GPU Cluster Configuration Notes

Introduction

This document contains notes on configuring a cluster of machines with NVIDIA GPUs running Ubuntu Linux 14.04 or later on a private network connected to a single master host that serves as the cluster's network gateway, file server, and name service master. SLURM is used for job management, OpenLDAP is used for name service management, and the existence of an externally managed Kerberos KDC is assumed for managing user authentication.

The sections of this document are not necessarily listed in a prescribed order, nor does the document attempt to provide all information necessary for obtaining an optimal cluster configuration. Feel free to submit suggestions/corrections as pull requests to the source repository.

The author categorically disclaims all responsibility for any adverse effects to your data center that may ensue as a result of following these instructions. :-)

Author & License

This work by Lev Givon is licensed under a Creative Commons Attribution 4.0 International License

General System Configuration

• After installing Ubuntu, it's possible that the system's console might not work because of misinteraction with the nouveau open source NVIDIA driver. To fix this, login to the machine over the network with ssh and blacklist the driver by adding a file to /etc/modprobe.d/ containing the line

  blacklist nouveau
  

  Recent NVIDIA CUDA packages should automatically do the above during installation, however.

• Add umask 0077 to /etc/bash.bashrc before creating any user accounts to enforce more private default file creation permissions.

• When creating user accounts, check that the created home directory and the various files created by default (e.g., .bashrc, .profile, etc.) are not world readable.

• If the master host contains an IPMI or BMC device for remote management exposed to the Internet, have your network administrator assign it a static IP address and remember to set an administrator password. The latter can typically be done through the web or via ipmitool.

• The IPMI devices of the remote management interfaces on the internal network do not need any passwords (the default username and password - ADMIN - can remain unchanged).

• To upgrade Ubuntu from the command line, install update-manager-core, edit /etc/update-manager/release-upgrades, and run do-release-upgrade as root.

Configuring Networking

• The below instructions assume that the worker nodes have private addresses in the 192.168.0.0/16 subnet.

• Activate ufw on the master host and deactivate it on the worker hosts.

• Leave the OpenSSH port on the master host open.

• Update /etc/default/ufw to contain the line

  DEFAULT_FORWARD_POLICY="ACCEPT"
  

• Update /etc/ufw/sysctl.conf to contain the lines

  net/ipv4/ip_forward=1
  net/ipv6/conf/default/forwarding=1
  net/ipv6/conf/all/forwarding=1
  

• Add the following lines to the top of /etc/ufw/before.rules (replace the multicast address as appropriate for the private network and the interface with whichever interface the gateway uses to communicate with the outside world):

  * nat
      :POSTROUTING ACCEPT [0:0]
      -A POSTROUTING -s 192.168.0.0/8 -o eth0 -j MASQUERADE
      COMMIT
  

• Add the following rules:

  ufw allow to 192.168.0.0/16
  ufw allow from 192.168.0.0/16
  

• After making the above modifications, restart ufw:

  ufw disable && ufw enable
  

• Install avahi-daemon on the master and configure avahi on all of the nodes (including the master) to assign a private hostname. This should only involve modifying the host-name and domain-name options in /etc/avahi/avahi-daemon.conf

• On the master, make sure that avahi only announces the private hostname on the internal Ethernet interface associated with the private network by setting the allow-interfaces option in /etc/avahi/avahi-daemon.conf accordingly.

• Put the hostname of each worker in its respective /etc/sysconfig/network file, e.g.,

  HOSTNAME=node02.local
  

• Add all of the worker host names and IP addresses to /etc/hosts on the master, e.g.:

  192.168.0.1 node01.local    node01
  192.168.0.2 node02.local    node02
  192.168.0.3 node03.local    node03
  192.168.0.4 node04.local    node04
  192.168.0.5 node05.local    node05
  

• Install isc-dhcp-server on the master and configure it to assign static private IP addresses to the workers; see the accompanying dhcpd.conf file for an example.

• If the machines have IPMI devices on the same physical Ethernet ports that are connected to the private network, make sure that they are assigned their own IP addresses via DHCP. It may be necessary to manually clear the IP address associated with the IPMI device in the machine's BIOS.

• Ostensibly, it is possible to use ipmitool to set the IPMI device LAN Select setting on SuperMicro motherboards (see this page for more information).

• To configure password-less login from any machine in the cluster to the other for all non-root users, make sure that /etc/ssh/ssh_config on all of the machines contains the following lines:

  HostbasedAuthentication yes
  EnableSSHKeysign yes
  

  To reduce latency, it is advisable to include the following lines:

  Compression no
  Ciphers blowfish-cbc
  

• /etc/ssh/shots.equiv on all of the nodes should contain the private names of each of the nodes.

