Using Amazon EC2 Capacity Blocks for ML with AWS PCS - AWS PCS
LimitationsCapacity Block expiration

Using Amazon EC2 Capacity Blocks for ML with AWS PCS

Amazon EC2 Capacity Blocks for ML is an Amazon EC2 purchasing option that enables you to pay in advance to reserve GPU-based accelerated computing instances within a specific date and time range to support short duration workloads. Instances that run inside a Capacity Block are automatically placed close together inside Amazon EC2 UltraClusters, for low-latency, petabit-scale, non-blocking networking. For more information, see Capacity Blocks for ML in the Amazon Elastic Compute Cloud User Guide.

You can use a launch template to have AWS PCS use a Capacity Block when it launches instances for a compute node group.

Note

AWS PCS only supports Capacity Blocks with Slurm version 24.05 or later.

Limitations

  • AWS PCS only supports Capacity Blocks with P5en, P5e, P5, and P4d instance families.

  • You can only associate a compute node group with 1 Capacity Block at a time.

  • You can't associate a compute node group with a capacity reservation group that combines multiple Capacity Blocks.

  • Capacity Blocks must be in a scheduled or active state to use with AWS PCS. You can't use Capacity Blocks in other states, such as payment-failed. For more information, see View Capacity Blocks in the Amazon Elastic Compute Cloud User Guide.

  • AWS PCS doesn't support Capacity Blocks in AWS GovCloud (US).

Capacity Block expiration

Capacity Blocks are limited to a specific date and time range. When a Capacity Block expires:

  • The compute node group associated with that Capacity Block continues to exist and remains associated with the same queues.

  • All instances in the compute node group are terminated and active jobs might fail, based on your Slurm settings.

  • AWS PCS can't launch new instances in the compute node group.

  • All queued or newly submitted jobs remain in pending state until another compute node group is attached to the queue or you update the compute node group to use a new launch template that specifies a new Capacity Block.