Contents
- Overview of Cisco usNIC
- Cisco usNIC Prerequisites
- Configuring Cisco usNIC
- Creating a Cisco usNIC Connection Policy using the Cisco Manager GUI
- Configuring a usNIC Ethernet Adapter Policy
- Modifying a usNIC using the Cisco UCS Manager GUI
- Creating a usNIC using the Cisco UCS Manager CLI
- Modifying a usNIC using the Cisco UCS Manager CLI
- Deleting a usNIC using the Cisco UCS Manager CLI
- Configuring the Linux Kernel for Cisco usNIC
- Installing Linux Software Packages for Cisco usNIC
- Source code for Linux Cisco usNIC software packages
- Manually Loading the Kernel Modules for Cisco usNIC
- Uninstalling Linux Software Packages for Cisco usNIC
- Upgrading the Linux Software Packages for Cisco usNIC
- Installing MPI
- Installing Community Open MPI
- Installing IBM Spectrum MPI
- Installing the Intel MPI Library
- Adding MPI to User Environments
- Environment for IBM Spectrum MPI Library
- Environment for Community Open MPI
- Environment for the Intel ® MPI Library
- Adding Libfabric to User Environments
- Verifying the Cisco usNIC Installation for Cisco UCS B-Series Blade Servers
- Troubleshooting Information
First Published:
Last Updated:
Text Part Number:
Overview of Cisco usNIC
The Cisco user-space NIC (Cisco usNIC) feature improves the performance of software applications that run on the Cisco UCS servers in your data center by bypassing the kernel when sending and receiving networking packets. The applications interact directly with a Cisco UCS VIC second generation or later adapter, which improves the networking performance of your high-performance computing cluster. To benefit from Cisco usNIC, your applications must use the Message Passing Interface (MPI) or Libfabric interface instead of sockets or other communication APIs.
Cisco usNIC offers the following benefits for your applications:
Provides a low-latency and high-throughput communication transport.
Employs the standard and application-independent Ethernet protocol.
Low jitter, or near constant latency, communications.
Standard Ethernet applications use user-space socket libraries, which invoke the networking stack in the Linux kernel. The networking stack then uses the Cisco eNIC driver to communicate with the Cisco VIC hardware. The following figure shows the contrast between a regular software application and an MPI application that uses usNIC.
Cisco usNIC Prerequisites
To benefit from Cisco usNIC, your configuration has the following prerequisites:
A supported Linux operating system distribution release. For more information on supported Linux operating system releases, please refer to the UCS Hardware and Software Compatibility Tool.
The UCS Driver ISO corresponding to the UCS server model, selected Linux operating system, and version of UCS firmware installed on the server as identified by the UCS Hardware and Software Compatibility Tool. For more information, see Downloading Cisco UCS VIC drivers.
A supported MPI implementation, such as IBM Sprectrum MPI, the open source Community Open MPI package, or version 4 or 5 of the Intel ® MPI Library. If the Intel ® MPI Library is used, the network must be configured with flow control enabled.
Configuring Cisco usNIC
The overall flow to configure Cisco usNIC is as follows:
Create or Modify a Service Profile to support usNIC
Install supported Linux OS (if not already installed)
Configure Linux kernel and OS to support usNIC
Install usNIC drivers and utilities
Install libfabric and MPI software
Verify the usNIC installation
Creating a Cisco usNIC Connection Policy using the Cisco Manager GUI
ProcedureYou can use the procedure described below or click Play on this video to watch how a Cisco usNIC Connection policy can be created.
Modifying a usNIC using the Cisco UCS Manager GUI
Procedure
Step 1 In the Navigation pane, click Servers. Step 2 On the Servers tab, expand . Step 3 Expand the service profile node where you want to configure the usNIC and click vNICs. Step 4 In the Work pane, click the Network tab. Step 5 In the vNICs area, choose a vNIC and click Modify. Step 6 In the Adapter Performance Profile area of the Modify vNIC dialog box, choose Linux from the Adapter Policy drop-down list. Step 7 In the Connection Policies area, click the usNIC radio button. Step 8 Choose the usNIC connection policy that you created from the usNIC Connection Policy drop-down list. Step 9 Click OK. Step 10 Click Save Changes. Step 11 In the Navigation pane, click the service profile that you just modified. Step 12 In the Work pane, click the Policies tab. Step 13 Expand the BIOS Policy bar and choose usNIC in the BIOS Policy drop-down list. Step 14 Click Save Changes.
