If you have solved, that cluster is that did not suffice you for your work. There is a question from what it to do and how many computers it is necessary to connect in cluster to reach objects in view. Also it is necessary to understand it from what computers to collect?

Cluster Beowulf - heterogeneous structure. The diversified computers constructed on various hardware platforms can enter into it. Moreover, on computers in cluster the most various operational systems can be established: Linux, Windows, FreeBSD. That we shall simplify a problem of construction of system to consider, that the hardware platform of computers of our future cluster is monotonous.

As to distinction in parameters (speed, memory) at the computers entering in cluster it is admissible. But in this case, you should consider these distinctions at a writing of parallel programs, distributing volume of the account depending on an opportunity of each separate computer. Otherwise cluster will work as the system consisting of PC with the minimal working parameters.

Let's begin with a choice of the cluster size. As cluster Beowulf - scaled system the question of quantity of units is not important. In process of growth of your needs you can add quantity of units at any time. If you for units will use the removed loading operational system on a network (for example, problems of cluster hosting) works on addition of unit of cluster will not be beyond technical connection of a new server in a network.

To take a great interest in quantity of units it is not necessary. The bottleneck in cluster is an environment of data transmission between units, that is throughput of a used network.

Thus, it is possible to recommend at initial construction clusterą it will be limited to four units. On the one hand, you if necessary can always increase cluster, on the other hand, the smaller quantity of units can give not so good result, predictably.

Cluster Beowulf consists of separate PCs (units) and a network (switch) uniting them. Except for OS, it is necessary to establish and adjust network drivers, compilers, specialized for distribution of problems and computing loading.

Units of cluster. A suitable choice at present are systems on the basis of processors Intel Pentium 4. It is necessary to establish on each unit not less 128MB operative memory. It is necessary to allocate one of PCs as central (the console of cluster) where it is possible to establish enough the big hard disk, probably more powerful processor and it is more than memory, than on other units. To do the console of cluster by more powerful server it is meaningful, if you want to have on this computer except for a command line more convenient operational an environment, for example KDE, Gnome, office programs, programs of visualization of given, etc.

It is meaningful to provide (protected) communication of this PC with an external world. In other words, the cluster network (a network consisting of cluster) should not their consoles of working units is inside of a corporate network. If it is necessary to provide access to the console of cluster from a corporate network and-or the Internet in this case, communication should go through the separate network card established in the main computer, and the separate switch.

At a complete set of working units quite probably to refuse hard disks - these units will load OS through a network from the central server, that, except for economy of means, allows to configure OS and all necessary ON only once (by the central server). If these units will simultaneously not be used as the user workplaces, there is no necessity to establish on them a videocard and monitors. Installation of units in racks is possible, that will allow to reduce a place borrowed in units, but will cost a little bit more dearly.

The organization of clusters on the basis of already existing networks of workstations is possible, that is workstations of users can be used as units of cluster at night and in days off. Systems of such type name COW (Cluster of Workstations). For cluster systems of type COW are not present an absolute must to stop cluster (and problems on it considered) for day time (working) time when behind central machines users work. Work of parallel programs will be slowed down certainly, but it not deadly. The Another matter if work of cluster it will be appreciable to brake and complicate work of users.

In the elementary case for communication between units of cluster one segment Ethernet (100Mbit/sec on twisted pair) is used. However cheapness of such network, owing to collisions turns around a greater overhead charge for interprocessor exchanges, and good productivity such of cluster it is possible to expect only on problems with very simple parallel structure and at very rare interactions between processes (for example, search of variants).

For reception of good productivity of interprocessor exchanges use Gigabit Ethernet. Thus for reduction of number of collisions or establish some "parallel" segments Ethernet, or connect units of cluster through the switch. Such structure of a network when each unit of cluster has more than one network card which by means of special drivers are united in one virtual network interface having total throughput is meant "parallel" segments. To avoid problems with configuration such virtual interface, it is necessary to use identical network cards by all cluster servers. Besides each parallel line of such interface should be an Ethernet-network constructed on separate (from other lines parallel to it) the switch.