPC-CNI Mode
- Cilium-Overlay Mode
- OPS Center
- Log Management
- Backup Center
- Remote Terminals
- TKE Serverless Cluster Guide
- TKE Registered Cluster Guide
- TKE Insight
- TKE Scheduling
- Cloud Native Service Guide
- Practical Tutorial
- Cluster
- Cluster Migration
- Serverless Cluster
- Scheduling
- Security
- Service Deployment
- Network
- DNS
- Self-Built Nginx Ingress Practice Tutorial
- Quick Start
- Custom Load Balancer
- Enabling CLB Direct Connection
- Optimization for High Concurrency Scenarios
- High Availability Configuration Optimization
- Observability Integration
- Access to Tencent Cloud WAF
- Installing Multiple Nginx Ingress Controllers
- Migrating from TKE Nginx Ingress Plugin to Self-Built Nginx Ingress
- Complete Example of values.yaml Configuration
- Using Network Policy for Network Access Control
- Deploying NGINX Ingress on TKE
- Nginx Ingress High-Concurrency Practices
- Nginx Ingress Best Practices
- Limiting the bandwidth on pods in TKE
- Directly connecting TKE to the CLB of pods based on the ENI
- Use CLB-Pod Direct Connection on TKE
- Obtaining the Real Client Source IP in TKE
- Using Traefik Ingress in TKE
- Release
- Logs
- Monitoring
- OPS
- Removing and Re-adding Nodes from and to Cluster
- Using Ansible to Batch Operate TKE Nodes
- Using Cluster Audit for Troubleshooting
- Renewing a TKE Ingress Certificate
- Using cert-manager to Issue Free Certificates
- Using cert-manager to Issue Free Certificate for DNSPod Domain Name
- Using the TKE NPDPlus Plug-In to Enhance the Self-Healing Capability of Nodes
- Using kubecm to Manage Multiple Clusters kubeconfig
- Quick Troubleshooting Using TKE Audit and Event Services
- Customizing RBAC Authorization in TKE
- Clearing De-registered Tencent Cloud Account Resources
- Terraform
- DevOps
- Auto Scaling
- KEDA
- Cluster Auto Scaling Practices
- Using tke-autoscaling-placeholder to Implement Auto Scaling in Seconds
- Installing metrics-server on TKE
- Using Custom Metrics for Auto Scaling in TKE
- Utilizing HPA to Auto Scale Businesses on TKE
- Using VPA to Realize Pod Scaling up and Scaling down in TKE
- Adjusting HPA Scaling Sensitivity Based on Different Business Scenarios
- Implementing elasticity based on traffic prediction with EHPA
- Implementing Horizontal Scaling based on CLB monitoring metrics using KEDA in TKE
- Containerization
- Microservice
- Cost Management
- Hybrid Cloud
- Fault Handling
- Disk Full
- High Workload
- Memory Fragmentation
- Cluster DNS Troubleshooting
- Cluster kube-proxy Troubleshooting
- Cluster API Server Inaccessibility Troubleshooting
- Service and Ingress Inaccessibility Troubleshooting
- Common Service & Ingress Errors and Solutions
- Engel Ingres appears in Connechtin Reverside
- CLB Ingress Creation Error
- Troubleshooting for Pod Network Inaccessibility
- Pod Status Exception and Handling
- Authorizing Tencent Cloud OPS Team for Troubleshooting
- CLB Loopback
- API Documentation
- History
- Introduction
- API Category
- Making API Requests
- Elastic Cluster APIs
- Resource Reserved Coupon APIs
- Cluster APIs
- AcquireClusterAdminRole
- CreateClusterEndpoint
- CreateClusterEndpointVip
- DeleteCluster
- DeleteClusterEndpoint
- DeleteClusterEndpointVip
- DescribeAvailableClusterVersion
- DescribeClusterAuthenticationOptions
- DescribeClusterCommonNames
- DescribeClusterEndpointStatus
- DescribeClusterEndpointVipStatus
- DescribeClusterEndpoints
- DescribeClusterKubeconfig
- DescribeClusterLevelAttribute
- DescribeClusterLevelChangeRecords
- DescribeClusterSecurity