Creating a usNIC using the Cisco UCS Manager CLI
Procedure
Command or Action Purpose
Step 1 UCS-A # scope service-profile server chassis-id / blade-id or rack_server-id Enters the service profile for the specified chassis, blade or UCS managed rack server ID.
Step 2 UCS-A /org/service-profile # show vnic Displays the vnics that are available on the server. A usNIC vNIC is available by default when you upgrade to Cisco UCS Manager, release 2.2. Step 3 UCS-A /org/service-profile # scope vnic vnic name Enters the vNIC mode for the specified vNIC. Step 4 UCS-A /org/service-profile/vnic # set adapter-policy Linux Specifies Linux and the adapter policy for the usNIC. Step 5 UCS-A /org/service-profile/vnic # enter usnic-conn-policy-ref usnic connection policy reference name Creates the usNIC connection policy reference for the vNIC with the specified name. The maximum size for the connection policy name is 16 characters. Step 6 UCS-A /org/service-profile/vnic/usnic-conn-policy-ref* # commit-buffer Commits the transaction to the system configuration. Step 7 UCS-A /org/service-profile/vnic/usnic-conn-policy-ref # top Enters the top-level mode. Step 8 UCS-A # scope org Enters the root organization mode. Step 9 UCS-A /org # create usnic-conn-policy usnic connection policy name Creates a usNIC connection policy with the specified name. Step 10 UCS-A /org/usnic-conn-policy* # set usnic-count number of usnics Specifies the number of Cisco usNICs to create. It is recommended that you enter 58 for this value.
Step 11 UCS-A /org/usnic-conn-policy* # set adaptor-profile usNIC Specifies the usNIC Ethernet adaptor profile for the usNIC connection policy. This usNIC adaptor profile is created by default when you upgrade from previous versions of Cisco UCS Manager to release 2.2. Step 12 UCS-A /org/usnic-conn-policy* # commit-buffer Commits the transaction to the system configuration. This example shows how to create a Cisco usNIC and specify its properties:
Server # scope org Server /org # create usnic-conn-policy usnic1 Server /org/usnic-conn-policy* # set usnic-count 58 Server /org/usnic-conn-policy* # set adaptor-profile usNIC Server /org/usnic-conn-policy* # commit-buffer Server /org/usnic-conn-policy # top Server # scope service-profile server 1/1 Server /org/service-profile # show vnic vNIC: Name Fabric ID Dynamic MAC Addr Virtualization Preference ------------------ --------- ------------------ ------------------------- eth0 A 00:25:B5:00:00:A1 NONE eth1 B 00:25:B5:00:00:A2 NONE eth2 A 00:25:B5:00:00:A3 NONE Server /org/service-profile # scope vnic eth0 Server /org/service-profile/vnic # set adapter-policy Linux Server /org/service-profile/vnic # enter usnic-conn-policy-ref usnic1 Server /org/service-profile/vnic/usnic-conn-policy-ref* # commit-buffer Server /org/service-profile/vnic/usnic-conn-policy-ref # exitModifying a usNIC using the Cisco UCS Manager CLI
Procedure
Command or Action Purpose
Step 1 UCS-A # scope service-profile server chassis-id / blade-id or rack_server-id Enters the service profile for the specified chassis, blade or UCS managed rack server ID.
