Search

Configuring InfiniBand and RDMA networks

download PDF
Red Hat Enterprise Linux 8

Configuring and managing high-speed network protocols and RDMA hardware

Red Hat Customer Content Services

Abstract

You can configure and manage Remote Directory Memory Access (RDMA) networks and InfiniBand hardware at an enterprise level by using various protocols. These include RDMA over Converged Ethernet (RoCE), the software implementation of RoCE (Soft-RoCE), the IP networks protocol such as iWARP, the software implementation of iWARP (Soft-iWARP), and the Network File System over RDMA (NFSoRDMA) protocol as a native support on RDMA-supported hardware. For low-latency and high-throughput connections, you can configure IP over InfiniBand (IPoIB) and Open Subnet Manager (OpenSM).

Providing feedback on Red Hat documentation

We appreciate your feedback on our documentation. Let us know how we can improve it.

Submitting feedback through Jira (account required)

  1. Log in to the Jira website.
  2. Click Create in the top navigation bar.
  3. Enter a descriptive title in the Summary field.
  4. Enter your suggestion for improvement in the Description field. Include links to the relevant parts of the documentation.
  5. Click Create at the bottom of the dialogue.

Chapter 1. Understanding InfiniBand and RDMA

InfiniBand refers to two distinct things:

  • The physical link-layer protocol for InfiniBand networks
  • The InfiniBand Verbs API, an implementation of the remote direct memory access (RDMA) technology

RDMA provides access between the main memory of two computers without involving an operating system, cache, or storage. Using RDMA, data transfers with high-throughput, low-latency, and low CPU utilization.

In a typical IP data transfer, when an application on one machine sends data to an application on another machine, the following actions happen on the receiving end:

  1. The kernel must receive the data.
  2. The kernel must determine that the data belongs to the application.
  3. The kernel wakes up the application.
  4. The kernel waits for the application to perform a system call into the kernel.
  5. The application copies the data from the internal memory space of the kernel into the buffer provided by the application.

This process means that most network traffic is copied across the main memory of the system if the host adapter uses direct memory access (DMA) or otherwise at least twice. Additionally, the computer executes some context switches to switch between the kernel and application. These context switches can cause a higher CPU load with high traffic rates while slowing down the other tasks.

Unlike traditional IP communication, RDMA communication bypasses the kernel intervention in the communication process. This reduces the CPU overhead. The RDMA protocol enables the host adapter to decide after a packet enters the network which application should receive it and where to store it in the memory space of that application. Instead of sending the packet for processing to the kernel and copying it into the memory of the user application, the host adapter directly places the packet contents in the application buffer. This process requires a separate API, the InfiniBand Verbs API, and applications need to implement the InfiniBand Verbs API to use RDMA.

Red Hat Enterprise Linux supports both the InfiniBand hardware and the InfiniBand Verbs API. Additionally, it supports the following technologies to use the InfiniBand Verbs API on non-InfiniBand hardware:

  • Internet Wide Area RDMA Protocol (iWARP): A network protocol that implements RDMA over IP networks
  • RDMA over Converged Ethernet (RoCE), which is also known as InfiniBand over Ethernet (IBoE): A network protocol that implements RDMA over Ethernet networks

Additional resources

Chapter 2. Configuring RoCE

Remote Direct Memory Access (RDMA) provides remote execution for Direct Memory Access (DMA). RDMA over Converged Ethernet (RoCE) is a network protocol that utilizes RDMA over an Ethernet network. For configuration, RoCE requires specific hardware and some of the hardware vendors are Mellanox, Broadcom, and QLogic.

2.1. Overview of RoCE protocol versions

RoCE is a network protocol that enables remote direct memory access (RDMA) over Ethernet.

The following are the different RoCE versions:

RoCE v1
The RoCE version 1 protocol is an Ethernet link layer protocol with ethertype 0x8915 that enables the communication between any two hosts in the same Ethernet broadcast domain.
RoCE v2
The RoCE version 2 protocol exists on the top of either the UDP over IPv4 or the UDP over IPv6 protocol. For RoCE v2, the UDP destination port number is 4791.

The RDMA_CM sets up a reliable connection between a client and a server for transferring data. RDMA_CM provides an RDMA transport-neutral interface for establishing connections. The communication uses a specific RDMA device and message-based data transfers.

Important

Using different versions like RoCE v2 on the client and RoCE v1 on the server is not supported. In such a case, configure both the server and client to communicate over RoCE v1.

2.2. Temporarily changing the default RoCE version

Using the RoCE v2 protocol on the client and RoCE v1 on the server is not supported. If the hardware in your server supports RoCE v1 only, configure your clients for RoCE v1 to communicate with the server. For example, you can configure a client that uses the mlx5_0 driver for the Mellanox ConnectX-5 InfiniBand device that only supports RoCE v1.

Note

Changes described here will remain effective until you reboot the host.

Prerequisites

  • The client uses an InfiniBand device with RoCE v2 protocol.
  • The server uses an InfiniBand device that only supports RoCE v1.

Procedure

  1. Create the /sys/kernel/config/rdma_cm/mlx5_0/ directory:

    # mkdir /sys/kernel/config/rdma_cm/mlx5_0/
  2. Display the default RoCE mode:

    # cat /sys/kernel/config/rdma_cm/mlx5_0/ports/1/default_roce_mode
    
    RoCE v2
  3. Change the default RoCE mode to version 1:

    # echo "IB/RoCE v1" > /sys/kernel/config/rdma_cm/mlx5_0/ports/1/default_roce_mode

2.3. Configuring Soft-RoCE

Soft-RoCE is a software implementation of remote direct memory access (RDMA) over Ethernet, which is also called RXE. Use Soft-RoCE on hosts without RoCE host channel adapters (HCA).

