Computer clustering involves the use of multiple computers, typically personal computers (PCs) or UNIX workstations, multiple storage devices, and redundant interconnections, to form what appears to users as a single integrated system (Cluster computing). Clustering has been available since the 1980s when it was used in Digital Equipment Corp's VMS systems. Today, virtually all leading hardware and software companies including Microsoft, Sun Microsystems, Hewlett Packard and IBM offer clustering technology. This paper describes why and how clustering is commonly used for parallel processing, batch processing, load balancing and high availability. Despite some challenges such as achieving transparency, mitigating network latency and the split-brain problem, clustering has proven to be a huge success for bringing scale and availability to computing applications. Hungry for even more efficient resource use, IT departments are now turning their eye on the next evolution of clustering called grid computing.

Parallel processing is the processing of program instructions by dividing them among multiple processors with the objective of running a program in less time. Parallel processing is normally applied for rendering and high computational based applications. Rather than using expensive specialized supercomputers for parallel processing, implementers have begun using a large cluster of small, commodity servers. Each server runs its own operating system, to take a number of jobs, process them, and send the output to the primary system (Shah, 1999). Clusters provide the ability to handle a large task in small bits, or lots and lots of small tasks across an entire cluster, making an entire system more affordable and more scalable.

The first PC cluster to be described in scientific literature was named Beowulf and was developed in 1994 at the NASA Goddard Space Flight Center (Beowulf clusters compared to Base One's batch job servers). Beowulf initially consisted of sixteen PCs, standard Ethernet, and Linux with modifications and achieved seventy million floating point operations per second). For only $40,000 in hardware, Beowulf had produced the processing power of a small supercomputer costing about $400,000 at that time. By 1996, researchers had achieved one billion floating point operations per second at a cost of less than $50,000. Later, in 1999, the University of New Mexico clustered 512 Intel Pentium III processors that was the 80th-fastest supercomputing system in the world, with a performance of 237 gigaflops.

Just as clustering has reduced the important of supercomputers for parallel processing, clusters are making the mainframe less relevant for batch applications. A batch job is a program that is assigned to the computer to run without further user interaction. Common batch-oriented applications include data mining, 3-D rendering and engineering simulations. Before clustering, batch applications were typically the domain of mainframes that involved high cost of ownership. Now with clusters and a scheduler, large batch jobs can easily be crunched on a less expensive cluster.

Load balancing is dividing the amount of work that a computer has to do between two or more computers so that more work gets done in the same amount of time and, in general, all users get served faster. For load balancing purposes, computers are used together in such a way that the traffic and load on the network are distributed over the computers in the cluster (D'Souza, 2001). Load balancing is commonly used in applications where the load on the system cannot be predicted and is variable from time to time.

One example where load balancing is often used is for web servers where two or more servers are configured in such a way that when one server gets overburdened with requests, they are passed on to the other servers in the cluster, thus evening out the work. In a business network for Internet applications, a cluster, often called a Web farm, might perform such services as providing centralized access control, file access, printer sharing, and backup for workstation users (Server farm). The servers may have individual operating systems or a shared operating system and can offer load balancing when there are many server requests. A Web page request is sent to a "manager" server that determines which of several identical or very similar Web servers to forward the request to for handling to allow traffic to be handled more quickly.

High availability refers to a system or component that is continuously operational for a desirably long length of time. To provide fault tolerance for high availability, the cluster is configured in such a way that the entire system responds to an unexpected error or a hardware failure in a graceful manner (D'Souza,...

Ideally, the resources of the failed hardware are taken over by the remaining nodes in the system and therefore there is no system downtime. This is implemented by mirroring the critical components in the system such as the storage sub-systems and the power supply. Experts suggest that clustering can help an enterprise achieve 99.999% availability in some cases (Cluster computing).
D'Souza (2001) states that one of the largest challenges in implementing a cluster is the bonding of the nodes together. For the systems to appear as a single entity to the users of the clustered system, the internal architecture and the operation of the system has to be transparent to the users. And, reducing the latencies of the network is required so that the response time of the network is acceptable. According to D'Souza, one of the major reasons why most switched-based networks are bogged down is due to the fact that there is a considerable amount of overhead prior to the actual transfer of data over the network.

Transparency and latency reduction are accomplished using a combination of specialized networking methodologies and the controlling software that is used in the cluster (D'Souza, 2001).

When it comes to networking, one of the most common methods is Channel Bonding. This process bonds together two or more Ethernet channels and, therefore, facilitates greater bandwidth between the nodes in the cluster. Channel Bonding is less costly than using a single Gigabit network or any other high-speed network. It is fairly easy to use channel bonding for higher performance when the number of machines is low and they can all connect through the same switch. However, when the number of processors goes up in the dozens or hundreds, the network topology requires a lot of calculating. In this situation, the Flat Neighborhood method is more appropriate. The Flat Neighborhood Network allows a node to talk to any other node on the cluster via a single switch vs. A single switch fabric. Thus, latencies are greatly minimized and the maximum bandwidth of the switch can be tapped. Therefore, in this system, a pair of PCs communicates through shared network neighborhood.

Yet another cluster challenge is the split-brain problem. This happens when a disruption in communication paths between nodes prevents normal internode communications within the cluster (Leng, Stanton and Saibak, 2001). When this happens, nodes may become isolated from the others, presenting the possibility that each groups of nodes may perform as independent nodes. Under this condition, multiple nodes may access the same volume of shared storage, causing data corruption from simultaneous data access.

According to Leng, Stanton and Zaibak (2001) there are several detection mechanisms and remedies for the split-brain challenge. One way is to configure multiple heartbeats for the cluster to eliminate a single point of failure in the communication path. A heartbeat is an "I am alive" message constantly sent out by a server. Heartbeats should never all be plugged into the same network interface card or other network devices such as a hub or switch. And, separate devices should not share a common power source. Public networks must be designed robustly with highly fault-tolerant hardware. And the use of Internet Protocol (IP) over Fibre Channel can be used to provide a heartbeat channel that allows nodes to exchange health information even when all the network connections fail entirely.

Once a server recognizes that there is a failure, it needs to perform cluster reconfiguration that includes input/output (I/O) control of the disksets (Sun product documentation). This is often referred to as resource-based fencing and it involves hardware mechanism to immediately disable or disallow access to shared resources. Resource-based fencing is necessary because a server that has failed may not be completely down and may still attempt to do I/O, because I/O is typically processed by interrupt routines. Thus, the server that takes over must be able reserve the diskset to prevent the sick server from continuing to read or write the disks. The disks or disk arrays block access to all nodes other than those to which they have issued a reservation. This reservation information is stored locally on the disk and remains there until the reservation is either released by the corresponding server, expires, or is broken by a reservation reset request.

In addition to resource-based fencing, there is also an approach called Shoot the Other Machine in the Head (STOMITH) fencing to cope with the split-brain issue (Burleson). In STOMITH systems, the errant cluster node is simply reset and forced to reboot. When it rejoins the cluster, it acquires resources in the normal way. In…