Cluster Configuration#
This documentation is part of NVIDIA DGX BasePOD: Deployment Guide Featuring NVIDIA DGX A100 Systems.
Note
Before you complete the steps in this documentation, complete Deployment.
Warning
The #
prompt indicates commands that you execute as the root user on a head node. The %
prompt indicates commands that you execute within cmsh.
Cluster Configuration Steps#
Log in to the Base Command Manager (BCM) head node assigned to externalnet.
1ssh <externalnet>
Install the cluster license by running the request-license command. Because HA is used, specify the MAC address of the first NIC of the secondary head node so that it can also serve the BCM licenses in the event of a failover.
This example is for a head node with Internet access. For air-gapped clusters, see “Off-cluster WWW access” in Section 4.3.3 of the NVIDIA Base Command Manager Installation Manual.
1# request-license 2Product Key (XXXXXX-XXXXXX-XXXXXX-XXXXXX-XXXXXX): 123456-123456-123456-123456 3...
Backup the default software image. The backup image can be used to create additional software images.
1# cmsh 2% softwareimage 3% clone default-image default-image-orig 4% commitWait for the ramdisk to be regenerated and the following text to be displayed.
1Wed Jul 26 09:00:53 2023 [notice] bcm10-headnode: Initial ramdisk for image default-image-orig was generated successfully
Backup the DGX software image. The backup image can be used to create additional software images.
1% softwareimage 2% clone dgx-os-6.0-a100-image dgx-os-6.0-a100-image-orig 3% commit
Wait for the ramdisk to be regenerated and the following text to be displayed.
1Wed Jul 26 09:01:11 2023 [notice] bcm10-headnode: Initial ramdisk for image dgx-a100-image-orig was generated successfully
Create the K8s software image by cloning the default software image. This software image will be further configured and provisioned onto the K8s control plane nodes. Wait for the ramdisk to be regenerated.
1% softwareimage 2% clone default-image k8s-master-image 3% commit
Add the required kernel modules to the k8s-master-image software image.
1% / 2% softwareimage 3% use k8s-master-image 4% kernelmodules 5% add mlx5_core 6% add bonding 7% softwareimage commit
Create the k8s-master node category and assign the k8s-master-image software image to it. All nodes assigned to the k8s-master category will be provisioned with the k8s-master-image software image.
1% category 2% clone default k8s-master 3% set softwareimage k8s-master-image 4% commit
Create the DGX nodes. node01 was created during head node installation. Clone node01 to create the DGX nodes, which will initially be named node02, node03, node04, and node05.
1% device 2% foreach --clone node01 -n node02..node05 () 3% commit
Rename the DGX nodes so they are more easily identified later.
1% use node02 2% set hostname dgx01 3% use node03 4% set hostname dgx02 5% use node04 6% set hostname dgx03 7% use node05 8% set hostname dgx04 9% device commit
Clone node01 to create the K8s control plane nodes, which will initially be named node05, node06 and node07.
1% device 2% foreach --clone node01 -n node06..node08 () 3% commit
Rename the K8s control plane nodes so they are more easily identifiable.
1% device 2% use node06 3% set hostname knode01 4% use node07 5% set hostname knode02 6% use node08 7% set hostname knode03 8% device commit
Rename node01. The purpose of this step is to specify that node01 is only a template.
1% device 2% use node01 3% set hostname template01 4% commit
Assign the DGX nodes to the correct node category; dgx-a100.
1% foreach -n dgx01..dgx04 (set category dgx-a100)
Assign the K8S nodes to the k8s-master node category.
1% foreach -n knode01..knode03 (set category k8s-master) 2% commit
Check the nodes and their categories. Extra options are used for device list to make the format more readable.