Important

The Soft-RoCE feature is provided as a Technology Preview only. Technology Preview features are not supported with Red Hat production Service Level Agreements (SLAs), might not be functionally complete, and Red Hat does not recommend using them for production. These previews provide early access to upcoming product features, enabling customers to test functionality and provide feedback during the development process.

See Technology Preview Features Support Scope on the Red Hat Customer Portal for information about the support scope for Technology Preview features.

Prerequisites

  • An Ethernet adapter is installed

Procedure

  1. Install the iproute, libibverbs, libibverbs-utils, and infiniband-diags packages:

    # yum install iproute libibverbs libibverbs-utils infiniband-diags
  2. Display the RDMA links:

    # rdma link show
  3. Load the rdma_rxe kernel module and add a new rxe device named rxe0 that uses the enp0s1 interface:

    # rdma link add rxe0 type rxe netdev enp1s0

Verification

  1. View the state of all RDMA links:

    # rdma link show
    
    link rxe0/1 state ACTIVE physical_state LINK_UP netdev enp1s0
  2. List the available RDMA devices:

    # ibv_devices
    
        device          	   node GUID
        ------          	----------------
        rxe0            	505400fffed5e0fb
  3. You can use the ibstat utility to display a detailed status:

    # ibstat rxe0
    
    CA 'rxe0'
    	CA type:
    	Number of ports: 1
    	Firmware version:
    	Hardware version:
    	Node GUID: 0x505400fffed5e0fb
    	System image GUID: 0x0000000000000000
    	Port 1:
    		State: Active
    		Physical state: LinkUp
    		Rate: 100
    		Base lid: 0
    		LMC: 0
    		SM lid: 0
    		Capability mask: 0x00890000
    		Port GUID: 0x505400fffed5e0fb
    		Link layer: Ethernet

Chapter 3. Configuring Soft-iWARP

Remote Direct Memory Access (RDMA) uses several libraries and protocols over an Ethernet such as iWARP, Soft-iWARP for performance improvement and aided programming interface.

3.1. Overview of iWARP and Soft-iWARP

Remote direct memory access (RDMA) uses the Internet Wide-area RDMA Protocol (iWARP) over Ethernet for converged and low latency data transmission over TCP. Using standard Ethernet switches and the TCP/IP stack, iWARP routes traffic across the IP subnets. This provides flexibility to efficiently use the existing infrastructure. In Red Hat Enterprise Linux, multiple providers implement iWARP in their hardware network interface cards. For example, cxgb4, irdma, qedr, and so on.

Soft-iWARP (siw) is a software-based iWARP kernel driver and user library for Linux. It is a software-based RDMA device that provides a programming interface to RDMA hardware when attached to network interface cards. It provides an easy way to test and validate the RDMA environment.

3.2. Configuring Soft-iWARP

Soft-iWARP (siw) implements the Internet Wide-area RDMA Protocol (iWARP) Remote direct memory access (RDMA) transport over the Linux TCP/IP network stack. It enables a system with a standard Ethernet adapter to interoperate with an iWARP adapter or with another system running the Soft-iWARP driver or a host with the hardware that supports iWARP.

Important

The Soft-iWARP feature is provided as a Technology Preview only. Technology Preview features are not supported with Red Hat production Service Level Agreements (SLAs), might not be functionally complete, and Red Hat does not recommend using them for production. These previews provide early access to upcoming product features, enabling customers to test functionality and provide feedback during the development process.

See Technology Preview Features Support Scope on the Red Hat Customer Portal for information about the support scope for Technology Preview features.

To configure Soft-iWARP, you can use this procedure in a script to run automatically when the system boots.

Prerequisites

  • An Ethernet adapter is installed

Procedure

  1. Install the iproute, libibverbs, libibverbs-utils, and infiniband-diags packages:

    # yum install iproute libibverbs libibverbs-utils infiniband-diags
  2. Display the RDMA links:

    # rdma link show
  3. Load the siw kernel module:

    # modprobe siw
  4. Add a new siw device named siw0 that uses the enp0s1 interface:

    # rdma link add siw0 type siw netdev enp0s1

Verification

  1. View the state of all RDMA links:

    # rdma link show
    
    link siw0/1 state ACTIVE physical_state LINK_UP netdev enp0s1
  2. List the available RDMA devices:

    # ibv_devices
    
     device                 node GUID
     ------              ----------------
     siw0                0250b6fffea19d61
  3. You can use the ibv_devinfo utility to display a detailed status:

    # ibv_devinfo siw0
    
        hca_id:               siw0
        transport:            iWARP (1)
        fw_ver:               0.0.0
        node_guid:            0250:b6ff:fea1:9d61
        sys_image_guid:       0250:b6ff:fea1:9d61
        vendor_id:            0x626d74
        vendor_part_id:       1
        hw_ver:               0x0
        phys_port_cnt:          1
            port:               1
                state:          PORT_ACTIVE (4)
                max_mtu:        1024 (3)
                active_mtu:     1024 (3)
                sm_lid:         0
                port_lid:       0
                port_lmc:       0x00
                link_layer:     Ethernet

Chapter 4. Configuring the core RDMA subsystem

The rdma service configuration manages the network protocols and communication standards such as InfiniBand, iWARP, and RoCE.