• /etc/ssh/ssh_known_hosts needs to contain the public host key for each host that one wishes to connect to; the host name and IP address need to be included as well.

• To enable password-less login for root on the private nodes,

  • create a /root/.shosts file that contains the private names of all of the machines in the cluster and make sure that /etc/ssh/sshd_config on each node contains the following option:

    IgnoreRhosts no
    

  • create public keys for the root user with no passphrase and dump the public keys into /root/.ssh/authorized_keys on each host

  • set PermitRootLogin without-password in /etc/ssh/sshd_config on all of the hosts

Setting up NFS

• Install nfs-server on the master and nfs-client on the worker hosts.

• To export the home directories on the master node, make sure that the line

  NEED_IDMAPD=yes
  

  is in /etc/default/nfs-common on both the master and client hosts.

• On the master, create a directory called /srv/nfs4/home on the master node, set its permissions to 755, and mount /home on it using the command

  mount --bind /home /srv/nfs4/home
  

  Modify the master's /etc/fstab file to contain

  /home /srv/nfs4/home none bind 0 0
  

• Modify /etc/exports on the master to contain

  /srv/nfs4      192.168.0.0/24(rw,fsid=0,nohide,no_subtree_check,no_root_squash)
  /srv/nfs4/home 192.168.0.0/24(rw,nohide,no_subtree_check,no_root_squash)
  

• Run exportfs -a on the master to export /srv/nfs4/home to the clients. Run showmount -e 192.168.0.1 on the clients to confirm that they can see the master's export list.

• Create the directory /mnt/server-home on the clients and modify their /etc/fstab files to contain

  192.168.0.1:/home /mnt/server-home nfs4 auto,_netdev,hard,intr 0 0
  

• Move /home to /local-home on all of the clients and create a link from /home to /mnt/server-home; mount /mnt/server-home on all of the clients.

• It may be possible to improve NFS performance by adjusting network interface settings and mount parameters. See this page for more information

Setting up LDAP

• Install openldap-servers and openldap-clients on the master.

• Use dpkg-reconfigure to reconfigure LDAP on Ubuntu. The default domain and base don't need to be changed.

• Make sure that /etc/nsswitch.conf is configured to look at ldap after files when looking up password, shadow, or group data:

  passwd:         files ldap [NOTFOUND=return] db
  group:          files ldap [NOTFOUND=return] db
  shadow:         files ldap [NOTFOUND=return] db
  

• If there is a need to reinstall the OS, the contents of the LDAP database can be dumped into an ldif format file using slapcat and loaded into the new server's database using something like

  ldapadd -v -x -W -D "cn=admin,o=nodomain" -c -f old.ldif
  

  where the domain is whatever is associated with the LDAP administrator.

Installing libuser

• libuser provides command-line tools for managing user accounts. Since the stock Ubuntu package isn't compiled with LDAP support, however, it needs to be manually built and installed as follows.

• Install libsasl-dev, libpython2.7-dev, libldap-dev, libpopt-dev, and libpam-dev. Make sure that the stock libuser1 package is not installed.

• Download the latest libuser source, unpack, and build as follows:

  ./configure --prefix=/usr/local --with-ldap=/usr/include \
  --with-popt=/usr/include --with-sasl=/usr/include
  make CFLAGS=-I/usr/include
  make install
  

• Update /usr/local/etc/libuser.conf to set the lines in the associated sections (replace the basedn, binddn, and password values as needed); also ensure that it is only readable by root.

  [defaults]
  modules = ldap
  create modules = ldap
  
  [ldap]
  server = ldap://127.0.0.1
  basedn = dc=nodomain
  
  binddn = cn=admin,dc=nodomain
  password = mypassword
  bindtype = simple
  

• Try adding a user using /usr/local/sbin/luseradd as root. If everything works properly, the new user should appear in the output of slapcat.

• Remember to add the Unix account used to administer the master machine to LDAP with luseradd - specify the existing uid, group, and home directory so that new ones are not created.