1% device list -f hostname:20,category:10,ip:20,status:15 2hostname (key) category ip status 3-------------------- ---------- -------------------- --------------- 4bcm10-headnode 10.227.48.8 [ UP ] 5dgx01 dgx-a100 10.227.48.5 [ DOWN ] 6dgx02 dgx-a100 10.227.48.6 [ DOWN ] 7dgx03 dgx-a100 10.227.48.7 [ DOWN ] 8dgx04 dgx-a100 10.227.48.4 [ DOWN ] 9knode01 k8s-master 10.227.48.4 [ DOWN ] 10knode02 k8s-master 10.227.48.4 [ DOWN ] 11knode03 k8s-master 10.227.48.4 [ DOWN ] 12template01 default 10.227.48.4 [ DOWN ]
Network Configuration#
Add a Network for InfiniBand (ibnet).
1% network 2% add ibnet 3% set domainname ibnet.cluster.local 4% set baseaddress 10.126.0.0 5% set netmaskbits 16 6% set mtu 2048 7% commit 8% add ibnet
Verify the results.
1% list -f name:20,type:10,netmaskbits:10,baseaddress:15,domainname:20 2name (key) type netmaskbit baseaddress domainname 3-------------------- ---------- ---------- --------------- -------------------- 4externalnet External 26 10.227.52.0 nvidia.com 5globalnet Global 0 0.0.0.0 cm.cluster 6ibnet Internal 16 10.126.0.0 ibnet.cluster.local 7internalnet Internal 26 10.227.48.0 eth.cluster 8ipminet Internal 26 10.227.20.64 ipmi.cluster
Ensure that head node interfaces are configured correctly.
1% device 2% use bcm10-headnode 3% interfaces 4% list 5Type Network device name IP Network Start if 6------------ -------------------- ---------------- ---------------- -------- 7bmc ipmi0 10.227.20.91 ipminet always 8physical ens10f0 10.227.52.8 externalnet always 9physical ens10f1 10.227.20.126 ipminet always 10physical ens1f1np1 [prov] 10.227.48.8 internalnet always
If any interfaces are missing or unconfigured, add any missing devices and configure their network as appropriate.
Reboot the head node if the network interfaces were changed.
1% / 2% device 3% use bcm10-headnode 4% reboot 5Reboot in progress for: bcm10-headnode
Configure Disk Layouts for Node Categories#
Part of using BCM for managing nodes in a DGX BasePOD is to define the disk partitions. Each DGX BasePOD node category includes K8s control plane and DGX node categories. The DGX categories are pre-configured with the correct disk partitions out of the box.
These steps detail how to configure the disk layout for the k8s-master category.
Augment the disksetup of the k8s-master category.
For the K8s control plane nodes, an EFI System Partition of 100 MB is created at the start of the disk, with the remainder of the disk dedicated to the OS as a single large partition. Note that this disk setup does not have a swap partition.
The configuration file references /dev/nvme0n1 as the block device used. This may need to be changed to match the specific device name used on systems intended as K8s control plane nodes.
Save the following text to /cm/local/apps/cmd/etc/htdocs/disk-setup/k8s-disksetup.xml, factoring in any necessary changes specific to the target systems as noted.
1<?xml version="1.0" encoding="ISO-8859-1"?> 2<diskSetup xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance"> 3<device> 4 5 <blockdev>/dev/nvme0n1</blockdev> 6 7 <partition id="a0" partitiontype="esp"> 8 <size>100M</size> 9 <type>linux</type> 10 <filesystem>fat</filesystem> 11 <mountPoint>/boot/efi</mountPoint> 12 <mountOptions>defaults,noatime,nodiratime</mountOptions> 13 </partition> 14 <partition id="a1"> 15 <size>max</size> 16 <type>linux</type> 17 <filesystem>xfs</filesystem> 18 <mountPoint>/</mountPoint> 19 <mountOptions>defaults,noatime,nodiratime</mountOptions> 20 </partition> 21</device> 22</diskSetup>
Assign this disk layout to the k8s-master node category.
1$ cmsh 2% category 3% use k8s-master 4% set disksetup /cm/local/apps/cmd/etc/htdocs/disk-setup/k8s-disksetup.xml 5% commit
Configure Node Network Interfaces#
Configure BCM to Allow MAC Addresses to PXE Boot#
Use the root (not cmsh) shell.
In /cm/local/apps/cmd/etc/cmd.conf, uncomment the AdvancedConfig parameter.
