Tag: GenAI

As mentioned in the first article in this series, the PowerScale multipath client driver is a kernel module, and that means it needs to be installed alongside a corresponding kernel version to that which it was built on. Version matching is strict, right down to the minor build version.

There are two installation options provided for the PowerScale multipath client driver:

• As a pre-built binary installation package for each of the supported Linux distribution versions listed below.
• Or via source code under the GPL 2 open source license, which can be compiled at a customer site.

This article covers the first option, and outlines the steps involved with the installation of the pre-built binary driver package for the following, currently supported Linux versions:

Linux distribution	Kernel version	Upstream driver version (minimum)	Multipath driver version	Package available
OpenSUSE 15.4	5.14.x	4.x	1.x	ü
Ubuntu 20.04	5.4.x	4.x	1.x	ü
Ubuntu 22.04	5.15.x	4.x	1.x	ü

Package installation is typically handled best by the client Linux distro’s native package manager. Since it’s a kernel module, installing and updating the driver typically requires a reboot. PowerScale engineering anticipate periodically releasing updated driver packages to keep up with Linux on the supported platforms list, as well as fix bugs and add additional functionality.

This multipath client driver runs on both physical and virtual machines, and both x86 CPU architectures and GPU-based platforms, such as the NVIDIA DGX range, are supported.

While there is no specific NFS or OneFS core configuration required on the PowerScale cluster side when using Linux clients with the Dell multipath driver, there are a couple of basic prerequisites. The following OneFS support matrix on the top right of this slide lays out which driver functionality is available in what release, from OneFS 9.5 to current.

Version	NFSv3, NFSv4.1 TCP	NFSv3 RDMA	NFSv4.1 RDMA	NVIDIA SuperPOD
OneFS 9.5	Yes	Yes	No	No
OneFS 9.7	Yes	Yes	Yes	No
OneFS 9.9	Yes	Yes	Yes	Yes

Also note that OneFS 9.9 is required for any NVIDIA SuperPOD deployments, because there are some performance optimizations in 9.9 specifically for that platform.

The following CLI commands can be run on the PowerScale cluster to verify its compatibility. The cluster’s current OneFS version can be easily determined using the following CLI command:

# uname -or

Isilon OneFS 9.9.0.0

Also, to confirm RDMA is supported and enabled:

# isi nfs settings global view | grep -i RDMA

   NFS RDMA Enabled: Yes

Additionally, both the dynamic and static network pools can be configured on the cluster for use with the multipath driver. If F710 nodes are being deployed in the cluster, OneFS 9.7 or later is required.

Note that when deploying an NVIDIA SuperPOD or BasePOD solution, the reference architecture mandates a PowerScale cluster composed of F710 all-flash nodes running OneFS 9.9 or later.

For a Linux client to successfully connect to a PowerScale cluster using the multipath driver, there are a few prerequisites that must be met, in addition to running one of the Linux versions listed above. These include:

• If RDMA is being configured, the client must contain an RDMA-capable Ethernet NIC, such as the Mellanox CX series.
• The Linux client should have the ‘trace-cmd’ package installed, along with NFS client related packages.

For example, on an Ubuntu system:

# sudo apt install trace-cmd nfs-common

The following CLI commands can be used to verify the kernel version, and other pertinent details, of a Linux client:

# uname -a

Similarly, depending on the flavor of Linux, the following commands will show the details of the particular distribution:

# lsb_release -a

Or:

# cat /etc/os-release

The multipath client driver is available for download on the Dell Support Site to any customer that has OneFS entitlement: For security purposes, the download files are signed with SHA256. Using Debian linux as an example, the driver components can be accessed as follows:

1. Verify the SHA256 Checksum Manually

The downloaded driver package’s authenticity can be manually verified via its SHA256 checksum as follows:

First, calculate the checksum on the signed driver package:

# sha256sum dellnfs-modules_4.0.24-Dell-Technologies.kver.5.4.0-190-generic_amd64.deb.signed

Then compare the output with the value in the accompanying checksum file.

2. Extract the Signed Package

First, if not already available, install the necessary extraction tools:

# sudo apt-get install ar

Next, extract the signed driver file:

# ar x dellnfs-modules_4.0.24-Dell-Technologies.kver.5.4.0-190-generic_amd64.deb.signed

This should yield multiple extracted files, including the following:

 control.tar.xz, data.tar.xz, and debian-binary.