Step 2 UCS-A /org/service-profile # show vnic Displays the vnics that are available on the server. A usnic vnic is available by default when you upgrade to Cisco UCS Manager, release 2.2. Step 3 UCS-A /org/service-profile # scope vnic vnic name Enters the vnic mode for the specified vNIC. Step 4 UCS-A /org/service-profile/vnic # enter usnic-conn-policy-ref usnic connection policy reference name Specifies the usnic connection policy reference for the vNIC that you want to use. Step 5 UCS-A /org/service-profile/vnic/usnic-conn-policy-ref* # commit-buffer Commits the transaction to the system configuration. This example shows how to modify Cisco usNIC properties:
Server # scope service-profile server 1/1 Server /org/service-profile # show vnic vNIC: Name Fabric ID Dynamic MAC Addr Virtualization Preference ------------------ --------- ------------------ ------------------------- eth0 A 00:25:B5:00:00:A1 SRIOV USNIC eth1 B 00:25:B5:00:00:A2 NONE eth2 A 00:25:B5:00:00:A3 NONE Server /org/service-profile # scope vnic eth0 Server /org/service-profile/vnic # enter usnic-conn-policy-ref usnic2 Server /org/service-profile/vnic/usnic-conn-policy-ref* # commit-buffer Server /org/service-profile/vnic/usnic-conn-policy-ref # exitDeleting a usNIC using the Cisco UCS Manager CLI
Procedure
Command or Action Purpose
Step 1 UCS-A # scope service-profile server chassis-id / blade-id or rack_server-id Enters the service profile for the specified chassis, blade or UCS managed rack server ID.
Step 2 UCS-A /org/service-profile # show vnic Displays the vNICs that are available on the server. A usNIC vNIC is available by default when you upgrade to Cisco UCS Manager, release 2.2. Step 3 UCS-A /org/service-profile # scope vnic vnic name Enters the vNIC mode for the specified vNIC. Step 4 UCS-A /org/service-profile/vnic # show usnic-conn-policy-ref usnic connection policy reference name Specifies the usNIC connection policy reference for the vNIC that you want to use. Step 5 UCS-A /org/service-profile/vnic # delete usnic-conn-policy-ref usnic connection policy reference name Deletes the specified usNIC connection policy reference. Step 6 UCS-A /org/service-profile/vnic/usnic-conn-policy-ref* # commit-buffer Commits the transaction to the system configuration. This example shows how to modify Cisco usNIC properties:
Server # scope service-profile server 1/1 Server /org/service-profile # show vnic vNIC: Name Fabric ID Dynamic MAC Addr Virtualization Preference ------------------ --------- ------------------ ------------------------- eth0 A 00:25:B5:00:00:A1 SRIOV USNIC eth1 B 00:25:B5:00:00:A2 NONE eth2 A 00:25:B5:00:00:A3 NONE Server /org/service-profile # scope vnic eth0 Server /org/service-profile/vnic # show usnic-conn-policy-ref usNIC Connection Policy Reference: usNIC Connection Policy Name ---------------------------- usnic2 Server /org/service-profile/vnic # delete usnic-conn-policy-ref usnic2 Server /org/service-profile/vnic* # commit-buffer Server /org/service-profile/vnic # exitConfiguring the Linux Kernel for Cisco usNIC
Before You BeginProcedure
Make sure that the following software and hardware components are installed on the Cisco UCS server:
A supported Linux operating system distribution release. For more information, see the UCS Hardware and Software Compatibility Tool.
GCC, G++, and Gfortran
DAT user library (if using Intel ® MPI)
libnl user library development package (either version 1 or version 3)
Cisco UCS VIC second generation or later adapter
Step 1 Ensure the kernel option CONFIG_INTEL_IOMMU is selected in the kernel. Enable the Intel IOMMU driver in the Linux kernel by manually adding 'intel_iommu =on' in the grub.conf file (for example, /boot/grub/grub.conf). For example, if your grub.conf file contains a "kernel" line such as kernel (hd0,0)/vmlinuz LANG=en_US.UTF-8 KEYTABLE=us, then you will add 'intel_iommu=on' to the end as shown below:
kernel (hd0,0)/vmlinuz LANG=en_US.UTF-8 KEYTABLE=us intel_iommu=on
For Red Hat Enterprise Linux use the grubby command line tool to add intel_iommu=on to the configuration file.
grubby --args="intel_iommu=on" --update-kernel /boot/vmlinuz-`uname -r`For SLES, use the grub2 command line tool to add intel_iommu=on to the GRUB_CMDLINE_LINUX_DEFAULT option found in /etc/default/grub configuration file then run the grub2 command below to apply the changes.
run grub2-mkconfig -o /boot/grub2/grub.cfgFor Ubuntu, use the grub2 command line tool to add intel_iommu=on to the GRUB_CMDLINE_LINUX_DEFAULT option found in /etc/default/grub configuration file then run the update-grub command below to apply the changes.
update-grubStep 2 Reboot your Cisco UCS server. You must reboot your server for the changes to take after you enable the Intel IOMMU.