1AdvancedConfig = { "DeviceResolveAnyMAC=1" } # modified value
Restart the CMDaemon to enable reliable PXE booting from bonded interfaces.
1# systemctl restart cmd
Restarting the CMDaemon will disconnect the cmsh session. Type connect to reconnect after the CMDaemon has restarted. Or enter exit and then restart cmsh.
Configure Provisioning Interfaces on the DGX Nodes#
The steps that follow are performed on the head node and should be run on all DGX systems.
Warning
Double check the MAC address for each interface, and the IP address for the bond0 interface. Mistakes here will be difficult to diagnose.
Use a cmsh for loop to quickly add the new physical interfaces and the bond0 interface. This will update all four DGX A100 systems.
1# cmsh 2% device 3% foreach -n dgx01..dgx04 (interfaces; add physical enp225s0f1; add physical enp97s0f1; add physical enp225s0f1np1; add physical enp97s0f1np1; commit) 4% foreach -n dgx01..dgx04 (interfaces; add bond bond0; set interfaces enp225s0f1 enp97s0f1 enp225s0f1np1 enp97s0f1np1; set network internalnet; set mode 4; set options miimon=100; commit)
Set the physical interface MAC addresses as appropriate, and set the ipmi0 and bond0 interfaces if they should be changed—this must be repeated on each DGX system (a single system shown here).
1# cmsh 2% device 3% use dgx01 4% interfaces 5% set enp225s0f1 mac B8:CE:F6:2F:08:69 6% set enp97s0f1 mac B8:CE:F6:2D:0E:A7 7% set enp225s0f1np1 mac B8:CE:F6:2F:08:69 8% set enp97s0f1np1 mac B8:CE:F6:2D:0E:A7 9% set ipmi0 ip 10.227.20.69 10% set bond0 ip 10.227.48.13 11% commit 12% list 13Type Network device name IP Network Start if 14------------ -------------------- ---------------- ---------------- -------- 15bmc ipmi0 10.227.20.69 ipminet always 16physical BOOTIF [prov] 10.227.48.4 internalnet always 17bond bond0 10.227.48.13 internalnet always 18physical enp225s0f1 0.0.0.0 always 19physical enp225s0f1np1 0.0.0.0 always 20physical enp97s0f1 0.0.0.0 always 21physical enp97s0f1np1 0.0.0.0 always
Using a foreach loop, the bond0 interface as the provisioninginterface and remove bootif.
1% / # go to top level of cmsh 2% device 3% foreach -n dgx01..dgx04 (set provisioninginterface bond0; commit; interfaces; remove bootif; commit)
Verify the configuration.
1% device 2% use dgx01 3% get provisioninginterface 4bond0 5% interfaces 6% list 7Type Network device name IP Network Start if 8------------ ---------------------- ---------------- ---------------- -------- 9bmc ipmi0 10.227.20.69 ipminet always 10bond bond0 [prov] 10.227.48.13 internalnet always 11physical enp225s0f1 (bond0) 0.0.0.0 always 12physical enp225s0f1np1 (bond0) 0.0.0.0 always 13physical enp97s0f1 (bond0) 0.0.0.0 always 14physical enp97s0f1np1 (bond0) 0.0.0.0 always
Configure Provisioning Interfaces on the Kubernetes Nodes#
All the following steps in this section must be run for each of the three Kubernetes (K8s) nodes.
Use a cmsh for loop to quickly add the new physical interfaces and the bond0 interface. This will update all three knodes.
1% / # got to top level of CMSH 2% device 3% foreach -n knode01..knode03 (interfaces; add physical ens1f1; add physical ens2f1; add physical ens1f1np1; add physical ens2f1np1; commit) 4% foreach -n knode01..knode03 (interfaces; add bond bond0; set interfaces ens1f1np1 ens2f1np1 ens1f1 ens2f1; set network internalnet; set mode 4; set options miimon=100)
Set the physical interface MAC addresses as appropriate, and set the ipmi0 and bond0 interfaces if they should be changed—this must be repeated on each knode system (a single system shown here).