3. Repackage the File

If needed, the extracted contents can be repackaged into a standard ‘.deb’ package file:

# ar rcs dellnfs-modules_4.0.24-Dell-Technologies.kver.5.4.0-190-generic_amd64.deb debian-binary control.tar.xz data.tar.xz

4. Install the Repackaged Debian File

Once repackaged, you can install the Debian package using the ‘dpkg’ utility. For example:

# sudo dpkg -i dellnfs-modules_4.0.24-Dell-Technologies.kver.5.4.0-190-generic_amd64.deb

Package installation is handled using the native package manager, and each of the supported Linux distributions uses the following format and package installation utility:

Linux Distribution	Package Manager	Package Utility
OpenSUSE	RPM	Zipper
Ubuntu	Deb	apt-get / dpkg

The RPM and DEB packages can either be obtained from the Dell download site or built manually at the customer site.

The multipath client driver is provided as a pre-built binary installation package for each of the supported Linux distribution versions. The process for package installation varies slightly across the different Linux versions, but the basic process is as follows:

Note that the multipath driver installation does require a reboot of the Linux system.

For Ubuntu, the following procedure describes how to install the multipath driver package:

1. Download the DEB package.
2. Verify that DEB package and Kernel Version Match.

Compare the package version and the kernel version to ensure they are an exact match. If they are not an exact  match, do not install the package. Instead, build the driver from source according to the instructions in the “Building and Installing the Driver’ section later in this document.

3. Install the DEB package.

# sudo apt-get install ./dist/dellnfs-modules_*-generic_amd64.deb

4. Check package is installed correctly.

# dpkg -l | grep dellnfs-modules

dellnfs-modules   2:4.0.22-dell.kver.5.4.0-150-generic amd64        NFS RDMA kernel modules

# dpkg -S /lib/modules/`uname -r`/updates/bundle/net/sunrpc/xprtrdma/rpcrdma.ko

dellnfs-modules: /lib/modules/5.4.0-150-generic/updates/bundle/net/sunrpc/xprtrdma/rpcrdma.ko

Regenerate the running kernel image.

# sudo update-initramfs -u -k `uname -r`

5. Reboot the Linux client.

# reboot

6. Confirm the module and services are running.

# dellnfs-ctl status

version: 4.0.22-dell
kernel modules: sunrpc rpcrdma compat_nfs_ssc lockd nfs_acl nfs nfsv3
services: rpcbind.socket rpcbind rpc-gssdrpc_pipefs: /run/rpc_pipefs

If the services are not up, run the ‘dellnfs-ctl reload’ command to start the services.

7. Verify an NFS mount

Attempt a mount, since NFS occasionally needs to create a symlink to rpc.statd. For example:

# mount -t nfs -o proto=rdma,port=20049,rsize=1048576,wsize=1048576,vers=3,nconnect=32,remoteports=10.231.180.95-10.231.180.98,remoteports_offset=1 10.231.180.95:/ifs/data/fio /mnt/test/

Created symlink /run/systemd/system/remote-fs.target.wants/rpc-statd.service → /lib/systemd/system/rpc-statd.service.

In the above, a symlink created on the first mount. Perform a reload to confirm the service is running correctly. For example:

# dellnfs-ctl reload

dellnfs-ctl: stopping service rpcbind.socket

dellnfs-ctl: umounting fs /run/rpc_pipefs

dellnfs-ctl: unloading kmod nfsv3

dellnfs-ctl: unloading kmod nfs

dellnfs-ctl: unloading kmod nfs_acl

dellnfs-ctl: unloading kmod lockd

dellnfs-ctl: unloading kmod compat_nfs_ssc

dellnfs-ctl: unloading kmod rpcrdma

dellnfs-ctl: unloading kmod sunrpc

dellnfs-ctl: loading kmod sunrpc

dellnfs-ctl: loading kmod rpcrdma

dellnfs-ctl: loading kmod compat_nfs_ssc

dellnfs-ctl: loading kmod lockd

dellnfs-ctl: loading kmod nfs_acl

dellnfs-ctl: loading kmod nfs

dellnfs-ctl: loading kmod nfsv3

dellnfs-ctl: mounting fs /run/rpc_pipefs

dellnfs-ctl: starting service rpcbind.socket

dellnfs-ctl: starting service rpcbind

 

Similarly, for OpenSUSE and SLES, the driver package installation steps are as follows:

1. Download the driver RPM package.
2. Verify that RPM package and Kernel Version Match.

Compare the package version and the kernel version to ensure they are an exact match. If they are not an exact  match, do not install the package. Instead, build the driver from source according to the instructions in the “Building and Installing the Driver’ section later in this document.

