High performance computing VM sizes

  • Article

Caution

This article references CentOS, a Linux distribution that is nearing End Of Life (EOL) status. Please consider your use and planning accordingly. For more information, see the CentOS End Of Life guidance.

Applies to: ✔️ Linux VMs ✔️ Windows VMs ✔️ Flexible scale sets ✔️ Uniform scale sets

HBv3-series VMs are optimized for HPC applications such as fluid dynamics, explicit and implicit finite element analysis, weather modeling, seismic processing, reservoir simulation, and RTL simulation. HBv3 VMs feature up to 120 AMD EPYC™ 7003-series (Milan) CPU cores, 448 GB of RAM, and no hyperthreading. HBv3-series VMs also provide 350 GB/sec of memory bandwidth, up to 32 MB of L3 cache per core, up to 7 GB/s of block device SSD performance, and clock frequencies up to 3.5 GHz.

All HBv3-series VMs feature 200 Gb/sec HDR InfiniBand from NVIDIA Networking to enable supercomputer-scale MPI workloads. These VMs are connected in a non-blocking fat tree for optimized and consistent RDMA performance. The HDR InfiniBand fabric also supports Adaptive Routing and the Dynamic Connected Transport (DCT, in addition to standard RC and UD transports). These features enhance application performance, scalability, and consistency, and their usage is strongly recommended.

Note

All HBv3 VMs have exclusive access to the physical servers. There is only 1 VM per physical server and there is no shared multi-tenancy with any other VMs for these VM sizes.

RDMA-capable instances

Most of the HPC VM sizes feature a network interface for remote direct memory access (RDMA) connectivity. This interface is in addition to the standard Azure Ethernet network interface available in the other VM sizes.

This secondary interface allows the RDMA-capable instances to communicate over an InfiniBand (IB) network, operating at HDR rates for HBv3 virtual machines. These RDMA capabilities can boost the scalability and performance of Message Passing Interface (MPI) based applications.

Note

SR-IOV support: In Azure HPC, currently there are two classes of VMs depending on whether they are SR-IOV enabled for InfiniBand. Currently, almost all the newer generation, RDMA-capable or InfiniBand enabled VMs on Azure are SR-IOV enabled. RDMA is only enabled over the InfiniBand (IB) network and is supported for all RDMA-capable VMs. IP over IB is only supported on the SR-IOV enabled VMs. RDMA is not enabled over the Ethernet network.

  • Operating System - Linux distributions such as CentOS, Ubuntu, SUSE are commonly used. Windows Server 2016 and newer versions are supported on all the HPC series VMs. See VM Images for a list of supported VM Images on the Marketplace and how they can be configured appropriately. The respective VM size pages also list out the software stack support.

  • InfiniBand and Drivers - On InfiniBand enabled VMs, the appropriate drivers are required to enable RDMA. See VM Images for a list of supported VM Images on the Marketplace and how they can be configured appropriately. Also see enabling InfiniBand to learn about VM extensions or manual installation of InfiniBand drivers.

  • MPI - The SR-IOV enabled VM sizes on Azure allow almost any flavor of MPI to be used with Mellanox OFED.

    Note

    RDMA network address space: The RDMA network in Azure reserves the address space 172.16.0.0/16. To run MPI applications on instances deployed in an Azure virtual network, make sure that the virtual network address space does not overlap the RDMA network.

Cluster configuration options

Azure provides several options to create clusters of HPC VMs that can communicate using the RDMA network, including:

  • Virtual machines - Deploy the RDMA-capable HPC VMs in the same scale set or availability set (when you use the Azure Resource Manager deployment model). If you use the classic deployment model, deploy the VMs in the same cloud service.

  • Virtual machine scale sets - In a virtual machine scale set, ensure that you limit the deployment to a single placement group for InfiniBand communication within the scale set. For example, in a Resource Manager template, set the singlePlacementGroup property to true. Note that the maximum scale set size that can be spun up with singlePlacementGroup=true is capped at 100 VMs by default. If your HPC job scale needs are higher than 100 VMs in a single tenant, you may request an increase, open an online customer support request at no charge. The limit on the number of VMs in a single scale set can be increased to 300. Note that when deploying VMs using Availability Sets the maximum limit is at 200 VMs per Availability Set.

    Note

    MPI among virtual machines: If RDMA (e.g. using MPI communication) is required between virtual machines (VMs), ensure that the VMs are in the same virtual machine scale set or availability set.

Deployment considerations

  • Azure subscription - To deploy more than a few compute-intensive instances, consider a pay-as-you-go subscription or other purchase options. If you're using an Azure Trial, you can use only a limited number of Azure compute cores.

  • Pricing and availability - Check VM pricing and availability by Azure regions.

  • Cores quota - You might need to increase the cores quota in your Azure subscription from the default value. Your subscription might also limit the number of cores you can deploy in certain VM size families, including the H-series. To request a quota increase, open an online customer support request at no charge. (Default limits may vary depending on your subscription category.)

    Note

    Contact Azure Support if you have large-scale capacity needs. Azure quotas are credit limits, not capacity guarantees. Regardless of your quota, you are only charged for cores that you use.

  • Virtual network - An Azure virtual network is not required to use the compute-intensive instances. However, for many deployments you need at least a cloud-based Azure virtual network, or a site-to-site connection if you need to access on-premises resources. When needed, create a new virtual network to deploy the instances. Adding compute-intensive VMs to a virtual network in an affinity group is not supported.

  • Resizing - Because of their specialized hardware, you can only resize compute-intensive instances within the same size family (H-series or N-series). For example, you can only resize an H-series VM from one H-series size to another. Additional considerations around InfiniBand driver support and NVMe disks may need to be considered for certain VMs.

Other sizes

Next steps