4.1. Renaming IPoIB devices

By default, the kernel names Internet Protocol over InfiniBand (IPoIB) devices, for example, ib0, ib1, and so on. To avoid conflicts, Red Hat recommends creating a rule in the udev device manager to create persistent and meaningful names such as mlx4_ib0.

Prerequisites

  • You have installed an InfiniBand device.

Procedure

  1. Display the hardware address of the device ib0:

    # ip link show ib0
    8: ib0: >BROADCAST,MULTICAST,UP,LOWER_UP< mtu 65520 qdisc pfifo_fast state UP mode DEFAULT qlen 256
        link/infiniband 80:00:02:00:fe:80:00:00:00:00:00:00:00:02:c9:03:00:31:78:f2 brd 00:ff:ff:ff:ff:12:40:1b:ff:ff:00:00:00:00:00:00:ff:ff:ff:ff

    The last eight bytes of the address are required to create a udev rule in the next step.

  2. To configure a rule that renames the device with the 00:02:c9:03:00:31:78:f2 hardware address to mlx4_ib0, edit the /etc/udev/rules.d/70-persistent-ipoib.rules file and add an ACTION rule:

    ACTION=="add", SUBSYSTEM=="net", DRIVERS=="?*", ATTR{type}=="32", ATTR{address}=="?*00:02:c9:03:00:31:78:f2", NAME="mlx4_ib0"
  3. Reboot the host:

    # reboot

Additional resources

4.2. Increasing the amount of memory that users are allowed to pin in the system

Remote direct memory access (RDMA) operations require the pinning of physical memory. As a consequence, the kernel is not allowed to write memory into the swap space. If a user pins too much memory, the system can run out of memory, and the kernel terminates processes to free up more memory. Therefore, memory pinning is a privileged operation.

If non-root users need to run large RDMA applications, it is necessary to increase the amount of memory to maintain pages in primary memory pinned all the time.

Procedure

  • As the root user, create the file /etc/security/limits.conf with the following contents:

    @rdma soft memlock unlimited
    @rdma hard memlock unlimited

Verification

  1. Log in as a member of the rdma group after editing the /etc/security/limits.conf file.

    Note that Red Hat Enterprise Linux applies updated ulimit settings when the user logs in.

  2. Use the ulimit -l command to display the limit:

    $ ulimit -l
    unlimited

    If the command returns unlimited, the user can pin an unlimited amount of memory.

Additional resources

  • limits.conf(5) man page

4.3. Configuring the rdma service

With the Remote Direct Memory Access (RDMA) protocol, you can transfer data between the RDMA enabled systems over the network by using the main memory. The RDMA protocol provides low latency and high throughput. To manage supported network protocols and communication standards, you need to configure the rdma service. This configuration includes high speed network protocols such as RoCE and iWARP, and communication standards such as Soft-RoCE and Soft-iWARP. When Red Hat Enterprise Linux detects InfiniBand, iWARP, or RoCE devices and their configuration files residing at the /etc/rdma/modules/* directory, the udev device manager instructs systemd to start the rdma service. Configuration of modules in the /etc/rdma/modules/rdma.conf file remains persistent after reboot. You need to restart the rdma-load-modules@rdma.service configuration service to apply changes.

Procedure

  1. Edit the /etc/rdma/modules/rdma.conf file and uncomment the modules that you want to enable:

    # These modules are loaded by the system if any RDMA devices is installed
    
    # iSCSI over RDMA client support
    ib_iser
    
    # iSCSI over RDMA target support
    ib_isert
    
    # SCSI RDMA Protocol target driver
    ib_srpt
    
    # User access to RDMA verbs (supports libibverbs)
    ib_uverbs
    
    # User access to RDMA connection management (supports librdmacm)
    rdma_ucm
    
    # RDS over RDMA support
    # rds_rdma
    
    # NFS over RDMA client support
    xprtrdma
    
    # NFS over RDMA server support
    svcrdma
  2. Restart the service to make the changes effective:

    # systemctl restart <rdma-load-modules@rdma.service>

Verification

  • After a reboot, check the service status:

    # systemctl status <rdma-load-modules@rdma.service>

4.4. Enabling NFS over RDMA on an NFS server

Remote Direct Memory Access (RDMA) is a protocol that enables a client system to directly transfer data from the memory of a storage server into its own memory. This enhances storage throughput, decreases latency in data transfer between the server and client, and reduces CPU load on both ends. If both the NFS server and clients are connected over RDMA, clients can use NFSoRDMA to mount an exported directory.

Prerequisites

  • The NFS service is running and configured
  • An InfiniBand or RDMA over Converged Ethernet (RoCE) device is installed on the server.
  • IP over InfiniBand (IPoIB) is configured on the server, and the InfiniBand device has an IP address assigned.

Procedure

  1. Install the rdma-core package:

    # dnf install rdma-core
  2. If the package was already installed, verify that the xprtrdma and svcrdma modules in the /etc/rdma/modules/rdma.conf file are uncommented:

    # NFS over RDMA client support
    xprtrdma
    # NFS over RDMA server support
    svcrdma
  3. Optional. By default, NFS over RDMA uses port 20049. If you want to use a different port, set the rdma-port setting in the [nfsd] section of the /etc/nfs.conf file:

    rdma-port=<port>
  4. Open the NFSoRDMA port in firewalld:

    # firewall-cmd --permanent --add-port={20049/tcp,20049/udp}
    # firewall-cmd --reload

    Adjust the port numbers if you set a different port than 20049.