3. Install the downloaded RPM package.

# zypper in ./dist/dellnfs-4.0.22-kernel_5.14.21_150400.24.97_default.x86_64.rpm
Loading repository data...
Reading installed packages...
Resolving package dependencies...

The following NEW package is going to be installed:
  dellnfs

1 new package to install.

4. Check the installed files.

# rpm -qa | grep dell
dellnfs-4.0.22-kernel_5.14.21_150400.24.100_default.x86_64

5. Reboot the Linux client.

# reboot

6. Verify the services are started.

# systemctl start rpcbind
# systemctl start nfs
# systemctl status nfs
nfs.service - Alias for NFS client
     Loaded: loaded (/usr/lib/systemd/system/nfs.service; disabled; vendor preset: disabled)
     Active: active (exited) since Wed 2023-12-13 15:11:09 PST; 2s ago
    Process: 15577 ExecStart=/bin/true (code=exited, status=0/SUCCESS)
   Main PID: 15577 (code=exited, status=0/SUCCESS)

7. Verify the client driver is loaded with ‘dellnfs-ctl’ script.

# dellnfs-ctl status
version: 4.0.22
kernel modules: sunrpc
services: rpcbind.socket rpcbind
rpc_pipefs: /var/lib/nfs/rpc_pipefs

8. Verify an NFS mount

Attempt a mount, since NFS occasionally needs to create a symlink to rpc.statd. For example:

# mount -t nfs -o proto=rdma,port=20049,rsize=1048576,wsize=1048576,vers=3,nconnect=32,remoteports=10.231.180.95-10.231.180.98,remoteports_offset=1 10.231.180.95:/ifs/data/fio /mnt/test/

Created symlink /run/systemd/system/remote-fs.target.wants/rpc-statd.service → /lib/systemd/system/rpc-statd.service.

In the above, a symlink is created on the first mount. Next, perform a reload to confirm the service is running correctly. For example:

# dellnfs-ctl reload

dellnfs-ctl: stopping service rpcbind.socket

dellnfs-ctl: umounting fs /run/rpc_pipefs

dellnfs-ctl: unloading kmod nfsv3

dellnfs-ctl: unloading kmod nfs

dellnfs-ctl: unloading kmod nfs_acl

dellnfs-ctl: unloading kmod lockd

dellnfs-ctl: unloading kmod compat_nfs_ssc

dellnfs-ctl: unloading kmod rpcrdma

dellnfs-ctl: unloading kmod sunrpc

dellnfs-ctl: loading kmod sunrpc

dellnfs-ctl: loading kmod rpcrdma

dellnfs-ctl: loading kmod compat_nfs_ssc

dellnfs-ctl: loading kmod lockd

dellnfs-ctl: loading kmod nfs_acl

dellnfs-ctl: loading kmod nfs

dellnfs-ctl: loading kmod nfsv3

dellnfs-ctl: mounting fs /run/rpc_pipefs

dellnfs-ctl: starting service rpcbind.socket

dellnfs-ctl: starting service rpcbind

Unlike installing the driver, removing it does not typically require a client reboot, and can also be performed via the appropriate package manager for the Linux flavor.

Also, be aware that there are no upgrade or patching systems for the driver. This means that if a client’s kernel version is updated, the module must be re-built or a matching package re-installed. And correspondingly, if there is an update to the dellnfs package, the module must be re-installed.

# sudo update-initramfs -u -k `uname -r`

Next, perform a reload to confirm the service is running correctly. For example:

# dellnfs-ctl reload

dellnfs-ctl: stopping service rpcbind.socket

dellnfs-ctl: umounting fs /run/rpc_pipefs

dellnfs-ctl: unloading kmod nfsv3

dellnfs-ctl: unloading kmod nfs

dellnfs-ctl: unloading kmod nfs_acl

dellnfs-ctl: unloading kmod lockd

dellnfs-ctl: unloading kmod compat_nfs_ssc

dellnfs-ctl: unloading kmod rpcrdma

dellnfs-ctl: unloading kmod sunrpc

dellnfs-ctl: loading kmod sunrpc

dellnfs-ctl: loading kmod rpcrdma

dellnfs-ctl: loading kmod compat_nfs_ssc

dellnfs-ctl: loading kmod lockd

dellnfs-ctl: loading kmod nfs_acl

dellnfs-ctl: loading kmod nfs

dellnfs-ctl: loading kmod nfsv3

dellnfs-ctl: mounting fs /run/rpc_pipefs

dellnfs-ctl: starting service rpcbind.socket

dellnfs-ctl: starting service rpcbind

Note that there are no upgrade or patching systems available for the ‘dellnfs’ multipath driver module. If a Linux client’s kernel version is updated, the module must be rebuilt or a matching package reinstalled. Similarly, if there is an update to the ‘dellnfs’ package, the module must be reinstalled.

