HPC Setup for ANSYS CFX Simulations

Hello @Justin Denne

Yes, your understanding is basically correct. Azure CycleCloud is a good fit for running ANSYS CFX as a traditional HPC workload, especially when you want multiple compute nodes, a scheduler, and automatic scaling. CycleCloud itself isn't the job scheduler; it provisions and manages the Azure infrastructure and integrates with schedulers such as Slurm, PBS Pro, and LSF.

For ANSYS CFX, I would start with an architecture like:

Your workstation
      |
      | Upload CFX case / submit job
      v
Slurm Head/Login Node
      |
      | MPI job
      v
+-------------------------------+
| Azure HPC Compute Nodes       |
| HB-series / other HPC VMs     |
| RDMA / InfiniBand             |
+-------------------------------+
      |
      v
Shared HPC Storage

CFX is a computational fluid dynamics workload, so for larger parallel simulations, consider Azure's HB-series HPC VMs rather than deploying ordinary general-purpose VMs. Current HB families are specifically optimized for workloads including computational fluid dynamics, and Azure recommends low-latency InfiniBand-enabled HPC VMs for tightly coupled MPI applications.

A reasonable proof-of-concept path would be:

1. Deploy Azure CycleCloud Workspace for Slurm.
2. Create a small Slurm cluster with a head/login node and initially one or two HPC compute nodes.
3. Install the supported ANSYS/CFX version on the compute image or configure its installation through the CycleCloud cluster configuration.
4. Configure your ANSYS licensing. This is important. Azure provides the infrastructure, but your ANSYS licensing must permit the required HPC execution/cores.
5. Configure shared storage for the CFX input/output files. Depending on performance requirements, this could be NFS initially and Azure Managed Lustre or Azure NetApp Files for larger environments.
6. Submit CFX through Slurm and benchmark it first on one node, then multiple nodes before scaling further.

CycleCloud can automatically provision compute nodes when jobs enter the scheduler queue and remove them when they're no longer required, which is one of its main advantages for expensive HPC VM sizes.

Since you're using a trial Azure account, however, check your VM quota before building the cluster. HPC VM families aren't necessarily available in every region or subscription, and Microsoft notes that HPC capacity can be constrained even when quota exists.

I also wouldn't build a large environment immediately. For someone moving from a local CFX workstation, I'd first create a small Slurm + CycleCloud proof of concept, validate ANSYS licensing and MPI operation, run the same CFX model locally and in Azure, and compare solver time and cost. Once that works, increase the node count.

Microsoft's documentation is a good starting point:

Azure CycleCloud documentation

Planning and sizing an Azure HPC cluster

Azure HPC architecture guidance

Deploying CycleCloud for HPC orchestration

For your ANSYS CFX 2021 installation specifically, I'd also verify its supported Linux/Windows OS, MPI version, and Azure/cloud licensing requirements with ANSYS before choosing the CycleCloud node image. That will determine the exact cluster configuration.

Please "Accept the Answer" if this information helped you. This will help us and others in the community.