Step 3 Verify that the running kernel has booted with the intel_iommu=on option. $ cat /proc/cmdline | grep iommuStep 4 Install the Cisco usNIC Linux drivers. For more information about installing the drivers, see "Installing Linux Drivers" section in the guide.
Note The Cisco usNIC packages do not support the upgrade or downgrade of an operating system. To update the operating system, first uninstall the usNIC packages, update the operating system, and then reinstall the usNIC drivers.
Alternatively, you can update the operating system, uninstall the usNIC drivers, and then reinstall the usNIC drivers.
Installing Linux Software Packages for Cisco usNIC
ProcedureThe following section lists the content of the usNIC folder, specific for each supported Linux operating system distribution that is included in the UCS Drivers ISO bundle. Documentation about known issues and installation instructions are also included in the README file in the usNIC folder.
kmod-usnic_verbs-{version}.x86_64.rpm—Linux kernel verbs driver for the usNIC feature of the Cisco VIC SR-IOV Ethernet NIC.
libdaplusnic-{version}.x86_64.rpm— User space library DAPL plugin for usNIC.
usnic_tools-{version}.x86_64.rpm — Utility programs for usNIC.
usd_tools-{version}.x86_64.rpm — Additional diagnostic tools for usNIC.
libfabric-cisco-{version}.x86_64.rpm— Libfabric package with built-in support for the Cisco usNIC transport.
libfabric-cisco-devel-{version}.x86_64.rpm — Development headers for the Libfabric package with built-in support for the Cisco usNIC transport.
libusnic_verbs-{version}.x86_64.rpm— A dummy library that causes the libibverbs library to skip Cisco usNIC Linux devices (because Cisco usNIC functionality is exposed through libfabric, not libibverbs).
Step 1 Upgrade to the latest version of the enic driver included in the Cisco UCS ISO for your Linux distribution as documented in Cisco UCS Virtual Interface Card Drivers for Linux Installation Guide. Step 2 Install the Cisco usNIC software packages from the Cisco UCS Drivers ISO for your Linux distribution. Step 3 Enable the Linux RDMA services. Once enabled, RDMA services will be started automatically after a system reboot. The exact command to enable Linux RDMA may vary between each operating system. For example, on Red Hat Enterprise Linux 6.X use the following command: # chkconfig rdma on
Note You may need to perform this step on some Linux operating systems distributions, such as RHEL 6.4.
Step 4 Reboot your server for the installation changes to take effect automatically. By rebooting at this point and performing the installation validation steps found below you will be confident your system is configured properly. Important:It is recommended that you install all the binary RPMs from the ISO. Building and installing source code packages is intended for advanced users who are familiar with Linux device driver development environments.
Manually Loading the Kernel Modules for Cisco usNIC
Assuming that the operating system was booted with Intel IOMMU support enabled, you can manually load the Cisco usNIC kernel modules with the following steps.
Before You BeginProcedure
Ensure you delete all the existing versions of the driver before you load the latest version of the driver. This will help you configure the system successfully.
Command or Action Purpose
Step 1 # rmmod enic
Unloads the existing enic driver module.
Note Make sure that you are not logged into the OS using the enic driver, for example, via SSH to a VIC IP interface. Otherwise, your network connection will get disconnected. Alternatively, you can log in to the server using the KVM to perform this step.
Step 2 # modprobe enic
Loads the enic driver module.
Step 3 # dmesg | grep 'Cisco VIC Ethernet NIC Driver'
Verify that the correct version of the enic driver was loaded. The output from this command should show a version string that matches the version of the enic RPM from the Cisco UCS Driver ISO that you just installed.
Step 4 # modprobe usnic_verbs
Loads the usnic_verbs driver module.
Step 5 # lsmod | grep usnic_verbs
Verify that the usnic_verbs module loaded successfully. If it did, you should see some output. If the usnic_verbs module did not load, you should see no output.
