What Are the Differences Between Public Resource Pools and Dedicated Resource Pools in ModelArts?

• Dedicated resource pools provide dedicated compute resources, which can be used for training jobs and model deployment. Dedicated resource pools deliver higher efficiency and cannot be shared with other users. Before using a dedicated resource pool, you must create one. When you create training jobs or deploy models on ModelArts, instances use the compute resources from the dedicated resource pool.

  Table 1 Description of dedicated resource pools

  Physical Resource Pool	Logical Resource Pool	Resource Queue (Original: Logical Subpool)
  Provides users with an independent compute cluster and network. Dedicated resource pools of different users are physically isolated from each other. Users submit requests using physical specifications and enjoy an exclusive, physically isolated dedicated resource pool. At the underlying level, each pool corresponds to a dedicated cluster. Its advantages include stronger isolation, physical isolation, a dedicated network, and support for network connectivity. Physical resource pools do not support elastic resources.	A logical resource pool is a logical quota. Users use logical flavors to deliver requests. A large underlying parent pool exists beneath the logical resource pools. This parent pool is a physical pool labeled os.modelarts.pool/schedulable: "" preconfigured by the administrator under the service provider tenant. The logical pool uses this parent pool for scheduling and matching based on resource quantity, specifications, and required job types. Upper-layer users are not aware of the parent pool and therefore cannot specify scheduling to a parent pool. They do not have exclusive ownership and only have the right to use this portion of resources. Its advantages are faster creation, scaling out, and scaling in, and support for elastic resources. Logical resource pools do not support connectivity with the tenant VPC network.	You can request resources from your own physical resource pools and create a resource queue with a parent pool that is a user-exclusive dedicated resource pool. The parent pool of a resource queue is the physical resource pool you own and is visible to you. To create resource queues, specify the parent pool specifications. You can also split resource queues from a ModelArts physical resource pool and set quotas, such as reserved and upper limits, for each queue. After splitting out a resource queue, submit training and inference tasks to it. These tasks must stay within the queue's quota to run.
  Summary of resource pool types: Logical resource pools and resource queues are abstract concepts that map to underlying physical resource pools, which serve as their parent entities. Users interact with both physical and logical resource pools at the same level. A physical resource pool is a physically isolated environment that provides exclusive resources for tenants. At the underlying level, each pool corresponds to a dedicated cluster. However, provisioning and capacity expansion in this pool are relatively slow. A logical resource pool is a quota-based environment that shares resources among tenants rather than providing exclusive access. It can be partitioned into smaller units, such as cards. At the underlying level, it does not correspond to a dedicated cluster, allowing for faster provisioning and capacity expansion. Users create resource queues from physical resource pools to allocate resources to teams or groups. Administrators manage these queues centrally.

  Figure 1 Relationships between logical resource pools, resource queues, and physical resource pools
  
  Figure 2 Logical resource pool
  

  Notes:

  You cannot select a parent pool before creating a logical resource pool. The arrows in the figure indicate which parent pools can be scheduled to the logical pool. Note that resource quantity verification is omitted here for display purposes.

  Example:

  Logical resource pool 1: can be scheduled to physical resource pool 1.

  Logical resource pool 2: can be scheduled to physical resource pools 1 and 2.

  Logical resource pool 3: can be scheduled to physical resource pool 3.

  Figure 3 Resource queue principles
  

  Notes:

  The minimum quota min guarantees the lowest amount of resources for the logical subpool, and max indicates the upper limit of resources available for the logical subpool.

  The sum of the minimum quotas of all resource queues must not exceed the number of nodes with the same flavor in the parent pool: min1 + min2 + min3 ≤ x.

  The maximum quota of each resource queue must not exceed the number of nodes with the same flavor in the parent pool: max1 ≤ x, max2 ≤ x, max3 ≤ x.

• Public resource pool: It provides large-scale public computing clusters, which are allocated based on job parameter settings. Resources are isolated by job. You can use ModelArts public resource pools to deliver training jobs, deploy models, or run DevEnviron instances and will be billed on a pay-per-use basis.

Differences between dedicated resource pools and public resource pools:

• Dedicated resource pools provide dedicated computing clusters and network resources for users. The dedicated resource pools of different users are physically isolated, while public resource pools are only logically isolated. Compared with public resource pools, dedicated resource pools feature better performance in isolation and security.
• When a dedicated resource pool is used for creating jobs and the resources are sufficient, the jobs will not be queued. When a public resource pool is used for creating jobs, there is a high probability that the jobs will be queued.
• A dedicated resource pool is accessible to your network. All running jobs in the pool can access storage and resources in your network. For example, if you select a dedicated resource pool with an accessible network when creating a training job, you can access SFS data after the training job is created.
• Dedicated resource pools allow you to customize the runtime environment of physical nodes, for example, you can upgrade GPU or NPU drivers. This function is not supported by public resource pools.