Usually, when it is necessary to make the hardware fault tolerance, simply add superfluous components. It can be executed by means of a standard architecture copy use which will work in a mode of expectation. However, more beautiful decision, use of specially developed architecture is.

Here the configuration with the several processors using the general memory is shown. Depending on details of a configuration and from operational system processors can work in three modes.

Parallel independent mode.

Each processor carries out the set of commands that leads to high speed and to zero fault tolerance. If in any processor there is a failure the processes which are carried out on it stop all; to restore them it is uneasy.

Mode of the virtual processor.

The given mode is fault tolerance. During each moment of time all processors execute the same commands, and results of their work are compared with each other. The so-called voting mechanism makes a decision concerning failure of the processor and automatically disconnects it. Switching-off the failure processor occurs absolutely imperceptibly for the user, productivity does not decrease, as other processors continue the work.

However the similar architecture contains individual points of refusal - components which refusal leads to a full stop of system. Such component in this case is memory.

The architecture, allowing to solve this problem, is resulted in a Fig. 2. Two systems are connected through the channels of input-output for reception large-unit full redundancy. For preservation of data integrity and grantings to all systems of cluster identical access to the general data file for a data storage are usually used multiport RAID-systems.

Usually one of two systems works as the basic server, and another is reserved. Using the special software, systems constantly check each other "You are still alive?". As soon as there is a failure in the basic server, safety at once takes all its functions on itself.

Multiprocessing systems have been developed for elimination of architecture lacks with the divided memory (individual points of refusal) without shared resources.

As shown in a Fig. 3, completely independent processors, everyone with own memory and peripheral devices, cooperate with each other through the high-speed local trunk.

In this case the trunk is dual so refusal of one trunk will not lead to crash of system. Refusal of one or even several modules of such system also will not cause system stop. The corresponding software is responsible for that the processes which are carried out on the given up processor, have been restarted on other processor.

Logic continuation of the architecture described just was recently developed technology of connections ServerNet. Being realization of a so-called system network, it provides high-speed connection of modules the processor-memory and peripheral devices in the any order.

This network is completely dynamically configured, and practically has no limit of expansibility. Flexibility of a network allows to carry out its configuration for the applications which are carrying out a plenty of calculations, data working with a lot.

Architecture ServerNet

Integrity of data can be reached due to a combination of methods multiprocessing processing with the general memory and cluster multiprocessing processing (the working part of cluster incurs all functions of a defective part). For achievement of full fault tolerance it is possible to make itself ServerNet superfluous and to connect all devices to both planes in parallel.

As a rule, servers Internet do not demand the big computing capacity, but huge volumes are necessary for a data storage and good lines of communication with high throughput for them.