使用 RHEL 8.8 时，可以为使用 Mellanox InfiniBand 硬件的用户做出选择。他们可以利用 Red Hat 附带的InfiniBand Support组安装（我选择在安装过程中执行）来启用 InfiniBand 网络，或者他们可以选择MLNX_OFED_LINUX-23.04-1.1.3.0-rhel8.8-x86_64.iso并执行他们的mlnxofedinstall脚本，这会用他们自己的软件包替换 Red Hat 软件包。
除了固件更新之外，有人知道使用 MLNX 路径相对于 Red Hat 路径来设置 InfiniBand 有什么优势吗？
MLNX 方法通常涉及更多的设置和配置，因此我对使用 MLNX 方法（考虑到我正在使用他们的硬件）而不是 Red Hat 提供的性能理由感兴趣。
以供参考：
Usage: ./mlnxofedinstall [OPTIONS]
Installation control:
--force Force installation
--tmpdir Change tmp directory. Default: /tmp
-k|--kernel <version>
Default on this system: 3.10.0-1160.92.1.el7.x86_64
-s|--kernel-sources <path>
Default on this system: /lib/modules/3.10.0-1160.92.1.el7.x86_64/build
-U|--update Update installed version
--with-32bit Install 32-bit libraries. Relevant for x86_64 platforms
--without-32bit Skip 32-bit libraries installation (Default). Relevant for x86_64 platforms
--distro Set Distro name for the running OS (e.g: rhel6.5, sles12sp3). Default: Use auto-detection
--skip-distro-check Do not check MLNX_OFED_LINUX vs Distro matching
--without-depcheck Run the installation without verifying that all required Distro's packages are installed
--check-deps-only Check for missing required Distro's packages and exit
--print-distro Print distribution name and exit
--add-kernel-support
Add kernel support (Run mlnx_add_kernel_support.sh) and install it
--add-kernel-support-build-only
Add kernel support (Run mlnx_add_kernel_support.sh) and exit
--kernel-extra-args '<args>'
pass <args> to kernel configure script (single paraeter, space separated)
- Pass it along with --add-kernel-support
--kmp Install kernel RPMs with KMP support if available
- Enabled by default with official MLNX_OFED_LINUX releases
- Disabled by default if new MLNX_OFED_LINUX was created with mlnx_add_kernel_support.sh
- Passing it along with --add-kernel-support will pass --kmp to mlnx_add_kernel_support.sh
--disable-kmp Install kernel RPMs without KMP support if available
- Enabled by default if new MLNX_OFED_LINUX was created with mlnx_add_kernel_support.sh
--skip-kmp-verify Do not check if the KMP rpms are compatiable with the target kernel
--speed-up-kmp Speed up KMP RPMs installation by skipping %post section and running the required operations once for all KMPs
--umad-dev-rw Grant non root users read/write permission for umad devices instead of default
--umad-dev-na Prevent from non root users read/write access for umad devices. Overrides '--umad-dev-rw'
--enable-mlnx_tune Enable Running the mlnx_tune utility
--enable-affinity Run mlnx_affinity script upon boot
--disable-affinity Disable mlnx_affinity script (Default)
--skip-unsupported-devices-check
Don't abort if system has an older, unsupported card
--enable-opensm Run opensm upon boot
--package-install-options
RPM install options to use when installing RPM packages (comma separated list)
--pre-install-<package> <path to script>
Run given script before given package's install
--post-install-<package> <path to script>
Run given script after given package's install
Firmware update control:
--without-fw-update Skip firmware update
--fw-update-only Update firmware. Skip driver installation
--force-fw-update Force firmware update
--fw-image-dir Firmware images directory to use instead of default package content
Package selection:
-c|--config <packages config_file>
Example of the config file can be found under docs
--all Install all available packages
--bluefield Install BlueField packages
--hpc Install minimum packages required for HPC
--basic Install minimum packages for basic functionality
--dpdk Install minimum packages required for DPDK
--ovs-dpdk Install DPDK and OVS packages
--vma Install minimum packages required for VMA
--xlio Install minimum packages required for XLIO
--guest Install minimum packages required by guest OS
--hypervisor Install minimum packages required by hypervisor OS
--with-fabric-collector
Enable installing fabric-collector package
--with-nvmf Enable NVMEoF support
--with-nfsrdma Enable NFSoRDMA support
--without-nfsrdma Disable NFSoRDMA support (Default)
Extra package filtering:
--kernel-only Install kernel space packages only
--user-space-only Filter selected packages and install only User Space packages
--without-<package> Do not install package
--with-<package> Force installing package
Miscellaneous:
-h|--help Display this help message and exit
--post-start-delay <sec>
Set openibd POST_START_DELAY parameter in seconds. (Default 0)
-p|--print-available Print available packages for current platform
And create corresponding ofed.conf file
--copy-ifnames-udev Copy compatibility udev rules for interface names
Output control:
-v|-vv|-vvv Set verbosity level
-q Set quiet - no messages will be printed
答案1
Red Hat 分布式版本通常落后于 Mellanox/NVIDIA 版本，并且不包含 Mellanox/NVIDIA 版本的所有实用程序。如果安装了 Red Hat 版本，我将其删除并安装 Mellanox/NVIDIA 版本。
当 Red Hat 发布的版本存在错误时，我开始在 Red Hat 7 上执行此操作。此错误导致它只能与可靠连接 (RC) 一起使用。
尝试使用不可靠连接 (UC) 或不可靠数据报 (UD) 运行代码时会遇到错误。
另一方面是固件和代码版本具有相互依赖性。您需要确保您拥有的固件包含在 OFED 版本的支持列表中。