RDMA network profiles

This page provides an overview of Remote Direct Memory Access (RDMA) network profiles in Google Cloud.

Overview

RDMA network profiles let you create Virtual Private Cloud (VPC) networks that provide low-latency, high-bandwidth RDMA communication between the memory or GPUs of VMs that are created in the network.

RDMA network profiles are useful for running AI workloads. For more information about running AI workloads in Google Cloud, see AI Hypercomputer overview .

You can create the following types of VPC networks by using RDMA network profiles:

Supported zones

RDMA network profiles are available in a limited set of zones. You can only create a Falcon VPC network or RoCE VPC network in a zone where the corresponding network profile is available.

To view the supported zones, see List network profiles .

Alternatively, you can view the supported zones for the machine type that you intend to create in the network. RDMA network profiles are available in the same zones as their supported machine types. For more information, see the following:

• For GPU machine types, see GPU availability by regions and zones .
• For other machine types, see Available regions and zones .

Specifications

VPC networks created with an RDMA network profile have the following specifications:

• Zonal constraint. Resources using a VPC network with an RDMA network profile are limited to the same zone as the RDMA network profile associated with the VPC network during the network creation. This zonal limit has the following effects:

  • All instances that have network interfaces in the VPC network must be created in the zone that matches the zone of the RDMA network profile used by the VPC network.

  • All subnets created in the VPC network must be located in the region that contains the zone of the RDMA network profile used by the VPC network.

• RDMA network interfaces only. A VPC network with an RDMA network profile supports attachments only from specific network interfaces:

  • Falcon VPC networks only support IRDMA network interfaces (NICs), which are only available on the H4D machine series.
  • RoCE VPC networks only support MRDMA NICs, which are only available on the A3 Ultra , A4 , and A4X machine series.

  All non-RDMA NICs of a virtual machine (VM) instance must be attached to a regular VPC network.

• 8896 byte MTU. For best performance, we recommend a maximum transmission unit (MTU) of 8896 bytes for VPC networks with an RDMA network profile. This allows the RDMA driver in the VM's guest operating system to use smaller MTUs if needed.

  If you create a VPC network with an RDMA network profile by using the gcloud CLI or the API, then 8896 bytes is the default MTU. If you use the Google Cloud console, then you must set the MTU to 8896 .

• Firewall differences. See the following information about firewall differences in VPC networks with an RDMA network profile:

  • VPC networks with an RDMA network profile use the following implied firewall rules, which are different from the implied firewall rules used by regular VPC networks:

    • Implied allow egress
    • Implied allow ingress

  • Cloud NGFW support depends on the type of VPC network:

    • RoCE VPC networks only support regional network firewall policies that have an RoCE firewall policy type. The set of parameters for rules within a supported regional network firewall policy are limited. For more information, see Cloud NGFW for RoCE VPC networks .
    • Falcon VPC networks don't support configuring Cloud NGFW rules or policies.

• No Connectivity Tests support. Connectivity Tests doesn't support VPC networks with an RDMA network profile.

• Other VPC features. VPC networks with an RDMA network profile support a limited set of other VPC features. For more information, see the following Supported and unsupported features section.

Supported and unsupported features

The following table lists which VPC features are supported by VPC networks with an RDMA network profile.

RoCE VPC network multi-NIC considerations

To support workloads that benefit from cross-rail GPU-to-GPU communication, RoCE VPC networks support VMs that have multiple MRDMA NICs in the network. Each MRDMA NIC must be in a unique subnet. Placing two or more MRDMA NICs in the same RoCE VPC network might affect network performance, including increased latency. MRDMA NICs use NCCL . NCCL attempts to align all network transfers, even for cross-rail communication. For example, it uses PXN to copy data through NVlink to a rail-aligned GPU before transferring it over the network.

What's next

• Network profiles for specific use cases
• Create a VPC network for RDMA NICs
• Cloud NGFW for RoCE VPC networks