  5. Restart the nfs-server service:

    # systemctl restart nfs-server

Verification

  1. On a client with InfiniBand hardware, perform the following steps:

    1. Install the following packages:

      # dnf install nfs-utils rdma-core
    2. Mount an exported NFS share over RDMA:

      # mount -o rdma server.example.com:/nfs/projects/ /mnt/

      If you set a port number other than the default (20049), pass port=<port_number> to the command:

      # mount -o rdma,port=<port_number> server.example.com:/nfs/projects/ /mnt/
    3. Verify that the share was mounted with the rdma option:

      # mount | grep "/mnt"
      server.example.com:/nfs/projects/ on /mnt type nfs (...,proto=rdma,...)

Chapter 5. Configuring an InfiniBand subnet manager

All InfiniBand networks must have a subnet manager running for the network to function. This is true even if two machines are connected directly with no switch involved.

It is possible to have more than one subnet manager. In that case, one acts as a master and another subnet manager acts as a slave that will take over in case the master subnet manager fails.

Most InfiniBand switches contain an embedded subnet manager. However, if you need a more up-to-date subnet manager or if you require more control, use the OpenSM subnet manager provided by Red Hat Enterprise Linux.

5.1. Installing the OpenSM subnet manager

OpenSM is a subnet manager and administrator that follows the InfiniBand specifications to initialize InfiniBand hardware where at least one instance of OpenSM service always runs.

Procedure

  1. Install the opensm package:

    # yum install opensm
  2. Configure OpenSM in case the default installation does not match your environment.

    With only one InfiniBand port, the host acts as the master subnet manager that does not require any custom changes. The default configuration works without any modification.

  3. Enable and start the opensm service:

    # systemctl enable --now opensm

Additional resources

  • opensm(8) man page

5.2. Configuring OpenSM using the simple method

OpenSM is an InfiniBand specification-based subnet manager and administrator that configures the InfiniBand fabric, a network topology to interconnect the InfiniBand nodes.

Prerequisites

  • One or more InfiniBand ports are installed on the server

Procedure

  1. Obtain the GUIDs for the ports using the ibstat utility:

    # ibstat -d mlx4_0
    
    CA 'mlx4_0'
       CA type: MT4099
       Number of ports: 2
       Firmware version: 2.42.5000
       Hardware version: 1
       Node GUID: 0xf4521403007be130
       System image GUID: 0xf4521403007be133
       Port 1:
          State: Active
          Physical state: LinkUp
          Rate: 56
          Base lid: 3
          LMC: 0
          SM lid: 1
          Capability mask: 0x02594868
          Port GUID: 0xf4521403007be131
          Link layer: InfiniBand
       Port 2:
          State: Down
          Physical state: Disabled
          Rate: 10
          Base lid: 0
          LMC: 0
          SM lid: 0
          Capability mask: 0x04010000
          Port GUID: 0xf65214fffe7be132
          Link layer: Ethernet
    Note

    Some InfiniBand adapters use the same GUID for the node, system, and port.

  2. Edit the /etc/sysconfig/opensm file and set the GUIDs in the GUIDS parameter:

    GUIDS="GUID_1 GUID_2"
  3. You can set the PRIORITY parameter if multiple subnet managers are available in your subnet. For example:

    PRIORITY=15

Additional resources

  • /etc/sysconfig/opensm

5.3. Configuring OpenSM by editing the opensm.conf file

OpenSM performance depends on the number of available InfiniBand ports on the device. You can customize it by editing /etc/rdma/opensm.conf file.

Prerequisites

  • Only one InfiniBand port is installed on the server.

Procedure

  1. Edit the /etc/rdma/opensm.conf file and customize the settings to match your environment.

    After updating an opensm package, the yum utility overrides the /etc/rdma/opensm.conf and creates a copy which is the new OpenSM configuration file /etc/rdma/opensm.conf.rpmnew. So, you can compare the previous and new files to identify changes and incorporate them manually in file opensm.conf.

  2. Restart the opensm service:

    # systemctl restart opensm

5.4. Configuring multiple OpenSM instances

OpenSM is an InfiniBand constrained subnet manager and administrator. To provide high performance, OpenSM uses switched fabric network topology that interconnects with InfiniBand network nodes.

Prerequisites

  • One or more InfiniBand ports are installed on the server.

Procedure

  1. Copy the /etc/rdma/opensm.conf file to /etc/rdma/opensm.conf.orig file:

    # cp /etc/rdma/opensm.conf /etc/rdma/opensm.conf.orig

    When you install an updated opensm package, the yum utility overrides the /etc/rdma/opensm.conf. With the copy created in this step, compare the previous and new files to identify changes and incorporate them manually in the instance-specific opensm.conf files.

  2. Create a copy of the /etc/rdma/opensm.conf file:

    # cp /etc/rdma/opensm.conf /etc/rdma/opensm.conf.1

    For each instance, create and append a unique and continuous number to a copy of the configuration file.

    After updating the opensm package, the yum utility stores the new OpenSM configuration file as /etc/rdma/opensm.conf.rpmnew. Compare this file with your customized /etc/rdma/opensm.conf.\* files, and manually incorporate the changes.

