GPU-HPN Capacity Reservation-Container Compute Service(ACS)-阿里云帮助中心

A GPU-HPN capacity reservation for an ACS cluster reserves GPU compute resources that support High-Performance Network (HPN). By associating a GPU-HPN capacity reservation with a virtual node, you can use affinity-based scheduling for GPU container workloads to ensure efficient resource utilization.

Background information

ACS generates the default Virtual Node based on the vSwitchIds in the acs-profile ConfigMap, and it does not consume any computing resources by itself.

GPU-HPN capacity reservation

GPU-HPN capacity reservations are available only through subscriptions. Once active, the resources are guaranteed for the entire subscription period. If you reserve multiple instances, you can associate them with multiple clusters simultaneously. Reserved nodes support Pod instances of different quantities and specifications. For example, in scenarios that require HPN for distributed GPU communication, such as large model training or fine-tuning, Pods consume a GPU-HPN capacity reservation based on the number of whole GPUs specified in the Pod, without constraints on CPU or memory size.

Billing and limitations

Capacity reservations are billed according to your subscription plan.

Billing starts when the reservation is successfully created and its status becomes active.

Billing stops after the reservation expires and the resources are automatically released.

Important

To enable GPU-HPN Capacity Reservation, submit a ticket.
An ACS Cluster only supports applying a reservation to offset costs for the same GPU resource type. If you purchase a reservation for one GPU model, you cannot apply it to a different GPU model. In this case, ACS bills you at the standard rate for the actual GPU type used.
Only Pods with the compute type set to high-performance network GPU (gpu-hpn) can use nodes from a GPU-HPN Capacity Reservation.

Create a GPU-HPN capacity reservation

Log on to the ACS console. In the left navigation pane, choose Capacity Reservation.

On the Capacity Reservations page, click Create GPU-HPN resource reservation and configure the following parameters.

Parameter	Description
Region	Select the region for the capacity reservation.
Zone	Select the zone for the capacity reservation.
Type	Currently, only instance is supported.
Category	Currently, only GPU is supported.
HPN zone	Select the HPN zone for the capacity reservation.
Instance type	The available options vary based on the console display.
Subscription duration	1 month, 1 year, or 3 years.
Down payment ratio (%)	A 0% down payment is required. You can use the resources immediately after purchase.
Discount	Discounts vary based on the subscription duration. A 1-month subscription has no discount. A 1-year subscription has a 12-month prepay discount. A 3-year subscription has a 36-month prepay discount.
Billing cycle	Monthly.
Number of installments	This depends on the subscription duration. For example, a 1-month subscription is paid in 1 installment, and a 1-year subscription is paid in 12 installments.
Quantity	The number of instances to reserve.

After you configure the parameters, click Buy Now. On the order confirmation page, click Proceed to Payment. On the payment page, click Subscribe to complete the purchase.

Associate a cluster

Log on to the ACS console. In the left navigation pane, choose Capacity Reservation.
On the Capacity Reservations page, find the reservation to associate and click Associate Cluster below its status.

Important
You can only associate reservations with an ACS Cluster, ACK Managed Cluster, ACK One Registered Cluster, or ACK One Distributed Workflow Argo Cluster. Other cluster types are not supported.
In the Resource Association dialog box, select or enter the cluster ID and specify the number of instances. Then, click OK.

Important
In the Resource Association dialog box, the cluster drop-down list displays only the IDs of ACS Clusters and ACK Managed Clusters. For an ACK One Registered Cluster or ACK One Distributed Workflow Argo Cluster, you must enter the cluster ID manually.

Query allocated resources

A GPU-HPN capacity reservation is represented as a Kubernetes node in the cluster. Use kubectl to view the resources allocated on the node.

Run the following command to view GPU-HPN nodes by label:

kubectl get node -l alibabacloud.com/node-type=reserved

Expected output:

NAME                                      STATUS   ROLES   AGE   VERSION
cn-wulanchabu-c.cr-rkccqmu0xz8rea1*****   Ready    agent   20m   v1.28.3-aliyun

Run the following command to query the resources allocated from the GPU-HPN Capacity Reservation. Replace the node name in the command with the actual name of the GPU-HPN node from the previous step.

kubectl describe node cn-wulanchabu-c.cr-rkccqmu0xz8rea1***** | grep Allocated -A 10

Expected output:

Allocated resources:
  (Total limits may be over 100 percent, i.e., overcommitted.)
  Resource              Requests    Limits
  --------              --------    ------
  cpu                   16 (8%)     16 (8%)
  memory                128Gi (7%)  128Gi (7%)
  ephemeral-storage     30Gi (0%)   30Gi (0%)
  nvidia.com/gpu        1           1
Events:                 <none>

Lines 5 to 8 show the allocated amounts and utilization for CPU, memory, ephemeral-storage, and GPUs, respectively.