1% / 2% device 3% use knode01 4% interfaces 5% set ens1f1 mac 04:3F:72:E7:64:97 6% set ens1f1np1 mac 04:3F:72:E7:64:97 7% set ens2f1 mac 0C:42:A1:79:9B:15 8% set ens2f1np1 mac 0C:42:A1:79:9B:15 9% add bond bond0 10% set ipmi0 ip 10.227.20.80 11% set bond0 ip 10.227.48.30 12% list 13% commit 14 15Type Network device name IP Network Start if 16------------ -------------------- ---------------- ---------------- -------- 17bmc ipmi0 10.227.20.80 ipminet always 18physical BOOTIF [prov] 10.227.48.4 internalnet always 19bond bond0 10.227.48.30 internalnet always 20physical ens1f1 (bond0) 0.0.0.0 always 21physical ens1f1np1 (bond0) 0.0.0.0 always 22physical ens2f1 (bond0) 0.0.0.0 always 23physical ens2f1np1 (bond0) 0.0.0.0 always
Set the bond0 interface as the provisioninginterface, and remove bootif. A for loop should be used here again.
1% / 2% device 3% foreach -n knode01..knode03 (set provisioninginterface bond0; commit; interfaces; remove bootif; commit)
Configure InfiniBand Interfaces on DGX Nodes#
The following procedure adds four physical InfiniBand interfaces, and must be run for each DGX node.
Use a cmsh for loop to quickly add the new physical Infiniband interfaces. This will update all four DGX nodes.
1% / # got to top level of CMSH 2% device 3% foreach -n dgx01..dgx04 (interfaces; add physical ibp12s0; set network ibnet; add physical ibp141s0; set network ibnet; add physical ibp186s0; set network ibnet; add physical ibp75s0; set network ibnet; commit)
Set the ip addresses for each physical Infiniband interface—this will need to be repeated on each DGX system (a single system shown here).
1% / # go to top level of CMSH 2% device 3% use dgx01 4% interfaces 5% set ibp12s0 ip 10.126.0.13 6% set ibp141s0 ip 10.126.2.13 7% set ibp186s0 ip 10.126.3.13 8% set ibp75s0 ip 10.126.1.13 9% commit 10% list 11Type Network device name IP Network Start if 12------------ ---------------------- ---------------- ---------------- -------- 13bmc ipmi0 10.227.20.69 ipminet always 14bond bond0 [prov] 10.227.48.13 internalnet always 15physical enp225s0f1 (bond0) 0.0.0.0 always 16physical enp225s0f1np1 (bond0) 0.0.0.0 always 17physical enp97s0f1 (bond0) 0.0.0.0 always 18physical enp97s0f1np1 (bond0) 0.0.0.0 always 19physical ibp12s0 10.126.0.13 ibnet always 20physical ibp141s0 10.126.2.13 ibnet always 21physical ibp186s0 10.126.3.13 ibnet always 22physical ibp75s0 10.126.1.13 ibnet always
Identify the Cluster Nodes#
Identify the nodes by setting the MAC address for the provisioning interface for each node to the MAC address listed in the site survey.
1% device 2% set dgx01 mac b8:ce:f6:2f:08:69 3% set dgx02 mac 0c:42:a1:54:32:a7 4% set dgx03 mac 0c:42:a1:0a:7a:51 5% set dgx04 mac 1c:34:da:29:17:6e 6% set knode01 mac 04:3F:72:E7:64:97 7% set knode02 mac 04:3F:72:D3:FC:EB 8% set knode03 mac 04:3F:72:D3:FC:DB 9% foreach -c dgx-a100,k8s-master (get mac) 10B8:CE:F6:2F:08:69 110C:42:A1:54:32:A7 120C:42:A1:0A:7A:51 131C:34:DA:29:17:6E 1404:3F:72:E7:64:97 1504:3F:72:D3:FC:EB 1604:3F:72:D3:FC:DB
If all the MAC addresses are set properly, commit the changes.
1% device commit 2% quit 3% commit
Next Steps#
After you complete the steps on this page, see Power On and Provision Cluster Nodes.