Installing MPI
Installing Community Open MPI
Procedure
Step 1 Download the latest version of Open MPI from https://www.open-mpi.org/software/ompi/current/. This URL will automatically redirect you to the current release. Step 2 Extract the Open MPI tarball with the following command: $ tar xf openmpi-VERSION.tar.bz2Step 3 In the directory that is created, run the configure command with the following options: $ cd openmpi-VERSION $ ./configure \ --prefix=INSTALL_DIRECTORY \ --with-usnic \ --with-libfabric=/opt/cisco/libfabric \ --without-memory-manager \ --enable-mpirun-prefix-by-default \ --enable-mca-no-build=btl-openib,common-verbs,oob-ud \ LDFLAGS="-Wl,-rpath -Wl,opt/cisco/libfabric/lib -Wl,--enable-new-dtags"Substitute an appropriate directory for "INSTALL_DIRECTORY" (e.g., /opt/openmpi/VERSION).
Note Note that the above configure command will build an Open MPI without verbs support (because older versions of the libibverbs library will issue needless warnings to stderr about usNIC devices). You can remove the --enable-mca-no-build option if you need your Open MPI to support the Linux verbs API.
Step 4 At the successful conclusion of configure, build Open MPI: $ make
Note Parallel builds are fully supported (which can greatly speed up build times) via the "-j" option. For example, on an idle 16-core server, you can "make -j 32" to allow "make" to utilize up to 32 concurrent build commands.
Step 5 At the successful conclusion of make, install Open MPI (you may need root or specific user permissions to write to your chosen INSTALL_DIRECTORY): $ make install
Adding MPI to User Environments
Before MPI applications can be compiled and launched, an MPI implementation must be added to each user's environment. It is recommended that you only add one MPI implementation to a user's environment at a time.
Environment for IBM Spectrum MPI Library
Spectrum MPI needs very little in terms of environment setup. You can invoke Spectrum MPI's commands either through an absolute path or by adding the directory with the Spectrum MPI commands in to your PATH, and then invoking them without an absolute path.
Environment for Community Open MPI
Community Open MPI requires that you add its installation binary and library paths to the environment variables. It is usually best to add these paths to the environment in your shell startup files so that they are automatically set upon login to all nodes in your cluster.
Specifically, you should prepend the PATH environment variable with INSTALL_DIRECTORY/bin, and prefix the LD_LIBRARY_PATH environment variable with INSTALL_DIRECTORY/lib. Optionally, you can also add the INSTALL_DIRECTORY/share/man to the MANPATH environment variable.
For example, if you configured Community Open MPI with an INSTALL_DIRECTORY of /opt/openmpi, if you are using Bash as your login shell, you can add these lines to your shell startup file (which is typically $HOME/.bashrc):
export PATH=/opt/openmpi/bin:$PATH export LD_LIBRARY_PATH=/opt/openmpi/lib:$LD_LIBRARY_PATH export MANPATH=/opt/openmpi/share/man:$MANPATHAlternatively, if you're using C shell as your login shell, you can add these lines to your shell startup file (which is typically $HOME/.cshrc):
set path=(/opt/openmpi/bin $path) setenv LD_LIBRARY_PATH /opt/openmpi/lib:$LD_LIBRARY_PATH if ("1" == "$?MANPATH") then setenv MANPATH /opt/openmpi/share/man:${MANPATH} else setenv MANPATH /opt/openmpi/share/man: endifYour system may require slightly different commands. Check the Open MPI Community FAQ (https://www.open-mpi.org/faq/) in the "Running MPI Jobs" section for more information about how to set your PATH, LD_LIBRARY_PATH, MANPATH environment variables.
Environment for the Intel ® MPI Library
In addition to the instructions provided by the Intel ® MPI Library documentation, additional environment variables must be set in each user's environment to enable Cisco usNIC functionality. Two scripts are installed by the libdaplusnic software package to help set the required environment variables. One script is for Bourne shell users; the other script is for C shell users:
/opt/cisco/intelmpi-usnic-vars.sh (For Bourne shell users)
/opt/cisco/intelmpi-usnic-vars.csh (For C shell users)
The appropriate script should be sourced as part of the users's shell startup / login sequence.
