This page describes all the steps to build a CentOS 5 'small' i386 cluster. Instructions are suitable also for installation of a multi-CPUs workstation, which resembles configuration on the server. You can use it as a tutorial, real configuration files and instructions are provided. Instructions can be easily modified for x86_64 platform.

Note: in the case of workstation installation skip server specific steps when requested, and all steps related to the "Golden Client".

The cluster consists of:

  • a server with two network interfaces: one will serve the nodes' private network, the other one will connect the server to the external network, the i386 server needs to be equipped with a harddisk with at least 40GB of disk space,

    Note that the configuration files are provided in a case of hard disk recognized as /dev/sda device.

  • at least 2 identical PC's - compute nodes - one on these will serve as a "Golden Client" (see installing and configuring systemimager), the others will serve as compute nodes (you need at least two single-cpu compute nodes to take advantage of parallel applications). The configuration described here requires i386 nodes with at least 14GB of disk space.

    Note the that configuration files are provided in a case of hard disk recognized as /dev/hda device.

  • an ethernet switch,

  • cables, screen, keyboard, mouse,

  • USB drive (optional: for kickstart installation of dulak-server), with "Write Protect" option.

This is only one of possible ways to organize the cluster, the others include:

  • having the cluster completely disconnected from the external network - the advantage of this solution is its security (no cracker's attacks possible), but the disadvantage that all the administration (software updates) must be performed on site.
  • having not only the server, but also the compute nodes connected to the public network; the disadvantage is its lack of security and the use of public bandwith; the advantage is its simpilicity of configuration (e.g. no need for DNS) and performing software updates.

The approach presented here is between these two described above: server is connected to the public network, so administration and software updates on the server (see http://yum.baseurl.org/) can be performed over the network, but it means also that the server must be kept secure. The software updates on the compute nodes must be still performed manually, unless you make the server to act as a local yum repository.

The described cluster/workstation installation assumes that:

  • you have an internet network connection,
  • root uses 0022 umask,
  • the compute nodes will reside on 10.3.0.0/24 private network,
  • domain is called dulak-cluster.fysik.dtu.dk (irrelevant in case of workstation installation),
  • server is called dulak-server and has an IP address of 10.3.0.2 (irrelevant in case of workstation installation),
  • Note: dulak-server serves the compute nodes on eth1 interface, and is connected to the public network using eth2 interface,
  • compute nodes are called "n001.dulak-cluster.fysik.dtu.dk" to "n100.dulak-cluster.fysik.dtu.dk",
  • "n001" serves as "Golden Client".

Please start with (please follow the order - it matters!) installing server.

Note: work still in progress!

Niflheim: Building_a_Cluster_-_Tutorial (last edited 2010-10-20 09:15:27 by localhost)