Building a Two-Node Linux Cluster with Heartbeat

The term "cluster" is actually not very well defined and could mean different things to different people. According to Webopedia, cluster refers to a group of disk sectors. Most Windows users are probably familiar with lost clusters--something that can be rectified by running the defrag utility.

However, at a more advanced level in the computer industry, cluster usually refers to a group of computers connected together so that more computer power, e.g., more MIPS (millions instruction per second), can be achieved or higher availability (HA) can be obtained.

Most super computers in the world are built on the concept of parallel processing--high-speed computer power is achieved by pulling the power from each individual computer. Made by IBM, "Deep Blue", the super computer that played chess with the world champion Garry Kasprov, was a computer cluster that consisted of several hundreds of RS6000s. In fact, many big time Hollywood movie animation companies, such as Pixar, Industrial Light and Magic, use computer clusters extensively for rendering (a process to translate all the information such as color, movement, physical properties, etc., into a single frame of picture).

In the past, a super computer was an expensive deluxe item that only few universities or research centers could afford. Started at NASA, Beowulf is a project of building clusters with "off-the-shelf" hardware (e.g., Pentium PCs) running Linux at a very low cost.

In the last several years, many universities world-wide have set up Beowulf clusters for the purpose of scientific research or simply for exploration of the frontier of super computer building.

Clusters in this category use various technologies to gain an extra level of reliability for a service. Companies such as Red Hat, TurboLinux and PolyServe have cluster products that would allow a group of computers to monitor each other; when a master server (e.g., a web server) goes down, a secondary server will take over the services, similar to "disk mirroring" among servers.

Because I do not have access to more than one real (or public) IP address, I set up my two-node cluster in a private network environment with some Linux servers and some Win9x workstations.

If you have access to three or more real/public IP addresses, you can certainly set up the Linux cluster with real IP addresses.

In the above network diagram (fig1.gif), the Linux router is the gateway to the Internet, and it consists of two IP addresses. The real IP, 24.32.114.35, is attached to a network card (eth1) in the Linux router and should be connected to either an ADSL modem or a cable modem for internet access.

The two-node Linux router consists of node1 (192.168.1.2) and node2 (192.168.1.3). Depending on your setup, either node1 or node2 can be your primary server, and the other will be your backup server. In this example, I will choose node1 as my primary and node2 as my backup. Once the cluster is set, with IP aliasing (read IP aliasing from the Linux Mini HOWTO for more detail), the primary server will be running with an extra IP address (192.168.1.4). As long as the primary server is up and running, services (e.g., DHCP, DNS, HTTP, FTP, etc.) on node1 can be accessed by either 192.168.1.2 or 192.168.1.4. In fact, IP aliasing is the key concept for setting up this two-node Linux cluster.

When node1 (the primary server) goes down, node2 will be take over all services from node1 by starting the same IP alias (192.168.1.4) and all subsequent services. In fact, some services can co-exist between node1 and node2 (e.g., FTP, HTTP, Samba, etc.), however, a service such as DCHP can have only one single running copy on the same physical segment. Likewise, we can never have two identical IP addresses running on two different nodes in the same network.

In fact, the underlining principle of a two-node, high-availability cluster is quite simple, and people with some basic shell programming techniques could probably write a shell script to build the cluster. We can set up an infinite loop within which the backup server (node2) simply keeps pinging the primary server, if the result is unsuccessful, and then start the floating IP (192.168.1.4) as well as the necessary dæmons (programs running at the background).

You need two Pentium class PCs with a minimum specification of a 100MHz CPU, 32MB RAM, one NIC (network interface card), 1G hard drive. The two PCs need not be identical. In my experiment, I used an AMD K6 350M Hz and a Pentium 200 MMX. I chose the AMD as my primary server as it can complete a reboot (you need to do a few reboots for testing) faster than the Pentium 200. With the great support of CFSL (Computers for Schools and Libraries) in Winnipeg, I got some 4GB SCSI hard drives as well as some Adaptec 2940 PCI SCSI controllers. The old and almost obsolete equipment is in good working condition and is perfect for this experiment.

node1

node2