Using the Intel ® MPI Library with usNIC requires the network to be configured with flow control enabled. This can be either IEEE 802.3x link-level flow control or IEEE 802.1Qbb Priority-based Flow Control (PFC). This feature is sometimes also called "no-drop." Refer to the configuration guide for the switches in your network for information about enabling flow control. If flow control is not enabled in the network, then applications utilizing the Intel® MPI Library may work correctly, but possibly with extremely degraded performance.
In deployments of the Intel ® MPI Library, the MPI traffic must have flow control enabled on all Cisco usNIC ports, and no-drop or platinum QoS system Class with default CoS value 5 in Cisco UCS Manager. Please refer to the Cisco UCS Manager Network Management guide section on "Quality of Service".
Adding Libfabric to User Environments
If you are developing Libfabric-specific applications, you may benefit from having the Libfabric test executables (such as fi_pingpong) and/or man pages in your environment. Two scripts are installed by the Cisco libfabric package to help set the required environment variables. One script is for Bourne shell users, the other is for C shell users:
The appropriate script should be sourced as part of the users's shell startup / login sequence.
Verifying the Cisco usNIC Installation for Cisco UCS B-Series Blade Servers
ProcedureAfter you install the required Linux drivers for Cisco usNIC, perform the following procedure at the Linux prompt to make sure that the installation completed successfully.
Note
The examples shown below are configurations verified on Linux operating system distribution RHEL 6.5.
Step 1 Search and verify if the usnic_verbs kernel module was loaded during the OS driver installation. $ lsmod | grep usnic_verbsThe following details are displayed when you enter the lsmod | grep usnic_verbs command. The kernel modules listed on your console may differ based on the modules that you have currently loaded in your OS.
usnic_verbs 73762 2 ib_core 74355 2 ib_uverbs,usnic_verbs enic 73723 1 usnic_verbsStep 2 View the configuration of Cisco usNIC-enabled NICs. $ /opt/cisco/usnic/bin/usnic_devinfoThe following section is a brief example of the results that are displayed when you execute the usnic_devinfo command. The results may differ based on your current installation. When the results are displayed on your console, ensure that the link state for each of the listed ports are shown as UP.
The following example shows two interfaces (usnic_1 and usnic_0) that are configured on a Cisco UCS VIC adapter. If you configured only one Cisco usNIC-enabled vNIC, you will see a listing for only usnic_0.usnic_0: Interface: eth3 MAC Address: 00:25:b5:31:32:10 IP Address: 10.10.10.2 Netmask: 255.255.255.0 Prefix len: 24 MTU: 9000 Link State: UP Bandwidth: 40 Gb/s Device ID: UCSB-MLOM-40G-03 [VIC 1340] [0x012c] Vendor ID: 4407 Vendor Part ID: 207 Firmware: 4.1(3S1) VFs: 58 CQ per VF: 6 QP per VF: 6 Interrupts per VF: 6 Max CQ: 348 Max CQ Entries: 65535 Max QP: 348 Max Send Credits: 4095 Max Recv Credits: 4095 Capabilities: Map per res: yes PIO sends: yes CQ interrupts: no usnic_1: Interface: eth4 MAC Address: 00:25:b5:31:32:20 IP Address: 10.10.10.31 Netmask: 255.255.255.0 Prefix len: 24 MTU: 9000 Link State: UP Bandwidth: 40 Gb/s Device ID: UCSB-MLOM-40G-03 [VIC 1340] [0x012c] Vendor ID: 4407 Vendor Part ID: 207 Firmware: 4.1(3S1) VFs: 58 CQ per VF: 6 QP per VF: 6 Interrupts per VF: 6 Max CQ: 348 Max CQ Entries: 65535 Max QP: 348 Max Send Credits: 4095 Max Recv Credits: 4095 Capabilities: Map per res: yes PIO sends: yes CQ interrupts: noStep 3 Run the usnic_check script to view the installed RPMs and their versions. $ /opt/cisco/usnic/bin/usnic_check enic RPM version 2.3.0.39 installed usnic_verbs RPM version 1.1.527.0.rhel6u5 installed libdaplusnic RPM version 2.0.39cisco1.0.527.0 installed libfabric RPM version 1.4.0cisco1.0.527.0.rhel6u5 installed libusnic_verbs RPM version 2.0.3.527.0.rhel6u5 installed
Note If you installed any components from the source code in the Cisco usNIC software packages, the usnic_check script will report the corresponding RPM as missing. It is recommended that you do not mix source code and binary package installations.
