There is no server which approaches for all possible problems. For example, the equipment for e-mail corporate system support and for a file server will be essential to differ, and cost can be similar. Therefore at a choice of a server pay attention not to cost of the equipment, and in what mode this server will carry out what problems. The unsuccessful choice can entail not only excessive expenses, but also threaten integrity and availability of the information and services. It can make impossible the enterprise normal functioning.

Purchase of a server it always the compromise between desirable productivity and financial opportunities. Certainly, even the small company can buy dear and productive equipment, whether but so it is necessary to do it if the most part of time this equipment will stand idle? We shall describe the general recommendations on servers selection for various problems.

Productivity of any server depends on following parameters:

• Type and productivity of processors.
• Volume and type of operative memory
• Productivity of a disk subsystem

CPU - heart of any computer system. In the market today there is the richest choice of processors from different manufacturers. For a successful choice from this variety it is necessary to understand technologies present in the market well enough.

Main parameters of processor system are: quantity of processors, their frequency and volume built in a cache - memories.

Owing to company Intel Frequency (quantity of operations which the processor is capable to execute for a second) the processor long time was considered as a unique parameter of productivity. Partly it is valid so - the slow processor quite can really make all system unproductive, not having had time to process all acting data. If to not consider other factors the mathematics is simple enough - the above frequency, the above productivity.

Cache-memory. One of the most essential parameters at work with databases. The cache is the memory built in the processor which serves for masking references to operative memory. The matter is that the processor in any case works much more quickly operative memory, and the difference makes not percent, and tens times. Accordingly, at insufficient volume a cache-memory the processor should pass steps and to wait while the necessary data will not be loaded from operative memory. It cannot be named a problem by transfer of data large volumes as thus data directly through the processor do not pass. The cache is important basically for work with dense files of the information (databases). The reason is simple - unlike simple data transmission at which linear reading is carried out, at work to databases there is practically casual reference to different points of hard disks and, at enough great volume of base, time spent for search, becomes unfairly long.

That time to reduce, recently requested data move (through operative memory) to a processor cache. As a rule, with databases enough plenty of users and the more a cache lumpsum works, the data can simultaneously obtain a lot of users.

Further the small deviation, devoted is necessary for a situation in the present market of processors for "lungs" and "average" servers. This market is divided between two companies - AMD and Intel with their rulers Opteron (AMD), Xeon and Itanium (Intel).

XEON (Intel).

The processor has appeared enough for a long time and has quite good productivity for the moderate money. Today in the market models with frequencies from 1,5 up to 3,66GHz and with volumes a cache-memory of the third level from 1 up to 8 Mb are presented. Lack of these processors consists that for connection of several processors one general half-duplex trunk which in case of the intensive reference to operative memory becomes "bottleneck" of system is used.

The trunk has not too high parameters for a server: word length of 128 bats and speed of 400 MHz, the maximal speed of data transmission makes 6,4 Gb/sec. In these conditions unique way to lower loading on the system trunk is the increase in volume a cache-memory, that we and observe. Models with indexes DP (for use in dual-processor servers) and MP (for four-processor servers) are issued.

Systems on the basis of XEON do not support more than 4 processors.

ITANIUM (Intel)

Rather recently appeared family of processors. Differ a little lower frequencies, very in great volume cash the third level - its volume can reach up to 9Mb and the expanded support of 64-bit architecture. Unfortunately, these processors have been ambiguously accepted by the market - their high price and complexity of platforms compatible creation to them have made their unattractive for wide application. The mite was brought also with refusal of their support by Microsoft. All these factors have defined position Itanium in the market as processor of the highest level applied to construction high-efficiency multiprocessing (from 64 up to 256 units) systems. Also use in structure of clusters, though because of costs on data transmission between processors cluster productivity always below, than high-grade multiprocessing system is justified.

OPTERON (AMD)

Family of the server processors presented by company AMD. In these processors it is not realized any essentially new technologies, short of high-grade support of 64-bit architecture.

They differ from series Xeon that processors are connected to the general switched memory. Each processor gets access to a demanded site of memory on the switched channel. Memories coherency at such architecture to provide easier. As a result such systems are better scaled, and their speed of the response as a rule appears above.

In the market models with frequencies from 1,4 up to 2,8 GHz with marks 1xx (uniprocessor a server and workstations), 2xx (a server and are presented to station up to 2 processors) and 8xx (support up to 8 processors). The small volume cash the second level (1 Mb) is compensated by the high-efficiency trunk Hyper Transport supporting frequency in 1GHz (800 MHz for Opteron the previous generation).

All above-stated is rather actual at a choice of multiprocessing system. Certainly, the concrete choice of architecture can be made only after the analysis of concrete problems shown to them, but as a whole it is possible to recommend the following.

Processors Xeon are sense to use for a file-servers and other systems which will simultaneously not process a plenty of fine inquiries. At such problems the processor "does not banish" through itself (so through the trunk) excessive volumes of the information, hence "bottleneck" of series Xeon cannot considerably affect productivity. Besides because of technological features on one server installation more than four such processors is impossible.

