AWS parallelcluster documentation change
Summary
Expanded AllocationStrategy options, added subnet priority documentation, added GPU health check limitations, and removed GPUDirect RDMA note
Security assessment
Changes primarily relate to feature enhancements (new allocation strategies) and operational guidance (GPU health check limitations). The GPU health check warning about high memory instances appears to address stability/performance rather than security. No explicit security vulnerabilities or protections are mentioned.
Diff
diff --git a/parallelcluster/latest/ug/Scheduling-v3.md b/parallelcluster/latest/ug/Scheduling-v3.md index 50e7362f9..f4d461ab4 100644 --- a//parallelcluster/latest/ug/Scheduling-v3.md +++ b//parallelcluster/latest/ug/Scheduling-v3.md @@ -585 +585,7 @@ Specify the allocation strategy for all the compute resources defined in Instanc -Valid values: `lowest-price` | `capacity-optimized` | `price-capacity-optimized` +Valid values: `lowest-price` | `capacity-optimized` | `price-capacity-optimized` | `prioritized` | `capacity-optimized-prioritized` + +CapacityType | Allowed strategies +---|--- +ONDEMAND | lowest-price, prioritized +SPOT | lowest-price, capacity-optimized, price-capacity-optimized, capacity-optimized-prioritized +CAPACITY_BLOCK | Not supported — AllocationStrategy cannot be configured @@ -620,0 +627,20 @@ Default: `lowest-price` +**`prioritized`** + + + * If you set `CapacityType = ONDEMAND`, Amazon EC2 Fleet honors the priority order that AWS ParallelCluster applies to the LaunchTemplate overrides when multiple subnets are specified. AWS ParallelCluster derives the override `priority` from the position of the target subnet in `SlurmQueues/Networking/SubnetIds` with the first Subnet getting the highest priority. The priorities are drived by AWS ParallelCluster in descending order from `SlurmQueues/Networking/SubnetIds`, with the first SubnetId having the highest priority and the last SubnetID having the lowest priority. + + * If you set `CapacityType = SPOT`, `prioritized` isn't available. + + + + +**`capacity-optimized-prioritized`** + + + * If you set `CapacityType = ONDEMAND`, `capacity-optimized-prioritized` isn't available. + + * If you set `CapacityType = SPOT`, Amazon EC2 Fleet optimizes for capacity first and then applies, on a best-effort basis, the priority order that AWS ParallelCluster assigns to LaunchTemplate overrides. The priorities are drived by AWS ParallelCluster in descending order from `SlurmQueues/Networking/SubnetIds`, with the first SubnetId having the highest priority and the last SubnetID having the lowest priority. All overrides that target the same subnet receive the same priority value. + + + + @@ -626,0 +653,2 @@ Default: `lowest-price` +**New in 3.14.0** : `prioritized` (for On-Demand) and `capacity-optimized-prioritized` (for Spot). + @@ -707,0 +736,2 @@ AWS ParallelCluster doesn't support `HealthChecks` / `Gpu` in nodes that use `al +Enabling GPU health checks when using instance types whose total GPU memory size is higher than 327680 MiB is discouraged. + @@ -721 +751 @@ Whether AWS ParallelCluster performs GPU health checks on compute nodes. The def - * The duration of the `Gpu` health check depends on the selected instance type, the number of GPUs in the instance, the total GPU memory and the number of `Gpu` health check targets (equivalent to the number of job GPU targets). For example, on a p4d.24xlarge, the typical duration is 3 minutes, whereas on a p6-b200.48xlarge it can take up to 10 minutes. + * The duration of the `Gpu` health check depends on the selected instance type, the number of GPUs in the instance, the total GPU memory and the number of `Gpu` health check targets (equivalent to the number of job GPU targets). For example, on a p4d.24xlarge, the typical duration is 3 minutes. @@ -1394,4 +1423,0 @@ If you're defining a custom security group in SecurityGroups, make sure that you -###### Note - -AWS ParallelCluster version 3.0.0 through 3.0.1: Support for GPUDirect RDMA is enabled for Slurm compute resources. Support for GPUDirect RDMA is supported by specific instance types (`p4d.24xlarge`) on specific operating systems ([Os](./Image-v3.html#yaml-Image-Os) is `alinux2`, `ubuntu1804`, or `ubuntu2004`). The default value is false. -