Step 4 Verify that the Cisco usNIC network packets are being transmitted correctly between the client and server hosts.
The following example shows the results that are displayed when you run the fi_pingpong program.Server-side: [server]$ /opt/cisco/libfabric/bin/fi_pingpong bytes #sent #ack total time MB/sec usec/xfer Mxfers/sec 64 10k =10k 1.2m 0.07s 17.84 3.59 0.28 256 10k =10k 4.8m 0.08s 66.23 3.87 0.26 1k 10k =10k 19m 0.10s 199.76 5.13 0.20 4k 10k =10k 78m 0.18s 466.60 8.78 0.11 Client-side: [client]$ /opt/cisco/libfabric/bin/fi_pingpong -p usnic SERVER_IP_ADDRESS bytes #sent #ack total time MB/sec usec/xfer Mxfers/sec 64 10k =10k 1.2m 0.07s 17.84 3.59 0.28 256 10k =10k 4.8m 0.08s 66.25 3.86 0.26 1k 10k =10k 19m 0.10s 199.77 5.13 0.20 4k 10k =10k 78m 0.18s 466.61 8.78 0.11
Note fi_pingpong is not a high-performance benchmark. It shows the general performance levels of your usNIC devices, but it is not highly tuned to show the absolute best performance. The fi_pingpong output should only be used to ensure that performance is in the general neighborhood of expected performance.
Step 5 Download, compile, and execute the ring_c test program to validate that the MPI traffic is correctly transmitted between the client and server hosts. You can obtain the ring_c test program from this link: https://raw.githubusercontent.com/open-mpi/ompi/v2.x/examples/ring_c.c.
The following example shows how to use the wget utility to obtain, compile, and execute the ring_c. Alternatively, you can use other methods of obtaining and running the test program.
Note Run the following commands with a single MPI implementation setup in your environment.
$ wget --no-check-certificate https://raw.githubusercontent.com/open-mpi/ompi/v2.x/examples/ring_c.c --2017-03-21 19:46:20-- https://raw.githubusercontent.com/open-mpi/ompi/v2.x/examples/ring_c.c Resolving raw.githubusercontent.com... 151.101.192.133, 151.101.64.133, 151.101.128.133, ... Connecting to raw.githubusercontent.com|151.101.192.133|:443... connected. HTTP request sent, awaiting response... 200 OK Length: 2416 (2.4K) [text/plain] Saving to: ‘ring_c.c’ ring_c.c 100%[======================>] 2.36K --.-KB/s in 0s 2017-03-21 19:46:20 (29.5 MB/s) - ‘ring_c.c’ saved [2416/2416] $ mpicc ring_c.c -o ring_c [no output] # IBM Spectrum MPI: $ /path/to/mpirun --usnic --host host1,host2 -n 2 ./ring_c # Community Open MPI: $ mpirun --mca btl usnic,vader,self --host host1,host2 -n 2 ./ring_c # The expected output from both IBM Spectrum MPI and Community Open MPI is: Process 0 sending 10 to 1, tag 201 (4 processes in ring) Process 0 sent to 1 Process 0 decremented value: 9 Process 0 decremented value: 8 Process 0 decremented value: 7 Process 0 decremented value: 6 Process 0 decremented value: 5 Process 0 decremented value: 4 Process 0 decremented value: 3 Process 0 decremented value: 2 Process 0 decremented value: 1 Process 0 decremented value: 0 Process 0 exiting Process 2 exiting Process 1 exiting Process 3 exiting ...
Note If desired, setup a different MPI implementation in your environment and re-run the mpicc and mpirun commands to verify that MPI implementation with Cisco usNIC functionality.
If the fi_pingpong program and the ring_c program executed successfully, you should now be able to run general MPI applications over Cisco usNIC.