  3. Edit the copy you created in the previous step, and customize the settings for the instance to match your environment. For example, set the guid, subnet_prefix, and logdir parameters.
  4. Optionally, create a partitions.conf file with a unique name specifically for this subnet and reference that file in the partition_config_file parameter in the corresponding copy of the opensm.conf file.
  5. Repeat the previous steps for each instance you want to create.
  6. Start the opensm service:

    # systemctl start opensm

    The opensm service automatically starts a unique instance for each opensm.conf.* file in the /etc/rdma/ directory. If multiple opensm.conf.* files exist, the service ignores settings in the /etc/sysconfig/opensm file as well as in the base /etc/rdma/opensm.conf file.

5.5. Creating a partition configuration

Partitions enable administrators to create subnets on InfiniBand similar to Ethernet VLANs.

Important

If you define a partition with a specific speed such as 40 Gbps, all hosts within this partition must support this speed minimum. If a host does not meet the speed requirements, it cannot join the partition. Therefore, set the speed of a partition to the lowest speed supported by any host with permission to join the partition.

Prerequisites

  • One or more InfiniBand ports are installed on the server

Procedure

  1. Edit the /etc/rdma/partitions.conf file to configure the partitions as follows:

    Note

    All fabrics must contain the 0x7fff partition, and all switches and all hosts must belong to that fabric.

    Add the following content to the file to create the 0x7fff default partition at a reduced speed of 10 Gbps, and a partition 0x0002 with a speed of 40 Gbps:

    # For reference:
    # IPv4 IANA reserved multicast addresses:
    #   http://www.iana.org/assignments/multicast-addresses/multicast-addresses.txt
    # IPv6 IANA reserved multicast addresses:
    #   http://www.iana.org/assignments/ipv6-multicast-addresses/ipv6-multicast-addresses.xml
    #
    # mtu =
    #   1 = 256
    #   2 = 512
    #   3 = 1024
    #   4 = 2048
    #   5 = 4096
    #
    # rate =
    #   2  = 2.5 GBit/s
    #   3  = 10   GBit/s
    #   4  = 30   GBit/s
    #   5  = 5   GBit/s
    #   6  = 20   GBit/s
    #   7  = 40   GBit/s
    #   8  = 60   GBit/s
    #   9  = 80   GBit/s
    #   10 = 120   GBit/s
    
    Default=0x7fff, rate=3, mtu=4, scope=2, defmember=full:
        ALL, ALL_SWITCHES=full;
    Default=0x7fff, ipoib, rate=3, mtu=4, scope=2:
        mgid=ff12:401b::ffff:ffff   # IPv4 Broadcast address
        mgid=ff12:401b::1           # IPv4 All Hosts group
        mgid=ff12:401b::2           # IPv4 All Routers group
        mgid=ff12:401b::16          # IPv4 IGMP group
        mgid=ff12:401b::fb          # IPv4 mDNS group
        mgid=ff12:401b::fc          # IPv4 Multicast Link Local Name Resolution group
        mgid=ff12:401b::101         # IPv4 NTP group
        mgid=ff12:401b::202         # IPv4 Sun RPC
        mgid=ff12:601b::1           # IPv6 All Hosts group
        mgid=ff12:601b::2           # IPv6 All Routers group
        mgid=ff12:601b::16          # IPv6 MLDv2-capable Routers group
        mgid=ff12:601b::fb          # IPv6 mDNS group
        mgid=ff12:601b::101         # IPv6 NTP group
        mgid=ff12:601b::202         # IPv6 Sun RPC group
        mgid=ff12:601b::1:3         # IPv6 Multicast Link Local Name Resolution group
        ALL=full, ALL_SWITCHES=full;
    
    ib0_2=0x0002, rate=7, mtu=4, scope=2, defmember=full:
            ALL, ALL_SWITCHES=full;
    ib0_2=0x0002, ipoib, rate=7, mtu=4, scope=2:
        mgid=ff12:401b::ffff:ffff   # IPv4 Broadcast address
        mgid=ff12:401b::1           # IPv4 All Hosts group
        mgid=ff12:401b::2           # IPv4 All Routers group
        mgid=ff12:401b::16          # IPv4 IGMP group
        mgid=ff12:401b::fb          # IPv4 mDNS group
        mgid=ff12:401b::fc          # IPv4 Multicast Link Local Name Resolution group
        mgid=ff12:401b::101         # IPv4 NTP group
        mgid=ff12:401b::202         # IPv4 Sun RPC
        mgid=ff12:601b::1           # IPv6 All Hosts group
        mgid=ff12:601b::2           # IPv6 All Routers group
        mgid=ff12:601b::16          # IPv6 MLDv2-capable Routers group
        mgid=ff12:601b::fb          # IPv6 mDNS group
        mgid=ff12:601b::101         # IPv6 NTP group
        mgid=ff12:601b::202         # IPv6 Sun RPC group
        mgid=ff12:601b::1:3         # IPv6 Multicast Link Local Name Resolution group
        ALL=full, ALL_SWITCHES=full;

Chapter 6. Configuring IPoIB

By default, InfiniBand does not use the internet protocol (IP) for communication. However, IP over InfiniBand (IPoIB) provides an IP network emulation layer on top of InfiniBand remote direct memory access (RDMA) networks. This allows existing unmodified applications to transmit data over InfiniBand networks, but the performance is lower than if the application would use RDMA natively.

Note

The Mellanox devices, starting from ConnectX-4 and above, on RHEL 8 and later use Enhanced IPoIB mode by default (datagram only). Connected mode is not supported on these devices.