Uninstalling the multipath client driver does not require a reboot and can be performed using the standard package manager for the pertinent Linux distribution. This unloads the loaded module and then removes the files. As such, uninstallation is a fairly trivial process.

The package manager commands for each Linux version to remove a package are as follows:

OS	Package removal command
Ubuntu	sudo apt-get autoremove <package_name>
OpenSUSE/SLES	zypper remove -u <package_name>

In the next article in this series, we’ll take a look at the specifics of multipath driver binary package installation.

In order for success with large AI model customization, inferencing, and training, GPUs require data served to them quickly and efficiently. Compute and storage must be architected and provisioned accordingly, in order to eliminate potential bottlenecks in the infrastructure.

To meet this demand, the new PowerScale multipath client driver enables performance aggregation of multiple PowerScale nodes through a single NFS mount point to one or many compute nodes. As a result, this driver, in conjunction with OneFS GPUDirect support, has enabled Dell to deliver the first Ethernet storage solution to be certified for NVIDIA’s DGX SuperPOD.

SuperPOD is an AI-optimized data center architecture that delivers the formidable computational power required to train deep learning (DL) models at scale, accelerating time to outcomes which drive future innovation.

Using DGX A200, B200, or H200 GPU-based compute in concert with a PowerScale F710 clustered storage layer, NVIDIA’s SuperPOD is able to deliver groundbreaking performance.

Deployed as a fully integrated scalable system, SuperPOD is purpose-built for solving challenging computational problems across a diverse range of AI workloads. These include streamlining supply chains, building large language models, and extracting insights from petabytes of unstructured data.

The performance envelope delivered by DGX SuperPOD enables rapid multi-node training of LLMs at significant scale. This integrated approach of provisioning, management, compute, networking, and fast storage, enables a diverse system that can span data analytics, model development, and AI inferencing, right up to the largest, most complex transformer-based AI workloads, deep learning systems, and trillion-parameter generative AI models.

To drive the throughput required for larger NVIDIA SuperPOD deployments, NFS client connectivity to a PowerScale cluster needs to utilize both RDMA and nconnect, in addition to GPUDirect.

While the native Linux NFS stack supports their use, it does not allow nconnect and RDMA to be configured simultaneously.

To address this, the multipath driver permits Linux NFS clients to use RDMA in conjunction with nconnect mount options, while also increasing the maximum nconnect limit from 16 to 64 connections. Additionally, the SuperPOD solution mandates the use of the ‘localports_failover’ NFS mount option, which only works with RDMA currently.

The Dell multipath client driver can be of considerable performance benefit for workloads with streaming reads and writes to and from individual high-powered servers, particularly to multiple files within a single NFS mount – in addition to SuperPOD and BasePOD AI workloads. Conversely, single file streams, and multiple concurrent writes to the same file across multiple nodes typically don’t benefit substantially from the multipath driver.

Without the multipath client driver, a single NFS mount can only route to one PowerScale storage node IP address.

By way of contrast, the multipath driver allows NFS clients to direct I/O to multiple PowerScale nodes for higher aggregate single-client throughput.

The multipath driver enables a single NFS mount point to route to multiple node IP addresses. A group of IP addresses consists of one logical NFS client with the remote endpoint (cluster) using multiple remote machines (nodes), implementing a distributed server architecture.

The principle NFS mount options of interest with the multipath client driver are:

Mount option	Description
nconnect	Allows the admin to specify the number of TCP connections the client can establish between itself and the NFS server. It works with remoteports to spread load across multiple target interfaces.
localports	Mount option that allows a client to use its multiple NICs to multiplex I/O.
localports_failover	Mount option allowing transports to temporarily move from local client interfaces that are unable to serve NFS connections.
proto	The underlying transport protocol that the NFS mount will use. Typically, either TCP or RDMA.
remoteports	Mount option that allows a client to target multiple servers/ NICS to multiplex I/O. Remoteports spreads the load to multiple file handles instead of taking a single file handle to avoid thrashing on locks.
version	The version of the NFS protocol that is to be used. The multipath driver supports NFSv3, NFSv4.1, and NFSv4.2. Note that NFSv4.0 is unsupported.