Troubleshooting Information
Problem
Verifying that Cisco usNIC packets are being transmitted correctly between client and server using fi_pingpong causes the following errors:
Solution
Make sure that valid Cisco usNIC connection policy is configured in usNIC Connection Policies and assigned to the vNICs in the service profile.
Make sure that IP addresses of the Cisco usNIC devices on both the server and client are configured correctly.
Make sure that the client pingpong is attempting to send packets to the correct server IP address of Cisco usNIC device.
Problem
Running the Cisco usNIC traffic using the mpirun causes the following errors:
MTU size mismatch error. See the following example:Example: # Enter the command below at the prompt on a single line from mpirun up to Sendrecv. # The backslash is included here as a line continuation and is not needed when the command is # entered at the prompt. $ mpirun --host node05,node06 -np 12 --mca btl usnic,vader,self \ --mca btl_usnic_if_include usnic_1 IMB-MPI1 Sendrecv The MTU does not match on local and remote hosts. All interfaces on all hosts participating in an MPI job must be configured with the same MTU. The usNIC interface listed below will not be used to communicate with this remote host. Local host: node05 usNIC interface: usnic_1 Local MTU: 8958 Remote host: node06 Remote MTU: 1458Solution
Make sure that the MTU size has been set correctly on the VLANs and QoS.
See: Configuring QoS System Classes with the LAN Uplinks Manager.
Problem
Installing a Cisco enic driver causes the following Cisco enic dependency errors:# rpm -ivh kmod-usnic_verbs-1.0.4.318.rhel6u5-1.x86_64.rpm error: Failed dependencies: ksym(enic_api_devcmd_proxy_by_index) = 0x107cb661 is needed by kmod-usnic_verbs-1.0.4.318.rhel6u5-1.x86_64 ksym(vnic_dev_alloc_discover) = 0xfb7e4707 is needed by kmod-usnic_verbs-1.0.4.318.rhel6u5-1.x86_64 ksym(vnic_dev_get_pdev) = 0xae6ae5c9 is needed by kmod-usnic_verbs-1.0.4.318.rhel6u5-1.x86_64 ksym(vnic_dev_get_res) = 0xd910c86b is needed by kmod-usnic_verbs-1.0.4.318.rhel6u5-1.x86_64 ksym(vnic_dev_get_res_bar) = 0x31710a7e is needed by kmod-usnic_verbs-1.0.4.318.rhel6u5-1.x86_64 ksym(vnic_dev_get_res_bus_addr) = 0x7be7a062 is needed by kmod-usnic_verbs-1.0.4.318.rhel6u5-1.x86_64 ksym(vnic_dev_get_res_count) = 0x759e4b07 is needed by kmod-usnic_verbs-1.0.4.318.rhel6u5-1.x86_64 ksym(vnic_dev_get_res_type_len) = 0xd122f0a1 is needed by kmod-usnic_verbs-1.0.4.318.rhel6u5-1.x86_64 ksym(vnic_dev_unregister) = 0xd99602a1 is needed by kmod-usnic_verbs-1.0.4.318.rhel6u5-1.x86_64 #Solution
Ensure that the correct enic driver has been installed. In addition, make sure of the following:
Specifically, you must ensure the following: the enic and usnic_verbs drivers must match. If you have a mismatch, you can get the above version errors.
Specifically, the enic and usnic_verbs that come in the Cisco UCS drivers ISO must be matched together. If you use an enic from one Cisco UCS driver ISO and usnic_verbs from another Cisco UCS driver ISO, it will result in the above version errors.
Problem
Intel IOMMU causes the following warnings:# rpm -ivh kmod-usnic_verbs-1.0.4.318.rhel6u5-1.x86_64.rpm Preparing... ########################################### [100%] 1:kmod-usnic_verbs ########################################### [100%] WARNING - Intel IOMMU does not appear to be enabled - please add kernel parameter intel_iommu=on to your boot configuration for USNIC driver to function. #Problem
Installing DAT user library can cause the following failed dependencies errors:
# rpm -ivh libdaplusnic-2.0.39cisco1.0.0.317-1.el6.x86_64.rpm error: Failed dependencies: dapl is needed by libdaplusnic-2.0.39cisco1.0.0.317-1.el6.x86_64 #Copyright © 2016, Cisco Systems, Inc. All rights reserved.