Skip to content

Latest commit

 

History

History
117 lines (86 loc) · 5.75 KB

beowulf-cluster.md

File metadata and controls

117 lines (86 loc) · 5.75 KB
Raw

STEPS TO CREATE A BEOWULF CLUSTER -

(My Host PC OS - Windows. There shouldn't be much difference in other OS, that's what I hope.)

  1. Install the latest version of Oracle VirtualBox

  2. Set up and install a stable Linux version (I used Ubuntu 16.04.4) in separate virtual machines (preferred 3-4)

  3. ENABLE BRIDGED NETWORKING BETWEEN THE NODES

    After installation, power off the virtual machines and for each node, do -

    Settings -> Network -> Check the "Enable Network Adaptor" option -> Select "Bridged Adaptor" and select any one of the options that works-> OK

  4. Make sure all the nodes have the same username. If they have different usernames, create a new user in all nodes with a common name and give sudo permissions to the new users. Install ssh-server on the master node and all the slave nodes by using sudo apt-get install openssh-server

  5. STATIC IP ALLOCATION

    Assign a static IP address to node by -

    • Run the following commands in terminal -
      • ifconfig -> gives the network name, address, netmask, network address(X.X.0.0) and broadcast address
      • route -m -> gives the gateway address
      • cat /etc/resolv.conf -> gives the dns-nameservers address
    • cd /etc/network and then open up interfaces using sudo gedit interfaces
    • Add the following lines -
    auto {insert-network-name}
iface {insert-network-name} inet static
address {insert-ip-address}
netmask {insert-network-mask}
network {insert-network-address}
broadcast {insert-broadcast-address}
gateway {insert-gateway-address}
dns-nameservers {insert-dns-nameservers-address}
    • Save the file and restart your network

  6. From home directory, open up hosts file in each node using -

cd /etc
sudo gedit hosts

Add the following info in all the nodes -

{host1 IP address} {master-node-name} 
{host2 IP address} {slave1-node-name} 
{host3 IP address} {slave2-node-name} 
.
.
.
  1. CREATE PASSWORDLESS SSH FOR LOG IN FROM master node TO slave nodes

METHOD 1

  • On the all the node, type ssh-keygen on the terminal. Just hit enter for all other options. This option will create a id_rsa.pub RSA public key for your node.
  • On the master node, type ssh-copy-id -i ~/.ssh/id_rsa.pub <first-node-name> on the terminal. It will prompt you for the password of the first-node, enter that. Repeat this step for all nodes. This will copy your RSA public key of your master node to the id_rsa.pub of your other nodes, enabling a passwordless SSH from master to all the slave node.
  • To check if this works, try ssh <node-name>. Try out for all nodes.
  • Note - Passwordless log in from slave to master is not possible with these steps and we won‘t be requiring that.

METHOD 2

  • On the master node, run the command ssh-keygen -t rsa

  • On the master node, do the following for slave node 1 -

    ssh <username>@<slave-node1-name> mkdir -p .ssh
cat .ssh/id_rsa.pub | ssh <username>@<slave-node1-name> 'cat >> .ssh/authorized_keys'
ssh <username>@<slave-node1-name> "chmod 700 .ssh; chmod 640 .ssh/authorized_keys"

  • Run ssh <slave-node1-name> and see if you can ssh without any password

  • Repeat the steps for all other slave nodes, excluding step 1 of METHOD 2. The ssh-keygen for master node needs to be done only once.

  1. SETTING UP NFS SERVER AND CLIENTS

  • NFS - Network File System. Used to view, store and update files between local and remote systems. Here, we’ll use it to provide a shared memory to the master and slave nodes.

  • On the master node, run the command sudo apt-get install nfs-kernel-server and install the NFS server. Create the folder that will be shared among the master and slave nodes using mkdir <foldername> . Move to etc folders and open up the exports file using cd /etc and sudo gedit exports . Add the following at the end -

    /home/<username>/<foldername> <slave-node1-name>(rw,sync,no_root_squash,no_subtree_check)
/home/<username>/<foldername> <slave-node2-name>(rw,sync,no_root_squash,no_subtree_check)
.
.
.

  • Go back to home directory, run exportfs -a in the terminal followed by sudo service nfs-kernel-server restart

  • On the slave nodes, run the command sudo apt-get install nfs-common. Create a file with the same name as the one on the master node <foldername>. Run sudo mount -t nfs <master-node-name>:/home/<username>/<foldername> ~/<foldername>. To check if it’s mounted or not, run df -h. To permanently mount it, go to /etc folder and open up fstab file using cd /etc and sudo gedit fstab. Add the following at the end -

    <master-node-name>:/home/<username>/<foldername> /home/<username>/<foldername>
  1. INSTALLING MPI ON ALL THE NODES
  • **NOTE - **MAKE SURE YOU INSTALL THE SAME VERSION OF MPI ON ALL NODES
  • Download the latest stable version of MPI tar file from here.
  • Extract the tar using tar -xzf <mpich-tar-filename>
  • Go to the extracted file
  • Run ./configure --disable-fortran
  • Run the command make
  • Run the command sudo make install
  • Check if build was successful using mpiexec --version
  1. RUNNING A SAMPLE MPI PROGRAM IN THE CLUSTER

  • We’ll just test our cluster and hence use an existing code that has already been compiled into an executable file in the examples file of MPI folder.

  • **On the master node - **Go to the MPI folder on home -> examples -> copy the cpi executable and paste it in the shared folder.

  • Navigate to the shared folder through the terminal and run mpi -np <> -hosts <slave1-node-name>,<slave2-node-name> ./cpi

And that’s it. We’ve formed a Beowulf cluster and allocated different tasks to each slave node from the master node via a shared memory. :)