- DescribeClusterStatus
- DescribeClusters
- DescribeEdgeAvailableExtraArgs
- DescribeEdgeClusterExtraArgs
- DescribeResourceUsage
- DisableClusterDeletionProtection
- EnableClusterDeletionProtection
- GetClusterLevelPrice
- GetUpgradeInstanceProgress
- ModifyClusterAttribute
- ModifyClusterAuthenticationOptions
- ModifyClusterEndpointSP
- UpgradeClusterInstances
- CreateBackupStorageLocation
- CreateCluster
- DeleteBackupStorageLocation
- DescribeBackupStorageLocations
- DescribeEncryptionStatus
- DisableEncryptionProtection
- EnableEncryptionProtection
- UpdateClusterKubeconfig
- UpdateClusterVersion
- Third-party Node APIs
- Network APIs
- Node APIs
- Node Pool APIs
- TKE Edge Cluster APIs
- CheckEdgeClusterCIDR
- DescribeAvailableTKEEdgeVersion
- DescribeECMInstances
- DescribeEdgeCVMInstances
- DescribeEdgeClusterInstances
- DescribeEdgeClusterUpgradeInfo
- DescribeTKEEdgeClusterStatus
- ForwardTKEEdgeApplicationRequestV3
- DescribeEdgeLogSwitches
- CreateECMInstances
- CreateEdgeCVMInstances
- CreateEdgeLogConfig
- DeleteECMInstances
- DeleteEdgeCVMInstances
- DeleteEdgeClusterInstances
- DeleteTKEEdgeCluster
- DescribeTKEEdgeClusterCredential
- DescribeTKEEdgeExternalKubeconfig
- DescribeTKEEdgeScript
- InstallEdgeLogAgent
- UninstallEdgeLogAgent
- UpdateEdgeClusterVersion
- DescribeTKEEdgeClusters
- CreateTKEEdgeCluster
- Cloud Native Monitoring APIs
- Scaling group APIs
- Super Node APIs
- Add-on APIs
- Other APIs
- Data Types
- Error Codes
- TKE API 2022-05-01
- FAQs
- Service Agreement
- Contact Us
- Glossary
- User Guide(Old)
- Release Notes and Announcements
- Release Notes
- Announcements
- Security Vulnerability Fix Description
- Host Operation System Release for Super Node Pods (Mitigated NodeLost Issue)
- TKE Native Node Sub-product Name Change Notice
- Announcement on Authentication Upgrade of Some TKE APIs
- Discontinuing Update of NginxIngress Addon
- qGPU Service Adjustment
- Version Upgrade of Master Add-On of TKE Managed Cluster
- Upgrading tke-monitor-agent
- Instructions on Cluster Resource Quota Adjustment
- Decommissioning Kubernetes Version
- Deactivation of Scaling Group Feature
- Notice on TPS Discontinuation on May 16, 2022 at 10:00 (UTC +8)
- Basic Monitoring Architecture Upgrade
- Starting Charging on Managed Clusters
- Instructions on Stopping Delivering the Kubeconfig File to Nodes
- Release Notes
- Product Introduction
- Purchase Guide
- Quick Start
- TKE General Cluster Guide
- TKE General Cluster Overview
- Purchase a TKE General Cluster
- High-risk Operations of Container Service
- Deploying Containerized Applications in the Cloud
- Open Source Components
- Permission Management
- Cluster Management
- Cluster Overview
- Cluster Hosting Modes Introduction
- Cluster Lifecycle
- Creating a Cluster
- Creating a Cluster (New)
- Changing the Cluster Operating System
- Deleting a Cluster
- Cluster Scaling
- Connecting to a Cluster
- Upgrading a Cluster
- Enabling IPVS for a Cluster
- Custom Kubernetes Component Launch Parameters
- Using KMS for Kubernetes Data Source Encryption
- Images
- Worker node introduction
- Normal Node Management
- Native Node Management
- Overview
- Native Node Parameters
- Purchasing Native Nodes
- Lifecycle of a Native Node
- Creating Native Nodes
- Modifying Native Nodes
- Deleting Native Nodes
- Self-Heal Rules
- Declarative Operation Practice
- Native Node Scaling
- In-place Pod Configuration Adjustment
- Enabling Public Network Access for a Native Node
- Management Parameters
- Enabling SSH Key Login for a Native Node