6.1. The IPoIB communication modes

An IPoIB device is configurable in either Datagram or Connected mode. The difference is the type of queue pair the IPoIB layer attempts to open with the machine at the other end of the communication:

  • In the Datagram mode, the system opens an unreliable, disconnected queue pair.

    This mode does not support packages larger than Maximum Transmission Unit (MTU) of the InfiniBand link layer. During transmission of data, the IPoIB layer adds a 4-byte IPoIB header on top of the IP packet. As a result, the IPoIB MTU is 4 bytes less than the InfiniBand link-layer MTU. As 2048 is a common InfiniBand link-layer MTU, the common IPoIB device MTU in Datagram mode is 2044.

  • In the Connected mode, the system opens a reliable, connected queue pair.

    This mode allows messages larger than the InfiniBand link-layer MTU. The host adapter handles packet segmentation and reassembly. As a result, in the Connected mode, the messages sent from Infiniband adapters have no size limits. However, there are limited IP packets due to the data field and TCP/IP header field. For this reason, the IPoIB MTU in the Connected mode is 65520 bytes.

    The Connected mode has a higher performance but consumes more kernel memory.

Though a system is configured to use the Connected mode, a system still sends multicast traffic using the Datagram mode because InfiniBand switches and fabric cannot pass multicast traffic in the Connected mode. Also, when the host is not configured to use the Connected mode, the system falls back to the Datagram mode.

While running an application that sends multicast data up to MTU on the interface, configures the interface in Datagram mode or configure the application to cap the send size of a packet that will fit in datagram-sized packets.

6.2. Understanding IPoIB hardware addresses

IPoIB devices have a 20 byte hardware address that consists of the following parts:

  • The first 4 bytes are flags and queue pair numbers
  • The next 8 bytes are the subnet prefix

    The default subnet prefix is 0xfe:80:00:00:00:00:00:00. After the device connects to the subnet manager, the device changes this prefix to match with the configured subnet manager.

  • The last 8 bytes are the Globally Unique Identifier (GUID) of the InfiniBand port that attaches to the IPoIB device
Note

As the first 12 bytes can change, do not use them in the udev device manager rules.

6.3. Configuring an IPoIB connection using nmcli commands

The nmcli command-line utility controls the NetworkManager and reports network status using CLI.

Prerequisites

  • An InfiniBand device is installed on the server
  • The corresponding kernel module is loaded

Procedure

  1. Create the InfiniBand connection to use the mlx4_ib0 interface in the Connected transport mode and the maximum MTU of 65520 bytes:

    # nmcli connection add type infiniband con-name mlx4_ib0 ifname mlx4_ib0 transport-mode Connected mtu 65520
  2. You can also set 0x8002 as a P_Key interface of the mlx4_ib0 connection:

    # nmcli connection modify mlx4_ib0 infiniband.p-key 0x8002
  3. To configure the IPv4 settings set a static IPv4 address, network mask, default gateway, and DNS server of the mlx4_ib0 connection:

    # nmcli connection modify mlx4_ib0 ipv4.addresses 192.0.2.1/24
    # nmcli connection modify mlx4_ib0 ipv4.gateway 192.0.2.254
    # nmcli connection modify mlx4_ib0 ipv4.dns 192.0.2.253
    # nmcli connection modify mlx4_ib0 ipv4.method manual
  4. To configure the IPv6 settings set a static IPv6 address, network mask, default gateway, and DNS server of the mlx4_ib0 connection:

    # nmcli connection modify mlx4_ib0 ipv6.addresses 2001:db8:1::1/32
    # nmcli connection modify mlx4_ib0 ipv6.gateway 2001:db8:1::fffe
    # nmcli connection modify mlx4_ib0 ipv6.dns 2001:db8:1::fffd
    # nmcli connection modify mlx4_ib0 ipv6.method manual
  5. To activate the mlx4_ib0 connection:

    # nmcli connection up mlx4_ib0

6.4. Configuring an IPoIB connection by using the network RHEL system role

You can use the network RHEL system role to remotely create NetworkManager connection profiles for IP over InfiniBand (IPoIB) devices. For example, remotely add an InfiniBand connection for the mlx4_ib0 interface with the following settings by running an Ansible playbook:

  • An IPoIB device - mlx4_ib0.8002
  • A partition key p_key - 0x8002
  • A static IPv4 address - 192.0.2.1 with a /24 subnet mask
  • A static IPv6 address - 2001:db8:1::1 with a /64 subnet mask

Perform this procedure on the Ansible control node.

Prerequisites

  • You have prepared the control node and the managed nodes
  • You are logged in to the control node as a user who can run playbooks on the managed nodes.
  • The account you use to connect to the managed nodes has sudo permissions on them.
  • An InfiniBand device named mlx4_ib0 is installed in the managed nodes.
  • The managed nodes use NetworkManager to configure the network.

Procedure

  1. Create a playbook file, for example ~/playbook.yml, with the following content:

    ---
    - name: Configure the network
      hosts: managed-node-01.example.com
      tasks:
        - name: Configure IPoIB
          ansible.builtin.include_role:
            name: rhel-system-roles.network
          vars:
            network_connections:
              # InfiniBand connection mlx4_ib0
              - name: mlx4_ib0
                interface_name: mlx4_ib0
                type: infiniband
    
              # IPoIB device mlx4_ib0.8002 on top of mlx4_ib0
              - name: mlx4_ib0.8002
                type: infiniband
                autoconnect: yes
                infiniband:
                  p_key: 0x8002
                  transport_mode: datagram
                parent: mlx4_ib0
                ip:
                  address:
                    - 192.0.2.1/24
                    - 2001:db8:1::1/64
                state: up

    If you set a p_key parameter as in this example, do not set an interface_name parameter on the IPoIB device.

