Cluster represents two or more computers (often named the units), united by means of network technologies on the basis of bus architecture or the switch and appearing before users as a uniform information resource. As cluster units can be chosen servers, workstations and even usual personal computers. Clustering advantage for increase of working capacity becomes obvious in case of failure of any unit: thus other unit of cluster can incur loading of faulty unit, and users will not notice interruption in access.

Scalability clusters allows to increase much more productivity of appendices for greater number of users of technologies (Fast/Gigabit Ethernet) on the basis of bus architecture or switch. Such supercomputer systems are the cheapest, as gather on the basis of standard completing elements ("off the shelf"), processors, switches, disks and external devices.

Clustering can be carried out at different levels of computer system, including hardware maintenance, operational systems, programs-utilities, control systems and appendices. Then the more levels are incorporated clustering technology, then the above system reliability, scalability and controllability of cluster. For this reason cluster server decisions are widely used in dedicated hosting solutions

Class I. The class of servers is under construction entirely of standard details which many sellers computer sell a component (the low prices, simple service, hardware components are accessible from various sources).

Class II. The system has the exclusive or not widely widespread details. This can reach very good productivity, but at higher cost.

As already it was specified above, clusters can exist in various configurations. The most used types of cluster are:

• Systems of high reliability.
• Systems for high-efficiency calculations.
• Multiline systems.

Borders between various types of clusters are washed away. Often existing cluster can have such properties or functions which are beyond the listed types. At configurating of big cluster, used as system of a general purpose, it is necessary to allocate the blocks which are carrying out all listed function.

Clusters for high-efficiency calculations are intended for parallel calculations. These clusters are usually collected from the big number of computers. Development such clusters is the complex process demanding on each step accurate coordination of such questions as installation, operation and simultaneous management of greater number of computers, technical requirements of parallel and high-efficiency access to the same system file (or to files) and interprocessor communication between units and coordination of work in a parallel mode. These problems is most easier are solved at maintenance of a uniform image of operational system for all cluster. However to realize the similar scheme it goes right far not always and usually it is usually applied only to not too greater systems.

Multiline systems are used for maintenance of the uniform interface to a number of resources, which can be increased any way in due course (or to be reduced) at a rate of. Most the general example of it is represented by group web servers.

In 1994, Thomas Sterling and Don Becker have created 16-central cluster from processors Intel DX4 connected by a network 10Mbit/sec Ethernet with duplication of channels. They have named it "Beowulf". Cluster has arisen in NASA Goddard Space Flight Center for support by necessary computing resources of project Earth and Space Sciences. Construction work above cluster have quickly turned that is known now under the name project Beowulf. The project became a basis of the general approach to construction parallel cluster computers. It describes multiprocessing architecture which can be used with success for parallel calculations.

Beowulf Cluster consists of one server unit (head host), and one or several subordinated units. Units are connected by means of a standard computer network. The system is under construction with use of standard equipment rooms a component, such as PC, started under Linux (Linux cluster or FreeBSD cluster - if started under FreeBSD), standard network adapters (for example, Ethernet) and switches. Beowulf uses: operational system Linux, systems of message transfer PVM, MPI, control systems of turns of tasks and other standard products. The server unit supervises all cluster and serves the files directed to client units.

The cluster systems architecture in a greater degree defines its productivity, than type of processors used in it. In the critical parameter influencing size of productivity of such system, the distance between processors is. Having connected together 10 personal computers, we shall receive system for carrying out of high-efficiency calculations. The problem will consist in a finding of the most effective way of connection of standard means with each other. As at increase in productivity of each processor in 10 times productivity of system as a whole in 10 times will not increase.

Let's consider a problem of construction symmetric 16-processor's system in which all processors are equal in rights. The most natural connection - in the form of a flat lattice.

The scheme of processors connection in the form of a flat lattice

At such type of connection the maximal distance between processors will appear equal 6. The theory shows, that if in system the maximal distance between processors is more 4 such system cannot work effectively. Therefore, at connection of 16 processors with each other the flat scheme is not effective. For reception of more compact configuration it is necessary to solve a problem about a finding of the figure having the maximal volume at the minimal area of a surface. In three-dimensional space such property the sphere possesses. But as it is necessary for us to construct central system, instead of a sphere it is necessary to use a cube (if number of processors equally 8) or a hypercube if the number of processors is more 8. Dimension of a hypercube will be defined depending on number of processors which are necessary for connecting. So, for connection of 16 processors 4-measured hypercube is required. For its construction it is necessary to take a usual 3-measured cube, to shift in one more direction and, having connected tops to receive a hypercube in the size 4.

Examples of hypercubes

The architecture of a hypercube is the second by efficiency, but the most evident. Other topology of communication networks are used also: three-dimensional, "ring", "star" and others.

Architecture of a ring with full communication on chords (Chordal Ring)

The most effective is the architecture with topology of "a thick tree" (fat-tree). The architecture "fat-tree" (hypertree) is offered by Charles E. Leiserson in 1985. Processors are localized in leaves of a tree while internal units of a tree are grouped in an internal network. Subtrees can communicate among themselves, not mentioning higher levels of a network.

"Fat-tree" cluster architecture

"Fat-tree" cluster architecture (the top view on the previous scheme)

As the way of connection of processors with each other influences productivity cluster more, than type of processors used in it it can appear more profitable to create system from greater number of cheap computers, than from smaller number dear. In cluster operational systems, standard for workstations, more often, freely distributed - Linux, FreeBSD, as a rule, are used. At work with clusters, also as well as with MPP systems, use so-called Massive Passing Programming Paradigm - a paradigm of programming with data transmission (more often - MPI). Cheapness of similar systems turns around a greater overhead charge for interaction of parallel processes among themselves, that strongly narrows a potential class of solved problems.