- FAQs for Native Nodes
- Supernode management
- Registered Node Management
- Memory Compression Instructions
- GPU Share
- Kubernetes Object Management
- Overview
- Namespace
- Workload
- Deployment Management
- StatefulSet Management
- DaemonSet Management
- CronJob Management
- Job Management
- Setting the Resource Limit of Workload
- Setting the Scheduling Rule for a Workload
- Setting the Health Check for a Workload
- Setting the Run Command and Parameter for a Workload
- Using a Container Image in a TCR Enterprise Instance to Create a Workload
- Configuration
- Auto Scaling
- Service Management
- Ingress Management
- Storage Management
- Policy Management
- Application and Add-On Feature Management Description
- Add-On Management
- Add-on Overview
- Add-On Lifecycle Management
- Cluster Autoscaler
- OOMGuard
- NodeProblemDetectorPlus Add-on
- NodeLocalDNSCache
- DNSAutoscaler
- COS-CSI
- CFS-CSI
- CFSTURBO-CSI
- CBS-CSI Description
- UserGroupAccessControl
- TCR Introduction
- TCR Hosts Updater
- DynamicScheduler
- DeScheduler
- Network Policy
- Nginx-ingress
- HPC
- Description of tke-monitor-agent
- tke-log-agent
- GPU-Manager Add-on
- Helm Application
- Application Market
- Network Management
- Container Network Overview
- GlobalRouter Mode
- VPC-CNI Mode
- VPC-CNI Mode
- Multiple Pods with Shared ENI Mode
- Pods with Exclusive ENI Mode
- Static IP Address Mode Instructions
- Non-static IP Address Mode Instructions
- Interconnection Between VPC-CNI and Other Cloud Resources/IDC Resources
- Security Group of VPC-CNI Mode
- Instructions on Binding an EIP to a Pod
- VPC-CNI Component Description
- Limits on the Number of Pods in VPC-CNI Mode
- Cilium-Overlay Mode
- OPS Center
- Log Management
- Backup Center
- Remote Terminals
- TKE Serverless Cluster Guide
- TKE Registered Cluster Guide
- TKE Insight
- TKE Scheduling
- Cloud Native Service Guide
- Practical Tutorial
- Cluster
- Cluster Migration
- Serverless Cluster
- Scheduling
- Security
- Service Deployment
- Network
- DNS
- Self-Built Nginx Ingress Practice Tutorial
- Quick Start
- Custom Load Balancer
- Enabling CLB Direct Connection
- Optimization for High Concurrency Scenarios
- High Availability Configuration Optimization
- Observability Integration
- Access to Tencent Cloud WAF
- Installing Multiple Nginx Ingress Controllers
- Migrating from TKE Nginx Ingress Plugin to Self-Built Nginx Ingress
- Complete Example of values.yaml Configuration
- Using Network Policy for Network Access Control
- Deploying NGINX Ingress on TKE
- Nginx Ingress High-Concurrency Practices
- Nginx Ingress Best Practices
- Limiting the bandwidth on pods in TKE
- Directly connecting TKE to the CLB of pods based on the ENI
- Use CLB-Pod Direct Connection on TKE
- Obtaining the Real Client Source IP in TKE
- Using Traefik Ingress in TKE
- Release
- Logs
- Monitoring
- OPS
- Removing and Re-adding Nodes from and to Cluster
- Using Ansible to Batch Operate TKE Nodes
- Using Cluster Audit for Troubleshooting
- Renewing a TKE Ingress Certificate
- Using cert-manager to Issue Free Certificates
- Using cert-manager to Issue Free Certificate for DNSPod Domain Name
- Using the TKE NPDPlus Plug-In to Enhance the Self-Healing Capability of Nodes
- Using kubecm to Manage Multiple Clusters kubeconfig
- Quick Troubleshooting Using TKE Audit and Event Services
- Customizing RBAC Authorization in TKE
- Clearing De-registered Tencent Cloud Account Resources
- Terraform
- DevOps
- Auto Scaling
- KEDA
- Cluster Auto Scaling Practices
- Using tke-autoscaling-placeholder to Implement Auto Scaling in Seconds