  2. Validate the playbook syntax:

    $ ansible-playbook --syntax-check ~/playbook.yml

    Note that this command only validates the syntax and does not protect against a wrong but valid configuration.

  3. Run the playbook:

    $ ansible-playbook ~/playbook.yml

Verification

  1. On the managed-node-01.example.com host, display the IP settings of the mlx4_ib0.8002 device:

    # ip address show mlx4_ib0.8002
    ...
    inet 192.0.2.1/24 brd 192.0.2.255 scope global noprefixroute ib0.8002
       valid_lft forever preferred_lft forever
    inet6 2001:db8:1::1/64 scope link tentative noprefixroute
       valid_lft forever preferred_lft forever
  2. Display the partition key (P_Key) of the mlx4_ib0.8002 device:

    # cat /sys/class/net/mlx4_ib0.8002/pkey
    0x8002
  3. Display the mode of the mlx4_ib0.8002 device:

    # cat /sys/class/net/mlx4_ib0.8002/mode
    datagram

Additional resources

  • /usr/share/ansible/roles/rhel-system-roles.network/README.md file
  • /usr/share/doc/rhel-system-roles/network/ directory

6.5. Configuring an IPoIB connection using nm-connection-editor

The nmcli-connection-editor application configures and manages network connections stored by NetworkManager using the management console.

Prerequisites

  • An InfiniBand device is installed on the server.
  • Corresponding kernel module is loaded
  • The nm-connection-editor package is installed.

Procedure

  1. Enter the command:

    $ nm-connection-editor
  2. Click the + button to add a new connection.
  3. Select the InfiniBand connection type and click Create.
  4. On the InfiniBand tab:

    1. Change the connection name if you want to.
    2. Select the transport mode.
    3. Select the device.
    4. Set an MTU if needed.
  5. On the IPv4 Settings tab, configure the IPv4 settings. For example, set a static IPv4 address, network mask, default gateway, and DNS server: infiniband IPv4 settings nm connection editor
  6. On the IPv6 Settings tab, configure the IPv6 settings. For example, set a static IPv6 address, network mask, default gateway, and DNS server: infiniband IPv6 settings nm connection editor
  7. Click Save to save the team connection.
  8. Close nm-connection-editor.
  9. You can set a P_Key interface. As this setting is not available in nm-connection-editor, you must set this parameter on the command line.

    For example, to set 0x8002 as P_Key interface of the mlx4_ib0 connection:

    # nmcli connection modify mlx4_ib0 infiniband.p-key 0x8002

Chapter 7. Testing InfiniBand networks

7.1. Testing early InfiniBand RDMA operations

InfiniBand provides low latency and high performance for Remote Direct Memory Access (RDMA).

Note

Apart from InfiniBand, if you use IP-based devices such as Internet Wide-area Remote Protocol(iWARP) or RDMA over Converged Ethernet (RoCE) or InfiniBand over Ethernet (IBoE) devices, see:

Prerequisites

  • You have configured the rdma service.
  • You have installed the libibverbs-utils and infiniband-diags packages.

Procedure

  1. List the available InfiniBand devices:

    # ibv_devices
    
        device                 node GUID
        ------              ----------------
        mlx4_0              0002c903003178f0
        mlx4_1              f4521403007bcba0
  2. Display the information of the mlx4_1 device:

    # ibv_devinfo -d mlx4_1
    
    hca_id: mlx4_1
         transport:                  InfiniBand (0)
         fw_ver:                     2.30.8000
         node_guid:                  f452:1403:007b:cba0
         sys_image_guid:             f452:1403:007b:cba3
         vendor_id:                  0x02c9
         vendor_part_id:             4099
         hw_ver:                     0x0
         board_id:                   MT_1090120019
         phys_port_cnt:              2
              port:   1
                    state:              PORT_ACTIVE (4)
                    max_mtu:            4096 (5)
                    active_mtu:         2048 (4)
                    sm_lid:             2
                    port_lid:           2
                    port_lmc:           0x01
                    link_layer:         InfiniBand
    
              port:   2
                    state:              PORT_ACTIVE (4)
                    max_mtu:            4096 (5)
                    active_mtu:         4096 (5)
                    sm_lid:             0
                    port_lid:           0
                    port_lmc:           0x00
                    link_layer:         Ethernet
  3. Display the status of the mlx4_1 device:

    # ibstat mlx4_1
    
    CA 'mlx4_1'
         CA type: MT4099
         Number of ports: 2
         Firmware version: 2.30.8000
         Hardware version: 0
         Node GUID: 0xf4521403007bcba0
         System image GUID: 0xf4521403007bcba3
         Port 1:
               State: Active
               Physical state: LinkUp
               Rate: 56
               Base lid: 2
               LMC: 1
               SM lid: 2
               Capability mask: 0x0251486a
               Port GUID: 0xf4521403007bcba1
               Link layer: InfiniBand
         Port 2:
               State: Active
               Physical state: LinkUp
               Rate: 40
               Base lid: 0
               LMC: 0
               SM lid: 0
               Capability mask: 0x04010000
               Port GUID: 0xf65214fffe7bcba2
               Link layer: Ethernet
  4. The ibping utility pings an InfiniBand address and runs as a client/server.