Opteron do not possess such frequency as processors Intel, but have a number of other advantages, namely - high throughput of the trunk and hardware support of 64-bit architecture. It allows them to address practically unlimited operative memory volume. Thus, optimum application of processors Opteron - a server of databases support. On one server it is possible to establish about 8 processors that provides fine productivity.

Some words about multinuclear processors. Equipment of one processor by several kernels is attempt to receive advantage cluster systems without its lacks (not enough fast switching of cluster units). Certainly, installation twice-cores processor on productivity negatively will not affect, but also advantages can not give. Twice-cores gives advantage only in parallel applications, there, where it is necessary to process a plenty of simultaneous inquiries.

Still anybody never complained that at it in a server, no less than in a workstation too much RAM. But memory for servers, unlike workstations costs much more dearly and has much greater value by way of productivity.

As has shown an expert, on the average for a server of databases than half from volume of the database should suffice 256 Mb for needs of operating system, approximately on 64 Mb on each user actively working with base plus not less.

On sale it is possible to meet modules of memory in volume in 256, 512,1024, 2048, 4096 Mb. For correct work of the equipment memory can be increased only by doubling the existing volume.

There are two basic standards of memory - DDR1 and DDR2. They differ in the speed of data transmission - for DDR1 it 266 MHz, 333 MHz, 400 MHz, for DDR2 - 533 MHz, 667 MHz, and 800 MHz. Here all is simple - the above frequency, the better. It is necessary to consider only, that these standards are among themselves incompatible and at purchase of the server focused on the further growth, it is desirable to get a platform supporting DDR2.

One more important point to which it is necessary to pay attention at a choice of memory - presence at it of function ECC (Error Correcting Code). Memory with this function automatically corrects the mistakes arising during work. Mistakes in work of memory render strong negative effect on productivity of a server and can lead to different consequences, down to loss of the information. ECC memory works a little bit more slowly usual (approximately on 5 %) and costs much more dearly. But at construction of data processing reliable systems is an obligatory component.

At a disk system choice it is necessary to start with that on what performance of problems the server will be focused that is that for it is more important - low time of information search and an fast processing opportunity of simultaneous big number inquiries to it, volume of carriers or cost.

The hard disks which are present today in the market, differ with the interface of connection (SCSI, SATA1-2, Fiber Channel, SAS), volume and speed of a spindle rotation.

The choice of the interface, besides, depends on the problems which are carried out by a server. For fast search of the necessary data in the dense information environment installation of disks with interface SCSI is desirable. These disks cost dearly enough, but speeds of a spindle rotation - 10 000 and 15 000 turns in a minute possess the lowest time of access to the information due to greater, than at SATA. Volume SCSI - disks does not exceed 300 Gb, speed of data transmission - 320 Mbit/sec (for standard Ultra320 SCSI). All this does by their quite good decision for application in the systems working with databases and engaged complex calculations.

SATA - disks, on the contrary, possess low speed of access, but the volume of disks reaches 500 Gb and their cost considerably below what SCSI. Speed of a spindle rotation - up to 7 200 turns. Such stores are optimum for storage of the information to which there are no constant frequent inquiries, for example for FTP-servers or the servers organization of the general access in Internet.

Fiber Channel - development of ideas SCSI. At use of this protocol data are transferred on the optical channel. This interface possesses the highest speed (up to 4 Gbit/sec), but for application demands special (and rather expensive) an infrastructure. Disks with this interface are applied in the systems focused on the maximal speed.

SAS (Serial Attached SCSI) - the new interface directed not only on increase of stores productivity, but also on unification of storage systems. Speed of data transmission - up to 3 Gbit/sec, probably consecutive connection up to 16 256 devices. Innovative feature SAS most, probably, - full compatibility with popular interface SATA because of the profitability. Thus, in one case it is possible to place simultaneously both high-efficiency SAS, and economic SATA stores. Besides SAS provides connection both standard 3.5", and 2,5" HDDs that does its by the extremely attractive to application in compact blade servers.

Without dependence from the used interface it is desirable to choose disks with the greatest possible for that interface in the speed of a spindle rotation.

The primary goal of a server in any organization - uninterrupted granting to users of the services. Unfortunately, no technics is ideal and sooner or later separate components of a server can fail. Completely to avoid that it is impossible, but quite really to make so that consequence of failure have not appeared, on the one hand, fatal, and with another - could be corrected in the minimal terms and with the minimal expenses.

If to speak about reliability of a data storage it can be raised by creation of failure-safe scheme RAID. Many system payments have built in RAID-controllers, but their reliability can appear insufficient. For creation really failure-safe RAID it is necessary to use only external RAID-controllers.

As to maintenance of work uninterrupted operation it is possible to recommend application of a server in which it is possible to establish reserve power units and supporting hot-swap of HDD. All this allows to replace defective components of system without its stop.