There are also several advanced mount options which can be useful to squeeze out some extra throughput, particularly with SuperPOD deployments. These options include ‘remoteport offsets’, which can help with loading up L1 cache, and’ spread reads and writes’, which can assist with load balancing.

The Dell multipath driver is available for download on the Dell Support Site to any customer that has OneFS entitlement:

https://www.dell.com/support/home/en-us/product-support/product/isilon-onefs/drivers

There is no license requirement for this driver, nor charge for it, and its provided as both pre-built Linux package, and customer-compliable source code. There’s a README file included with the code that provides basic instruction.

This multipath client driver runs on both physical and virtual machines, and across several popular Linux distros. The following matrix shows the currently supported variants, plus the availability of a pre-compiled driver package and/or self-compilation option.

Linux distribution	Kernel version	Upstream driver version (minimum)	Multipath driver version	Package available	Self- compile
OpenSUSE 15.4	5.14.x	4.x	1.x	ü	ü
Ubuntu 20.04	5.4.x	4.x	1.x	ü	ü
Ubuntu 22.04	5.15.x	4.x	1.x	ü	ü

While the multipath driver’s major release version—1.x—is correct in the table, the second digit release number will be frequently incremented as updated versions of the multipath client driver are released.

By design, the multipath driver only supports newer and most recent versions of the popular Linux distributions. Older Linux kernel versions often do not support full NFS client functionality, particularly for the ‘–remoteports’ and/or ‘–localports’ mount configuration options. Additionally, older and end-of-life Linux versions can often present significant security risks, especially once current vulnerability patches and hotfixes are no longer being made available.

Both x86 CPU architectures and GPU-based platforms, such as the NVIDIA DGX range, are supported.

Linux system	Processor type	Example
Physical	CPU	Dell PE R760
Physical	GPU	Dell PE XE9680 NVIDIA DGX H100
Virtual machine	CPU	VMware ESXi
Virtual machine	GPU	VMware vDGA

While there is no specific NFS or OneFS core configuration required on the PowerScale cluster side for multipath driver support, there are a couple of basic prerequisites The following OneFS support matrix on the top right of this slide lays out which driver functionality is available in what release, from OneFS 9.5 to current.

Version	NFSv3, NFSv4.1 TCP	NFSv3 RDMA	NFSv4.1 RDMA	NVIDIA SuperPOD
OneFS 9.5	Yes	Yes	No	No
OneFS 9.7	Yes	Yes	Yes	No
OneFS 9.9	Yes	Yes	Yes	Yes

Also note that OneFS 9.9 is required for any NVIDIA SuperPOD deployments, because there are some performance optimizations in 9.9 specifically for that platform.

Additionally, both the dynamic and static network pools can be configured on the cluster for use with the multipath driver. If F710 nodes are being deployed in the cluster, OneFS 9.7 or later is required.

Note that when deploying an NVIDIA SuperPOD or BasePOD solution, the reference architecture mandates a PowerScale cluster composed of F710 all-flash nodes running OneFS 9.9 or later.

 

For a Linux client to successfully connect to a PowerScale cluster using the multipath driver it currently must be running one of the following Linux flavors:

Supported Linux Distribution	Kernel Version
OpenSUSE 15.4	5.14.x
Ubuntu 20.04	5.4.x
Ubuntu 22.04	5.15.x

By design, the multipath driver only supports newer and most recent versions of the popular Linux distributions. Older Linux kernels often don’t include full NFS client functionality, particularly for the ‘–remoteports’ and ‘–localports’ mount options. You will also likely notice the conspicuous absence of Red Hat Enterprise Linux from this matrix. However, engineering do anticipate supporting both RHEL 8 and 9 in a near-future version.

There are also a couple of additional client prerequisites that must be met:

• If RDMA is being configured, the client must contain an RDMA-capable Ethernet NIC, such as the Mellanox CX series.
• The Linux client should have the ‘trace-cmd’ package installed, along with NFS client related packages.

In the next article in this series, we’ll take a look at the specifics of the multipath driver binary package installation.