Add a constant for supported NVIDIA OpenRM Architecture and the unsup…

…port instance types

CHANGELOG.md

@@ -19,7 +19,8 @@ CHANGELOG
 - Add support for Python 3.11, 3.12 in pcluster CLI and aws-parallelcluster-batch-cli.
 - Upgrade Python to version 3.12 and NodeJS to version 18 in ParallelCluster Lambda Layer.
 - Build network interfaces using network card index from `NetworkCardIndex` list of EC2 DescribeInstances response, 
-  instead of looping over `MaximumNetworkCards` range.  
+  instead of looping over `MaximumNetworkCards` range. 
+- Fail cluster creation when using instance types p3, g3, p2 and g2 because their GPU architecture is not compatible with Open Source Nvidia Drivers (OpenRM) introduced as part of 3.8.0 release.
 
 3.8.0
 ------
@@ -47,7 +48,7 @@ CHANGELOG
 - Upgrade NVIDIA driver to version 535.129.03.
 - Upgrade CUDA Toolkit to version 12.2.2.
 - Use Open Source NVIDIA GPU drivers (OpenRM) as NVIDIA kernel module for Linux instead of NVIDIA closed source module.
-  - This change removes support for p3, p2, g3, and g2 instances with gpu architecture not supported by OpenRM.
+  - This change removes support for p3, g3, p2 and g2 instances with GPU architecture not supported by OpenRM. The open source driver only works on platforms that have the GSP (GPU System Processor).
 - Remove support of `all_or_nothing_batch` configuration parameter in the Slurm resume program, in favor of the new `Scheduling/ScalingStrategy` cluster configuration.
 - Changed cluster alarms naming convention to '[cluster-name]-[component-name]-[metric]'.
 - Change default EBS volume types in ADC regions from `gp2` to `gp3`, for both the root and additional volumes.

cli/src/pcluster/constants.py

@@ -31,6 +31,7 @@
 DELETION_POLICIES_WITH_SNAPSHOT = DELETION_POLICIES + ["Snapshot"]
 SUPPORTED_ARCHITECTURES = ["x86_64", "arm64"]
 SUPPORTED_OSES_FOR_ARCHITECTURE = {"x86_64": SUPPORTED_OSES, "arm64": SUPPORTED_OSES}
+NVIDIA_OPENRM_UNSUPPORTED_INSTANCE_TYPES = ["p3", "p3dn", "p2", "g3", "g3s", "g2"]
 SLURM = "slurm"
 AWSBATCH = "awsbatch"
 

cli/src/pcluster/validators/ec2_validators.py

@@ -18,6 +18,7 @@
 from pcluster.aws.aws_resources import CapacityReservationInfo
 from pcluster.aws.common import AWSClientError
 from pcluster.config.common import CapacityType
+from pcluster.constants import NVIDIA_OPENRM_UNSUPPORTED_INSTANCE_TYPES
 from pcluster.utils import get_resource_name_from_resource_arn
 from pcluster.validators.common import FailureLevel, Validator
 
@@ -146,15 +147,17 @@ def _validate(self, instance_type: str, image: str):
                     ),
                     FailureLevel.ERROR,
                 )
-        unsupported = ["p3", "p2", "g3", "g2"]
+
         if (
             image_info
             and "AWS ParallelCluster AMI" in image_info.description
-            and instance_type.split(".")[0] in unsupported
+            and instance_type.split(".")[0] in NVIDIA_OPENRM_UNSUPPORTED_INSTANCE_TYPES
         ):
             self._add_failure(
-                f"The instance type '{instance_type}' is not supported by OpenRM drivers. "
-                f"A custom AMI must be used.",
+                f"The instance type '{instance_type}' is not supported by NVIDIA OpenRM drivers. "
+                f"OpenRM can only be used on any Turing or later GPU architectures. "
+                f"Please consider using a different instance type or building a custom AMI "
+                f"with closed source NVIDIA drivers.",
                 FailureLevel.ERROR,
             )
 

cli/tests/pcluster/validators/test_ec2_validators.py

@@ -322,7 +322,10 @@ def test_instance_type_memory_info_validator(mocker, instance_type, instance_typ
         (
             "p3.2xlarge",
             "ami-0185634c5a8a37250",
-            "The instance type 'p3.2xlarge' is not supported by OpenRM drivers. A custom AMI must be used.",
+            "The instance type 'p3.2xlarge' is not supported by NVIDIA OpenRM drivers. "
+            "Only Ampere, Turing, and Hopper architectures are supported. "
+            "Please consider using a different instance type or building a "
+            "custom AMI with closed source NVIDIA drivers.",
             {
                 "ImageId": "ami-0185634c5a8a37250",
                 "Architecture": "x86_64",