- Installing metrics-server on TKE
- Using Custom Metrics for Auto Scaling in TKE
- Utilizing HPA to Auto Scale Businesses on TKE
- Using VPA to Realize Pod Scaling up and Scaling down in TKE
- Adjusting HPA Scaling Sensitivity Based on Different Business Scenarios
- Implementing elasticity based on traffic prediction with EHPA
- Implementing Horizontal Scaling based on CLB monitoring metrics using KEDA in TKE
- Containerization
- Microservice
- Cost Management
- Hybrid Cloud
- Fault Handling
- Disk Full
- High Workload
- Memory Fragmentation
- Cluster DNS Troubleshooting
- Cluster kube-proxy Troubleshooting
- Cluster API Server Inaccessibility Troubleshooting
- Service and Ingress Inaccessibility Troubleshooting
- Common Service & Ingress Errors and Solutions
- Engel Ingres appears in Connechtin Reverside
- CLB Ingress Creation Error
- Troubleshooting for Pod Network Inaccessibility
- Pod Status Exception and Handling
- Authorizing Tencent Cloud OPS Team for Troubleshooting
- CLB Loopback
- API Documentation
- History
- Introduction
- API Category
- Making API Requests
- Elastic Cluster APIs
- Resource Reserved Coupon APIs
- Cluster APIs
- AcquireClusterAdminRole
- CreateClusterEndpoint
- CreateClusterEndpointVip
- DeleteCluster
- DeleteClusterEndpoint
- DeleteClusterEndpointVip
- DescribeAvailableClusterVersion
- DescribeClusterAuthenticationOptions
- DescribeClusterCommonNames
- DescribeClusterEndpointStatus
- DescribeClusterEndpointVipStatus
- DescribeClusterEndpoints
- DescribeClusterKubeconfig
- DescribeClusterLevelAttribute
- DescribeClusterLevelChangeRecords
- DescribeClusterSecurity
- DescribeClusterStatus
- DescribeClusters
- DescribeEdgeAvailableExtraArgs
- DescribeEdgeClusterExtraArgs
- DescribeResourceUsage
- DisableClusterDeletionProtection
- EnableClusterDeletionProtection
- GetClusterLevelPrice
- GetUpgradeInstanceProgress
- ModifyClusterAttribute
- ModifyClusterAuthenticationOptions
- ModifyClusterEndpointSP
- UpgradeClusterInstances
- CreateBackupStorageLocation
- CreateCluster
- DeleteBackupStorageLocation
- DescribeBackupStorageLocations
- DescribeEncryptionStatus
- DisableEncryptionProtection
- EnableEncryptionProtection
- UpdateClusterKubeconfig
- UpdateClusterVersion
- Third-party Node APIs
- Network APIs
- Node APIs
- Node Pool APIs
- TKE Edge Cluster APIs
- CheckEdgeClusterCIDR
- DescribeAvailableTKEEdgeVersion
- DescribeECMInstances
- DescribeEdgeCVMInstances
- DescribeEdgeClusterInstances
- DescribeEdgeClusterUpgradeInfo
- DescribeTKEEdgeClusterStatus
- ForwardTKEEdgeApplicationRequestV3
- DescribeEdgeLogSwitches
- CreateECMInstances
- CreateEdgeCVMInstances
- CreateEdgeLogConfig
- DeleteECMInstances
- DeleteEdgeCVMInstances
- DeleteEdgeClusterInstances
- DeleteTKEEdgeCluster
- DescribeTKEEdgeClusterCredential
- DescribeTKEEdgeExternalKubeconfig
- DescribeTKEEdgeScript
- InstallEdgeLogAgent
- UninstallEdgeLogAgent
- UpdateEdgeClusterVersion
- DescribeTKEEdgeClusters
- CreateTKEEdgeCluster
- Cloud Native Monitoring APIs
- Scaling group APIs
- Super Node APIs
- Add-on APIs
- Other APIs
- Data Types
- Error Codes
- TKE API 2022-05-01
- FAQs
- Service Agreement
- Contact Us
- Glossary
- User Guide(Old)
Overview
Add-on description
GPU Manager is an all-in-one GPU manager. It is implemented based on the Kubernetes Device Plugin system. This manager provides features such as assigning shared GPU, querying GPU metrics, and preparing GPU-related devices before running a container. It supports your use of GPU devices in Kubernetes clusters.
