are based on Debian and the Debian Diskless Cluster Howto by Gabe Ortiz. The present scripts automate most of the steps described in that Howto. The difference to the just mentioned setup is, that it mounts the NFS root filesystem as read-only and puts Aufs (an overlay-filesystem) on top of it.

I tested the resulting system in a 50 node cluster, where no problems showed up. The only fundamental disadvantage of this solution is, that the system image for worker nodes cannot be updated online. This is, because "Aufs" does not expect that the underlying read-only file system is updated. That means, that you have to shutdown all nodes in the cluster, do the update and to boot them up again (there is no free lunch, sorry).

WARNING: Currently, the project is not that mature, i.e. it fulfills primarily my personal needs, but I would love to see it becoming more general.

If you are interested in a more mature project, then have a look at FAI.

The following sections describe how to setup a diskless Debian cluster system.

• packages assumed to be installed

  • debootstrap
  • tftpd-hpa
  • syslinux
  • nfs-kernel-server
  • nfs-common
  • rsyslog

• services which are assumed to be already present

  • DHCP server
  • TFTP server (tftpd-hpa)
  • NFS server (nfs-kernel-server nfs-common)
  • NIS server

• the nodes must have PXE (Preboot Execution Environment) enabled and the boot order must be changed to PXE first

The setup consists of two parts, the setup of a master node and the setup to bootstrap an image used as a base for the worker nodes.

The instructions assume that you already have a working cluster installation.

1. change in the base directory where the scripts reside
2. configure "config.sh"
3. run "prepare_master.sh"
4. run "prepare_nfsroot.sh"

For those who do not already have DHCP entries for each worker node. The file to adapt is /etc/dhcp/dhcpd.conf. An entry for each worker node (IP, Mac, hostname) looks like follows:

subnet @NETWORK@ netmask @NETMASK@ {
    filename "pxelinux.0";
    next-server @TFTP_SERVER_IP@;
    option subnet-mask @NETMASK@;
    option broadcast-address @BROADCAST@;
    option routers @ROUTER@;
}

host @WORKER_HOSTNAME@ {
    hardware ethernet @WORKER_MAC_ADDRESS@;
    fixed-address @WORKER_IP_ADDRESS@;
    option host-name "@WORKER_HOSTNAME@";
}

• diskless means here that the base system does not require any local disk space
• designed by following the KISS principle
• tries to be as non-intrusive as possible, i.e. the goal is to integrate well in already existing cluster setups (e.g. it is assumed that an DHCP-server already exists)
• local disks (if available) can completely be used by user applications (e.g. to place data on the disk such you get maximum read/write performance)
• heavily changing directories (like /home) should reside in a separate (writable) medium (like NFS or local discs), otherwise RAM is filled up
• the common shared base system ensures that all cluster nodes have exactly the same configuration (except e.g. the hostname and IP-address, of course)
• the base system image cannot be updated online, "Aufs" does not expect that the underlying read-only filesystem changes (if you try that out, the system may continue to work, but it will definitely crash later on)
• the master is a single point of failure
• works only with Debian Squeeze right now

This project definitely does need more work and help. Patches and ideas are very welcome.