How to activate and manage RAY resource groups - Lindorm

Lindorm’s RAY resource group provides distributed computing services and supports end-to-end AI workload processing. This resource group is fully compatible with the standard RAY computing model and programming interfaces. It integrates deeply with Lindorm’s multi-model storage engines to efficiently handle data pre-processing, training, and inference tasks. This topic describes how to activate, manage, and bill RAY resource groups.

Important

The RAY resource group is currently in invitational preview. To request access, contact Lindorm technical support (DingTalk ID: s0s3eg3).

Prerequisites

You have enabled Lindorm LindormTable.
You have activated Lindorm compute engine.

Billing method

The RAY resource group provides resident mode. Fees include the following two parts:

Resident resource fee: Billed based on the configured Head and Worker node resources using CU billing.
Elastic resource fee: Worker nodes support auto scaling based on workload. Scaled Worker nodes are billed by usage duration using CU billing.

Activate a RAY resource group

Log on to the Lindorm console. In the upper-left corner of the page, select the region of the instance. On the Instances page, click the ID of the target instance or click View Instance Details in the Actions column for the instance.
In the Configurations section of the Instance Details page, click Resource Groups in the Compute Engine operation column.

On the Resource Group Details page, click Create Resource Group and configure the following settings:

Resource Group Type: Select RAY.
Resource Group Name: Enter a name for the resource group. Use only lowercase letters and digits. The name must be no longer than 63 characters. Example: raycg.

Running Mode: The runtime mode of the RAY resource group. You can select Resident or Non-resident.

Resident: The Ray cluster runs in resident mode. Submit RAY jobs to the specified RAY cluster. When no jobs are running, the cluster runs with the minimum resource retention. When jobs are submitted, the cluster dynamically allocates resources based on job demand. Suitable for scenarios that submit jobs frequently and require a stable interactive access address.

If you select the Resident mode, you must configure the Head node and Worker Group settings.

Head node configuration

Configuration item	Parameter description
Head resource type	RAY resource groups support two resource types: CPU and GPU.
Head Resource Specifications	For the CPU resource type, select your CPU and memory quota, such as 4-core 8 GB, 4-core 16 GB, or 8-core 32 GB. Choose based on your cluster size. Default: 4-core 16 GB. For the GPU resource type, availability depends on instance type and inventory. Contact Lindorm technical support (DingTalk ID: s0s3eg3) when you need GPU resources.
Head disk size	Disk space for the Head node. Used to store logs, memory overflow files, and resource files used during job execution. Default: 30 GB.

Worker Group configuration

You can select one or more Worker Groups based on your scenario. Each Worker Group can use a different resource specification.

Configuration item	Parameter description
Worker resource type	RAY resource groups support two resource types: CPU and GPU.
Worker Resource Specifications	For the CPU resource type, select your CPU and memory quota, such as 4-core 8 GB, 4-core 16 GB, or 8-core 32 GB. Choose Worker Group specification based on your actual job requirements. Default: 4-core 16 GB. For the GPU resource type, availability depends on instance type and inventory. Contact Lindorm technical support (DingTalk ID: s0s3eg3) when you need GPU resources.
Worker disk space	Disk space for the Worker node. Used to store logs, memory overflow files, and resource files used during job execution. Default: 30 GB.
Minimum number of workers	Minimum number of running replicas in the Worker Group. When no jobs are running, the group runs with this resource quota.
Maximum number of workers	Maximum number of running replicas in the Worker Group. When jobs are running, this is the maximum number of Worker nodes that can be allocated.

Non-resident (Serverless): Computing resources are allocated and released on demand. A temporary Ray cluster is created automatically for each job submission and reclaimed after the job completes. Suitable for low-frequency batch processing and cost-sensitive scenarios.

If you select the Non-resident (Serverless) mode, configure the following parameters:

Parameter	Description
CPU limit (cores)	Maximum CPU resource usage. Minimum value: 100.
Memory limit (GB)	Maximum memory resource usage. Minimum value: 400.
Authorized users	Users who are allowed to submit jobs to this resource group. `*` indicates all users.

Click OK to create the RAY resource group. Creation takes about 20 minutes.

Manage RAY resource groups

Log on to the Lindorm console. In the upper-left corner of the page, select the region of the instance. On the Instances page, click the ID of the target instance or click View Instance Details in the Actions column for the instance.
In the Configurations section of the Instance Details page, click Resource Groups in the Compute Engine operation column.
On the Resource Group Details page, hover your mouse over WebUI in the Actions column of the RAY resource group to view its WebUI address. Example: http://alb-57k7r581oht8rd****.cn-hangzhou.alb.aliyuncsslb.com/ray/raycg/dashboard/.
In your browser, go to the resource group’s WebUI address to view its runtime status.
- In the navigation bar at the top of the WebUI, switch between tabs to view the job list (Jobs), cluster status (Cluster), actor list (Actors), and cluster logs (Logs).
- On the Cluster tab, view CPU, memory, GPU, Object Store, and other resource usage across all cluster nodes.
(Optional) On the Resource Group Details page, you can also Delete an existing resource group.

Note
RAY resource groups do not currently support modification or restart operations.