    1. To start server mode on a host, use the -S parameter on port number -P with -C InfiniBand certificate authority (CA) name:

      # ibping -S -C mlx4_1 -P 1
    2. To start client mode on another host, send some packets -c on port number -P using -C InfiniBand certificate authority (CA) name with -L Local Identifier (LID):

      # ibping -c 50 -C mlx4_0 -P 1 -L 2

Additional resources

  • ibping(8) man page

7.2. Testing an IPoIB using the ping utility

After you configured IP over InfiniBand (IPoIB), use the ping utility to send ICMP packets to test the IPoIB connection.

Prerequisites

  • The two RDMA hosts are connected in the same InfiniBand fabric with RDMA ports
  • The IPoIB interfaces in both hosts are configured with IP addresses within the same subnet

Procedure

  • Use the ping utility to send five ICMP packets to the remote host’s InfiniBand adapter:

    # ping -c5 192.0.2.1

7.3. Testing an RDMA network using qperf after IPoIB is configured

The qperf utility measures RDMA and IP performance between two nodes in terms of bandwidth, latency, and CPU utilization.

Prerequisites

  • You have installed the qperf package on both hosts.
  • IPoIB is configured on both hosts.

Procedure

  1. Start qperf on one of the hosts without any options to act as a server:

    # qperf
  2. Use the following commands on the client. The commands use port 1 of the mlx4_0 host channel adapter in the client to connect to IP address 192.0.2.1 assigned to the InfiniBand adapter in the server.

    1. Display the configuration of the host channel adapter:

      # qperf -v -i mlx4_0:1 192.0.2.1 conf
      
      conf:
          loc_node   =  rdma-dev-01.lab.bos.redhat.com
          loc_cpu    =  12 Cores: Mixed CPUs
          loc_os     =  Linux 4.18.0-187.el8.x86_64
          loc_qperf  =  0.4.11
          rem_node   =  rdma-dev-00.lab.bos.redhat.com
          rem_cpu    =  12 Cores: Mixed CPUs
          rem_os     =  Linux 4.18.0-187.el8.x86_64
          rem_qperf  =  0.4.11
    2. Display the Reliable Connection (RC) streaming two-way bandwidth:

      # qperf -v -i mlx4_0:1 192.0.2.1 rc_bi_bw
      
      rc_bi_bw:
          bw             =  10.7 GB/sec
          msg_rate       =   163 K/sec
          loc_id         =  mlx4_0
          rem_id         =  mlx4_0:1
          loc_cpus_used  =    65 % cpus
          rem_cpus_used  =    62 % cpus
    3. Display the RC streaming one-way bandwidth:

      # qperf -v -i mlx4_0:1 192.0.2.1 rc_bw
      
      rc_bw:
          bw              =  6.19 GB/sec
          msg_rate        =  94.4 K/sec
          loc_id          =  mlx4_0
          rem_id          =  mlx4_0:1
          send_cost       =  63.5 ms/GB
          recv_cost       =    63 ms/GB
          send_cpus_used  =  39.5 % cpus
          recv_cpus_used  =    39 % cpus

Additional resources

  • qperf(1) man page

Legal Notice

Copyright © 2024 Red Hat, Inc.
The text of and illustrations in this document are licensed by Red Hat under a Creative Commons Attribution–Share Alike 3.0 Unported license ("CC-BY-SA"). An explanation of CC-BY-SA is available at http://creativecommons.org/licenses/by-sa/3.0/. In accordance with CC-BY-SA, if you distribute this document or an adaptation of it, you must provide the URL for the original version.
Red Hat, as the licensor of this document, waives the right to enforce, and agrees not to assert, Section 4d of CC-BY-SA to the fullest extent permitted by applicable law.
Red Hat, Red Hat Enterprise Linux, the Shadowman logo, the Red Hat logo, JBoss, OpenShift, Fedora, the Infinity logo, and RHCE are trademarks of Red Hat, Inc., registered in the United States and other countries.
Linux® is the registered trademark of Linus Torvalds in the United States and other countries.
Java® is a registered trademark of Oracle and/or its affiliates.
XFS® is a trademark of Silicon Graphics International Corp. or its subsidiaries in the United States and/or other countries.
MySQL® is a registered trademark of MySQL AB in the United States, the European Union and other countries.
Node.js® is an official trademark of Joyent. Red Hat is not formally related to or endorsed by the official Joyent Node.js open source or commercial project.
The OpenStack® Word Mark and OpenStack logo are either registered trademarks/service marks or trademarks/service marks of the OpenStack Foundation, in the United States and other countries and are used with the OpenStack Foundation's permission. We are not affiliated with, endorsed or sponsored by the OpenStack Foundation, or the OpenStack community.
All other trademarks are the property of their respective owners.
Red Hat logoGithubRedditYoutubeTwitter

Learn

Try, buy, & sell

Communities

About Red Hat Documentation

We help Red Hat users innovate and achieve their goals with our products and services with content they can trust.

Making open source more inclusive

Red Hat is committed to replacing problematic language in our code, documentation, and web properties. For more details, see the Red Hat Blog.

About Red Hat

We deliver hardened solutions that make it easier for enterprises to work across platforms and environments, from the core datacenter to the network edge.

© 2024 Red Hat, Inc.