Installing A Cluster

Network configuration

Head nodes works as NAT router for slave nodes

eth0 connects to outside world eth1 is internal

Internal network is 10.200.0.0/255.255.0.0

OS installation

Normal server Linux install with the following packages:

• openssh-server
• tftpd-hpa
• dhcp3-server
• nfs-kernel-server
• debootstrap
• libpmi
• mpich2
• slurm-llnl
• slurm-llnl-slurmdbd
• syslinux

Setting up a chroot for the node root

• Boot to a live CD and mount the root drive as /new
• Copy the root fs to a chroot rsync -av /new/ /new/nodes/precise/ --exclude /nodes
• Enter the chroot chroot /nodes/precise/
• Change the following lines in /etc/initramfs-tools/initramfs.conf

 MODULES=netboot
 BOOT=nfs

The DEVICE= and NFSROOT= lines may also be of use if the tftpboot configuration isn't working

• Update the initramfs update-initramfs -c -k all

NFS

On the head node, add the following to /etc/exports

/nodes/lucid    10.200.0.0/24(ro,async,no_root_squash,no_subtree_check)
/home           10.200.0.0/24(rw,async,root_squash,no_subtree_check)

see man 5 exports for more information. There are also more nfs settings in

/etc/default/nfs-common 
/etc/default/nfs-kernel-server

Netbooting

Setting up dhcp

Edit /etc/dhcp3/dhcpd.conf as follows:

 ddns-update-style none;
 option domain-name "iac.isu.edu";
 option domain-name-servers 134.50.254.5, 134.50.57.57;
 authoritative;

 default-lease-time 600;
 max-lease-time 1200;

 subnet 10.0.200.0 netmask 255.255.255.0  
 {
  range 10.0.34.1 10.0.34.200; #we really don't even need a range
  option routers 10.0.200.1;
  #address of the TFTP server, optional if the same as the dhcp server
  next-server 10.0.200.1;
  filename "pxelinux.0";
 }
 #eth0 on brems2
 host brems2_0
 {
  hardware ethernet 00:50:45:5C:10:54;
  fixed-address 10.0.200.2;
  option host-name "brems2";
 }

Edit /etc/default/dhcp3-server

INTERFACES=eth1

This will avoid dhcp serving on the outside network!

service dhcp3-server restart

Setting up tftp

Edit /etc/default/tftpd-hpa

TFTP_USERNAME="tftp"
TFTP_DIRECTORY="/var/lib/tftpboot"
#Keep TFTP on the inside network
TFTP_ADDRESS="10.0.200.1:69"
TFTP_OPTIONS="--secure"

Set up the filesystem to over pxe/tftpd:

mkdir -p /var/lib/tftpboot/pxelinux.cfg
cp /usr/lib/syslinux/pxelinux.0 /var/lib/tftpboot/

Make a file similiar to the following as /var/lib/tftpboot/boot.msg

Booting Brems!!

Put the following in /var/lib/tftpboot/pxelinux.cfg/default. The files are referenced relative to /var/lib/tftpboot/

TIMEOUT 5
DISPLAY boot.msg
DEFAULT vmlinuz
APPEND root=/dev/nfs initrd=initrd.img netboot=nfs nfsroot=10.0.200.1:/nodes/precise,nolock,ro nomodeset

Set up a kernel and initrd in tftpboot

cp /nodes/precise/boot/vmlinuz-3.2.0-24-generic /var/lib/tftpboot/vmlinuz
chmod a+r /var/lib/tftpboot/vmlinuz
cp /nodes/precise/boot/initrd.img-3.2.0-24-generic /var/lib/tftpboot/initrd.img

Scheduler installation

The Quick Start Administrator Guide is very helpful.

1. Install

• slurm-llnl
• slurm-llnl-slurmdbd
• slurm-llnl-doc
  • mkdir /var/run/slurm-llnl

Munge

Munge is an authentication framework recommended by slurm. All the configuration it needs is:

root@brems:# /usr/sbin/create-munge-key
Generating a pseudo-random key using /dev/urandom completed.
root@brems:# /etc/init.d/munge start

Adding a node to the cluster

1. Set node to PXE boot
2. Add a new entry to /etc/dhcp3/dhcpd.conf
3. Reload dhcpd
4. Add new host to /etc/hosts file
5. Copy /etc/hosts to /nodes/lucid/etc/hosts
6. Create a new var cp -r /nodes/lucidvar/template /nodes/lucidvar/newhost
7. Boot the node
8. Add the ssh key to system-wide known-hosts ssh-keyscan newhost >> /etc/ssh/ssh_known_hosts
9. Add the host to cluster ssh
10. Add the node to slurm config

Customising initrd

Sometimes a customised initrd is necessary.

Extracting

Create a directory to extract into:

mkdir init_test; cd init_test

Extract an existing initrd

gzip -d < /var/lib/tftpboot/initrd.img | cpio -iv

Edit the file tree as needed. You can add files (modules for instance), or editing boot script (/init is run by default).

Then package up the directory into a new initrd:

find ./ | cpio -ov -H newc | gzip > ../initrd.new