Instance Overview

A CES instance is a computing resource in the GPU Cloud that you can log in to and use directly. Like a server, you can prepare the runtime environment, access data, and run training or inference tasks on an instance. Depending on the task status, you can start, stop, restart, renew, or delete the instance. If your tasks require direct control of the OS environment, temporary debugging, running standalone tasks, or managing a small number of fixed resources, prioritize CES instances. If you need to run containerized workloads over the long term, centrally schedule multiple nodes, or manage applications using Kubernetes-native capabilities, CKS clusters are recommended.

Select Instance Type

CES instances are classified as GPU Instances and CPU Instances. Before creating one, select an instance type based on whether the task requires GPU acceleration.

Instance Type	Applicable Scenarios	Resource Types Available at Creation
GPU Instance	Deep learning training, model fine-tuning, inference services, and GPU-accelerated computing.	Bare metal.
CPU Instance	General computing tasks such as data preprocessing, scheduling services, development environments, lightweight services, and file storage clients.	Virtual machines or bare metal.

GPU Instances are suitable for tasks that require direct access to GPU resources. CPU Instances are suitable for tasks that do not rely on GPU acceleration. If a CPU task has requirements for resource scale, hardware isolation, or RDMA networking, you can select a Bare Metal Instance when creating it.

What You Need to Decide Before Creating

Before creating an instance, we recommend confirming the following items to avoid rework caused by mismatched specifications, Logon Credential, or data access methods after creation.

Item	Details to Confirm
Task type	Confirm whether the task requires GPUs, whether it needs to run on a single machine or collaboratively across multiple nodes, and whether the instance environment needs to be retained long term.
Billing method	Pay-as-you-go is suitable for short-term tasks or tasks with an uncertain duration; subscription is suitable for resources intended for long-term, stable use.
Region and network	The instance is created in a specified region and requires you to select its network. When subsequently using an EIP, NAT Gateway, or VPC, ensure that it matches the region and network associated with the instance.
Instance specification	The specification determines information such as vCPUs, memory, GPUs, RDMA, and price. When selecting a specification, consider the task’s resource requirements, available inventory, and budget.
Image and drivers	The image determines the base operating system. GPU tasks typically also require an appropriate GPU driver version based on the framework, CUDA version, or task requirements.
Logon method	Password login and key pair login are supported. Before selecting key pair login, prepare an available key pair.
Data access	To access shared datasets, model weights, or training outputs, you can plan the GPFS file system mount path in advance. For object-based access, you can use OSS.
Monitoring and alerts	If you need to continuously monitor resource usage or configure alerts, it is recommended that you install monitoring when creating the instance.

When creating instances in batches, if you need to distinguish the node order by HostName, you can select Custom HostName and enable HostName with a custom sequence as needed, making it easier for multi-node tasks to identify node roles and for subsequent operations and maintenance.

How to Use It After Creation

After the instance is created and ready, it is typically used as follows:

Connect to the instance using the Logon Credential selected when creating the instance.
Prepare the runtime environment, for example, by installing dependencies, configuring frameworks, or pulling images.
Prepare the data required for the task. Select an appropriate storage location and complete the necessary preparations based on whether the data needs to be shared between instances, transferred to or from the local environment, or retained long term.
Run common compute tasks such as training, inference, or data processing.
Use Monitor to view resource utilization, use Logs to view task output and troubleshoot errors, and use Alerts to receive timely notifications of resource or task exceptions.
After the task is complete, Stop, Renew, or Delete the instance depending on whether the resource is still needed.

Common Management Tasks

After an instance is created, go to the Instance page and manage it under the GPU Instance or CPU Instance tab. The available actions may vary depending on the status and billing method.

Task	Applicable Scenario	Notes
Start, stop, or restart an instance	Manage the instance power status, resume operation, or temporarily release runtime resources.	Stopping or restarting interrupts running tasks.
Mount or unmount a file system	Enable an instance to access shared data in a GPFS file system.	Before unmounting, ensure that tasks no longer read from or write to the mount path.
Find mount points or manage clients	Check the file system mount status on instances in batches, or manage file storage clients.	Before performing batch operations, confirm the scope of instances.
Reinstall the instance OS	The operating system environment is abnormal and needs to be restored.	The current system reinstallation does not support selecting an image again and deletes system disk data.
Modify deletion protection	Reduce the risk of accidentally deleting an instance.	If deletion protection is enabled, disable it before deleting the instance.
Renew an instance	A Subscription instance is nearing expiration and its usage period needs to be extended.	Renewal incurs charges. Before proceeding, confirm the renewal period and fees.
Delete an instance	The instance is no longer needed.	Before deletion, confirm the tasks, data, mount relationships, and billing impact.

Security and Data Risks

Key pairs and account passwords are both login credentials. Keep them secure, and do not include them in public repositories, images, script templates, or logs.
Data stored locally on an instance or on system disks or data disks may be unrecoverable after deletion, reinstallation, or unexpected operations. Save important data to persistent storage.
Delete, Stop, restart, reinstallation, and batch operations may interrupt tasks. Before proceeding, confirm the task status and scope of impact.
To reduce the risk of accidental deletion, you can enable deletion protection for important instances.

Get Started

Compute

Storage

Cloud networks

Observability

Instance Overview

Select Instance Type

What You Need to Decide Before Creating

How to Use It After Creation

Common Management Tasks

Security and Data Risks

​Select Instance Type

​What You Need to Decide Before Creating

​How to Use It After Creation

​Common Management Tasks

​Security and Data Risks

Select Instance Type

What You Need to Decide Before Creating

How to Use It After Creation

Common Management Tasks

Security and Data Risks