Add-on features
Topology assignment: provides an assignment feature based on GPU topology. When you assign an application with more than one GPU card, it can select the fastest topology link method to assign the GPU device.
GPU sharing: allows you to submit tasks with less than one GPU card, and supplies QoS assurance.
Querying application GPU metrics: You can access the
/metric path of the CVM port (by default this is 5678) to provide GPU metrics collection feature for Prometheus, and can access the /usage path to perform state querying of readable containers.Kubernetes objects deployed in a cluster
Kubernetes Object Name | Type | Recommended Resource Reservation | Namespaces |
gpu-manager-daemonset | DaemonSet | Each node: 1-core CPU, 1 Gi memory | kube-system |
gpu-quota-admission | Deployment | Each node: 1-core CPU, 1 Gi memory | kube-system |
Use Cases
When the GPU application is running in a Kubernetes cluster, it can prevent the waste of resources caused by requesting a whole card in scenarios such as AI training, allowing full utilization of resources.
Limits
This add-on is implemented through Kubernetes Device Plugin. It can be directly used on clusters of Kubernetes version 1.10 and above.
Each GPU card is split into 100 shares. The usage of GPU can only be a between 0.1 to 0.9, or an integer. VRAM resources are assigned memory with 256MiB as the smallest unit.
To use GPU-Manager, the cluster must contain GPU model nodes.
Directions
Installing the add-on
1. Log in to the TKE console and select Cluster in the left sidebar.
2. On the Cluster Management page, click the ID of the target cluster to go to the cluster details page.
3. In the left sidebar, click Add-on Management to go to the Add-on List page.
4. On the "Add-On List" page, click Create. On the "Create an Add-On" page that appears, select GpuManager.
5. Click Done to complete the process.
Creating fine-grained GPU workloads
After the GpuManager add-on is successfully installed, you can use the following two methods to create fine-grained GPU workloads.
Method 1: creating through the TKE Console
1. Log in to the TKE console and select Clusters in the left sidebar.
2. Select the cluster for which you want to create the GPU application to go to the workload management page, and click Create.
3. On the Create a Workload page, set the configuration as needed. You can configure a fine-grained GPU workload in GPU Resource, as shown below:
Method 2: creating through YAML
Note:
When submitting, use YAML to set the GPU resource usage for the container. The core resources must have
tencent.com/vcuda-core entered in "resource". The VRAM resources must have tencent.com/vcuda-memory entered in "resource".A YAML example is as follows:
P4 device using 1 card:
apiVersion: v1kind: Pod...spec:containers:- name: gpuresources:tencent.com/vcuda-core: 100
5GiB VRAM application using 0.3 cards:
apiVersion: v1kind: Pod...spec:containers:- name: gpuresources:tencent.com/vcuda-core: 30tencent.com/vcuda-memory: 20
Yes
No
Was this page helpful?