Setting up Kerberos Authentication

• Install the krb5-workstation package on the master server and configure /etc/krb5.conf to refer to the appropriate KDC. The accompanying krb5.conf file is specific to Columbia University.
• Install pam-krb5. Note that this is the module used by Debian, not by RedHat.
• After installing pam-krb5, it may be necessary to adjust the minimum_uid parameter in the pam configuration files.
• Add .k5login files to the users' directories containing the appropriate principal. For Columbia University, this should be [email protected] (where abc123 is the CUIT-assigned UNI of the user in question) to enable users to access the machine using the Kerb password associated with their UNI.
• Add users authorized to access the machine to the AllowUsers line in /etc/ssh/sshd_config.
• To store the password of an account locally in /etc/shadow (e.g., to ensure that the user can login even if Kerberos or LDAP are not functioning),
  • temporarily disable Kerberos and LDAP authentication using pam-auth-update,
  • create a temporary local password using mkpasswd -m sha-512 -S somesaltstring -s <<< TempPassword
  • add a line for the account to /etc/passwd with vipw and a line containing the encrypted password to /etc/shadow with vipw -s,
  • modify the password to whatever the user wants using /usr/bin/passwd,
  • update the account's local groups if so desired by editing /etc/group using vigr and vigr -s, and
  • re-enable Kerberos and LDAP authentication using pam-auth-update.

Installing CUDA

• Ubuntu provides its own NVIDIA GPU driver and CUDA packages. Although you can use them, the ones provided by NVIDIA are usually more up to date; read on if you want to use them.

• For versions of Ubuntu for which a .deb package is available:

  • Download and install the "deb (network)" Ubuntu package from NVIDIA's website.
  • After refreshing the system's package information using apt-get update, install the cuda-VERSION metapackage (e.g., cuda-7-5) to install all of the requisite drivers and libraries. Reboot the machine after installation.

• For more recent versions of Ubuntu for which no .deb package is available (e.g., Ubuntu 16.04 as of April 2016):

  • Ensure that the most recent NVIDIA kernel drivers are installed; you can find them by installing aptitude and running the command aptitude search nvidia
  • Download and install the "runfile (local)" file from NVIDIA's website for the most recent release of Ubuntu.
  • Make the file executable and run it with the --override option.
  • When prompted by the installer as to whether to install the "Accelerated Graphics Driver", enter n.
  • Install the CUDA software in /usr/local/cuda-VERSION with a link from /usr/local/cuda to that directory, where VERSION is the version of CUDA being installed.
  • After installation is complete, ensure that that all of the contents of the /usr/local/cuda-VERSION directory are world-readable (and executable where appropriate).
  • Create a file named /etc/profile.d/cuda.sh containing the line export PATH=$PATH:/usr/local/cuda/bin
  • Create a file named /etc/ld.so.conf.d/cuda.conf containing the line /usr/local/cuda/lib64
  • Run the command sudo source /etc/profile.d/cuda.sh
  • Run the command sudo ldconfig

• If the /dev/nvidia* devices fail to initialize when the machine boots and there appears to be a kernel module error in the output of dmesg, try installing a more recent version of the device drivers (you may need to obtain it from a third party ppa).

• Ensure that nvidia-persistenced has been installed and is running - this will keep GPUs warm so as to avoid delays in startup. On Ubuntu 16.04, it may be necessary to create a startup script manually; see the init subdirectory in this repo for details.

• Add /usr/local/cuda/bin to PATH in /etc/bash.bashrc so that all users can access the CUDA binaries without having to modify their own .bashrc scripts.

• On Ubuntu 16.04, comment out the line that contains the following text in the file /usr/local/cuda-7.5/include/host_config.h:

  #error -- unsupported GNU version! gcc versions later than ... not
  supported!
  

  using a C++ line comment symbol (//) so that CUDA works properly with gcc 5.

Configuring SLURM

• Install slurm-llnl and munge on all hosts.

• Generate a MUNGE key on the master by running create-munge-key.

• Modify various directory/file permissions as indicated in the MUNGE Wiki.

• On Ubuntu 14.04, update /etc/default/munge to circumvent this bug.

• For Ubuntu 15.04 or later, see this issue.

• Copy the MUNGE key on the master to /etc/munge on the worker hosts.

• Start MUNGE using service munge start

• Install the accompanying slurm.conf and gres.conf files to /etc/slurm-llnl; modify both files as appropriate. To find the number of CPUs (or hyperthreads, if supported), sockets, cores per socket, and threads per core, run the lscpu utility; to find the GPU device files to list in gres.conf, run ls -l /dev/nvidia?.

• Note that slurm.conf must be the same on all nodes, but gres.conf should be customized in accordance with the actual number of GPUs on a host.

• On Ubuntu 16.04, it may be necessary to include the following lines in slurm.conf:

  SelectType=select/cons_res
  SelectTypeParameters=CR_CPU_Memory
  

• Run update-rc.d slurm-llnl enable to ensure that SLURM starts on reboot. On Ubuntu 14.04, it may be necessary to restart SLURM manually after a reboot if GPU initialization does not complete before the system tries to start SLURM.

• To prevent users on the master node from accessing any GPUs on that machine without using SLURM, include the following in /etc/bash.bashrc

  export CUDA_VISIBLE_DEVICES=