Most of the time, your computer is bored. Start a program like xload or top that monitors your system use, and you will probably find that your processor load is not even hitting the 1.0 mark. If you have two or more computers, chances are that at any given time, at least one of them is doing nothing. Unfortunately, when you really do need CPU power - during a C++ compile, or encoding Ogg Vorbis music files - you need a lot of it at once. The idea behind clustering is to spread these loads among all available computers, using the resources that are free on other machines.
The basic unit of a cluster is a single computer, also called a "node". Clusters can grow in size - they "scale" - by adding more machines. A cluster as a whole will be more powerful the faster the individual computers and the faster their connection speeds are. In addition, the operating system of the cluster must make the best use of the available hardware in response to changing conditions. This becomes more of a challenge if the cluster is composed of different hardware types (a "heterogeneous" cluster), if the configuration of the cluster changes unpredictably (machines joining and leaving the cluster), and the loads cannot be predicted ahead of time.
Basically there are 3 types of clusters, Fail-over, Load-balancing and HIGH Performance Computing, The most deployed ones are probably the Failover cluster and the Load-balancing Cluster.
Fail-over Clusters consist of 2 or more network connected computers with a separate heartbeat connection between the 2 hosts. The Heartbeat connection between the 2 machines is being used to monitor whether all the services are still in use: as soon as a service on one machine breaks down the other machines try to take over.
With load-balancing clusters the concept is that when a request for say a web-server comes in, the cluster checks which machine is the least busy and then sends the request to that machine. Actually most of the times a Load-balancing cluster is also a Fail-over cluster but with the extra load balancing functionality and often with more nodes.
The last variation of clustering is the High Performance Computing Cluster: the machines are being configured specially to give data centers that require extreme performance what they need. Beowulfs have been developed especially to give research facilities the computing speed they need. These kind of clusters also have some load-balancing features; they try to spread different processes to more machines in order to gain performance. But what it mainly comes down to in this situation is that a process is being parallelized and that routines that can be ran separately will be spread on different machines instead of having to wait till they get done one after another.
Most common known examples of loadbalancing and failover clusters are webfarms, databases or firewalls. People want to have a 99,99999% uptime for their services, the internet is open 24/24 7/7/ 365/365 not unlike in the old days when you could shut down your server when the office closed.
People that are in need of cpu cycles often can afford to schedule downtime for their environments, as long as they can use the maximum power of their machines when they need it.
Traditionally Supercomputers have only been built by a selected number of vendors: a company or organization that required the performance of such a machine had to have a huge budget available for its Supercomputer. Lots of universities could not afford the costs of a Supercomputer by themselves, therefore other alternatives were being researched by them. The concept of a cluster was born when people first tried to spread different jobs over more computers and then gather back the data those jobs produced. With cheaper and more common hardware available to everybody, results similar to real Supercomputers were only to be dreamed of during the first years, but as the PC platform developed further, the performance gap between a Supercomputer and a cluster of multiple personal computers became smaller.
There are different ways of doing parallel processing: (N)UMA, DSM, PVM and MPI are all different kinds of Parallel Processing schemes. Some of them are implemented in hardware, others in software, others in both.
(N)UMA ((Non-)Uniform Memory Access), machines for example have shared access to the memory where they can execute their code. In the Linux kernel there is a NUMA implementation that varies the memory access times for different regions of memory. It then is the kernel's task to use the memory that is the closest to the CPU it is using.
DSM aka Distributed Shared memory, has been implemented in both software and hardware , the concept is to provide an abstraction layer for physically distributed memory.
PVM and MPI are the tools that are most commonly being used when people talk about GNU/Linux based Beowulfs.
MPI stands for Message Passing Interface. It is the open standard specification for message passing libraries. MPICH is one of the most used implementations of MPI. Next to MPICH you also can find LAM, another implementation of MPI based on the free reference implementation of the libraries.
PVM or Parallel Virtual Machine is another cousin of MPI that is also quite often being used as a tool to create a Beowulf. PVM lives in user space so no special kernel modifications are required: basically each user with enough rights can run PVM.
The openMosix software package turns networked computers running GNU/Linux into a cluster. It automatically balances the load between different nodes of the cluster, and nodes can join or leave the running cluster without disruption of the service. The load is spread out among nodes according to their connection and CPU speeds.
Since openMosix is part of the kernel and maintains full compatibility with Linux, a user's programs, files, and other resources will all work as before without any further changes. The casual user will not notice the difference between a Linux and an openMosix system. To her, the whole cluster will function as one (fast) GNU/Linux system.
openMosix is a Linux-kernel patch which provides full compatibility with standard Linux for IA32-compatible platforms. The internal load-balancing algorithm transparently migrates processes to other cluster members. The advantage is a better load-sharing between the nodes. The cluster itself tries to optimize utilization at any time (of course the sysadmin can affect the automatic load-balancing by manual configuration during runtime).
This transparent process-migration feature makes the whole cluster look like a BIG SMP-system with as many processors as available cluster-nodes (of course multiplied with X for X-processor systems such as dual/quad systems and so on). openMosix also provides a powerful optimized File System (oMFS) for HPC-applications, which unlike NFS provides cache, time stamp and link consistency.