PAI AI computing resources provide high-availability, high-performance GPU capacity for training and inference workloads. These resources power DSW, DLC, and EAS. Purchase computing resources in a resource pool, allocate capacity to teams through resource quotas (Quotas), and bind quotas to workspaces before use.
Resource types
Cloud-native resources
Cloud-native resources apply to PAI-DSW, PAI-DLC, and PAI-EAS. They come in two types:
|
Type |
Description |
Use cases |
|
General-purpose computing resources |
Built on cloud products such as ECS, ECI, and EGS, these resources provide flexible and stable capacity. After you activate PAI, the system automatically creates a public resource quota for immediate use. |
Routine model training and inference, and small-to-medium-scale development and experimentation. |
|
Lingjun AI computing resources |
High-performance capacity designed for large-scale deep learning. Hardware-software co-optimization delivers high-performance network interconnect and storage acceleration. |
Large model training, thousand-GPU-scale distributed training, and workloads that demand high network and storage bandwidth. |
Big data engine resources
Big data engine resources apply to PAI-Designer. They include two types:
|
Type |
Description |
References |
|
MaxCompute |
An enterprise-grade cloud data warehouse that supports efficient analysis and processing of massive datasets. |
|
|
Realtime Compute for Apache Flink |
A real-time big data analytics platform built on Apache Flink with sub-second latency. |
Key concepts
|
Concept |
Description |
|
Resource pool |
The entry point for purchasing and managing computing resources. In a resource pool, you create resource groups, purchase GPU nodes, and manage renewals and unsubscriptions. |
|
Resource quota (Quota) |
A unit that allocates capacity from a resource pool. You can create a parent-child quota hierarchy (Quota Tree) to distribute resources by team or project. After you bind a quota to a workspace, DSW, DLC, and EAS in that workspace can use the allocated capacity. |
|
Resource group |
A logical grouping within a resource pool. Each resource group corresponds to a set of computing nodes of the same type. A single resource quota can draw capacity from multiple resource groups. |
Workflow
Follow this workflow to set up cloud-native AI computing resources:
-
Purchase resources: Create a resource group and purchase computing resources in AI Computing Resources > Resource Pool.
General-purpose computing resources: For details, see Create a resource group and purchase general computing resources
Lingjun AI computing resources: For details, see Create a resource group and purchase Lingjun resources
Create resource quotas: Allocate capacity to different teams in AI Computing Resources > Resource Quota (Quota). For details, see Create a resource quota.
Bind to a workspace: Bind a resource quota to a workspace so that workspace members can use the allocated capacity for AI development and training.
Use the resource quota: When creating a DSW instance, DLC task, or EAS service, select the resource quota you created as the compute source.
Big data engine resources use a different quota management approach. For MaxCompute, see MaxCompute resource quotas. For Realtime Compute for Apache Flink, see Manage fully managed Flink resources.
Advanced configurations
Scheduling policies: Configure queue scheduling policies to improve dequeue efficiency and resource utilization. For details, see Scheduling policies.
Preemption policies: Enable capacity preemption so tasks can use idle capacity from sibling or child quotas. For details, see Preemption policy.
Monitoring and alerting: View resource usage through CloudMonitor and ARMS, and configure alert notifications. For details, see Monitoring and alerting.
Integrated training and inference resource management: Run inference services and training tasks on the same GPU cluster to improve overall utilization. For details, see Unified inference and training resource management.
Billing
AI computing resources are billed by subscription (monthly or annually). Fees depend on the node specification, quantity, and subscription duration. For details, see Billing of AI computing resources.