In Elastic Algorithm Service (EAS), resource groups are used to isolate cluster resources. When you create an online model service, you can choose to deploy the service in a public resource group or dedicated subscription resource group.
The following figure compares the features of public resource groups and dedicated subscription resource groups.
|Public resource groups||Dedicated subscription resource groups|
|Pay-as-you-go (post-payment)||Subscription (pre-payment)||Pay-as-you-go (post-payment)|
Public resource groups
Public resource groups provide CPU resources.
1. User guide
When you deploy a model on the page, select a public resource group. Then, the system deploys the model to the public resource group. For more information on how to deploy a model in EAS, see How to deploy a model in EAS.
When you deploy a model in EASCMD, you need to set the resource field to CPU.
2. Billing and stopOnce a model service is deployed and is running, billing starts. For more information, see Billing of deployed public resource groups. Remember to stop unused model services promptly to avoid being charged.
You can stop billing as follows:Click Stop in the Actions column of the service in the service list to deactivate the service and stop billing. Note: Make sure that the deactivated service is no longer used to avoid unnecessary business losses.
Dedicated subscription resource groups
- Subscription is provided to reduce costs.
- Dedicated subscription resource groups support processor customization. This enables support for more types of models.
- You can connect a dedicated subscription resource group to your Virtual Private Cloud (VPC) network. This maximizes the efficiency and minimizes the costs when you call your online model services.
You can create a dedicated subscription resource group and deploy online model services in this resource group. The resource group is hosted by EAS.
1. Basic informationThe following table lists the basic information of dedicated subscription resource groups.
|Type||Subscription resource group|
|Operations on resource groups||Scale-out and renewal|
2. User guide
(1) Create a resource group
- Log on to the EAS console and then click New Resource Group.
- You are then directed to the purchase page. Select a resource type, quantity, and subscription duration.
- After you pay for the order, the system starts allocating resources and deploying the resource group. You can then go back to the console to check the status of the resource group.
Note: Due to issues such as insufficient resources in the current region, the system may fail to create the resource group. Your payment for the subscription will be refunded to your account.
- You can check all resource groups created under your account and the relevant resource group information on the EAS page. You can modify a resource group name so that it can be used for identification.
- After the resource group is created, you can then select the resource group when you deploy an online model service.
(2) Resource group details page
You can perform the following operations on the resource group details page:
- View the resource group ID. The resource group ID is the unique identifier of a resource group, which is required for EASCMD deployment.
- Modify and view the resource group name for easy identification and memory. You can use a string that indicates a service. When deploying model services on a web page, you need to select this resource group based on this name.
- Check the status of the servers in the resource group, and the resources provided by each server.
- Check all services deployed in the current resource group and manage these services.
- Scale out the resource group, renew the resource group, and activate VPC direct connect.
(3) Operations on resource groups
- Scale out: You can scale out the number of servers in a resource group as long as your subscription has not expired.
- Renew (only applicable to subscription resource groups): You can increase the subscription duration of a resource group.
- Activate VPC direct connect: You can connect a resource group to your VPC network in the same region through Elastic Network Interface (ENIs). For more information, see Connect a resource group to a VPC network.
(4) RAM user authorization
You can use an Alibaba Cloud account to view, access, create, and purchase resource groups, and perform operations on resource groups.
If a Resource Access Management (RAM) user needs to perform the preceding operations, the RAM user must be authorized by the Alibaba Cloud user. Otherwise, the RAM user cannot perform the preceding operations.For more information about permission definitions and authorization methods for RAM users, see EAS permission management.
In particular, to create resource groups, a RAM user needs both the permissions authorized based on EAS permission management and finance permissions. Otherwise, the RAM user cannot place an order for payment when purchasing a resource group. For more information, see 3. Other considerations in EAS permission management.
3. Resource types of dedicated subscription resource groups
Currently, 14 resource types can be used to create resource groups, which vary depending on regions. The following table lists the server model for each resource type and the GPU and CPU configurations corresponding to this type.
|Server model||GPU configuration||CPU configuration|
|ecs.c5.6xlarge||/||24 cores, 48 GB|
|ecs.g5.6xlarge||/||24 cores, 96 GB|
|ecs.gn5i-c4g1.xlarge||1 * NVIDIA Tesla P4||4 cores, 16 GB|
|ecs.gn5i-c8g1.2xlarge||1 * NVIDIA Tesla P4||8 cores, 32 GB|
|ecs.gn6i-c4g1.xlarge||1 * NVIDIA Tesla T4||4 cores, 15 GB|
|ecs.gn6i-c8g1.2xlarge||1 * NVIDIA Tesla T4||8 cores, 31 GB|
|ecs.gn6i-c16g1.4xlarge||1 * NVIDIA Tesla T4||16 cores, 62 GB|
|ecs.gn6i-c24g1.6xlarge||1 * NVIDIA Tesla T4||24 cores, 93 GB|
|ecs.gn5-c4g1.xlarge||1 * NVIDIA P100||4 cores, 30 GB|
|ecs.gn5-c8g1.2xlarge||1 * NVIDIA P100||8 cores, 60 GB|
|ecs.gn5-c28g1.7xlarge||1 * NVIDIA P100||28 cores, 112 GB|
|ecs.gn6v-c8g1.2xlarge||1 * NVIDIA V100||8 cores, 32 GB|
For more information about the pricing, see EAS billing.