The course HARDWARE 2: Servers, networks and communication

7. The specific hardware for network servers.

7.1. Introduction to the servers - 7.2. External specificities - 7.3. Internal specificities - 7.4. Basic configuration of a server - 7.5. Memories servers - 7.6. Internal ports - 7.7. Processors - 7.8. The techniques multiprocessors

7.1. Introduction to the servers networks

Using of heavy network (Windows NT / 2000 / 2003 / 2008, Linux or Netware Novell) imposes one or more host computer. These computers can be of all types, including mainframe. We will be interested here in the servers networks of the type PC X86. A server may be a normal PC for a small network and a server can be used like workstation the top-of-the-range one. Nevertheless, the majority of the servers are specific computers. The characteristics of these computers must answer various criteria related to application and data security installed on the server. 

With the evolution of data processing in the company, the host computer and the peripheral installation become paramount, the least stop of the installation immediately causes the stop of the factory with the consequences than one can imagine. That this stop is of software problem type, related to a malfunction of a server or even on a recalcitrant switch is without real importance. The result is the same one for the company: loss of production, loss of data, ... A network installation should not stop. If it made its preparatory work of the possible causes of breakdown (and the immediate solutions), a good maintenance man must "walk" in the factory. The least stop, especially if it lasts, is paid cash, to minimize the durations as much.

With the following chapter, we will make a turn on the side of the hard disks, their types and of their regrouping (RAID). Nevertheless, the servers of large powers are not limited to hard disks or external protections. Internal architecture is also different. If a server should never stop, it is necessary also that its power is sufficient for the application. For recall, a server should not function with more than 10 % of load on average, under penalty of slowing down the application and the users. This does not actuate muscular servers inevitably. According to the application, the number of connected PC, the function of the server (file, printer, program), the configuration will be selected consequently. Let us see the various solutions which were adopted for (to try) obtaining result.

7.2. External specificities

External protection to avoid to stop of a server by an indelicate hand are established on the servers of high range.

• Blocking by keys of the buttons of lighting, reset, access to the disk drive... although this is not most important.
• Connecting cables, must be fixed on the case or in a cable shelf, to prevent that the cables are not torn off by traction... It must be labeled.

7.3. Internal specificities

• fast disks hard to give information as soon as possible after a request of a connected customer, if possible multi-sessions, which implies connections SCSI.
• data security in the event of "crash landing disk" (breakdown supplements of a hard disk): unfolding of the discs by systems RAID
• data security in the event of breakdown of power supply or fall of tension: UPS
• data security in the event of breakdown supplements of a computer (daily safeguard)
• installation of discs, electronic board "hot plug". This means that the peripheral can be replaced whereas the server continues to function.

By seeing the list, this does not appear too difficult. Let us avoid the external aspect. This is related to the case on the one hand and... with the technicians, administrators networks for the aspect wiring: no wiring with goes quickly in the passages. Moreover, the computer rooms conceived for the servers networks are generally installed with a false floor, which makes it possible to pass the cables and provided below with a system fire-resistant with gas to freon, an inert gas. At the price of a bottle, avoid lighting your cigarette in these rooms. Temperature is generally under controlled of 18° centigrate.

Each server is connected to a UPS, uninterruptible power supply (inverter). The inverter is of able to control the server (to stop it) in the event of or fall current failure of tension via a specialized program. It is also often the case of the HUBS and switch and external peripherals additional.

The internal aspect is more complex. The hard disks are generally SCSI. This type of connection is hardly more powerful in station than hard disks E-IDE, except at the access times. On the other hand, connections SCSI are more powerful in multi-read. In the event of crash landing disc, one uses systems redundant of the discs (RAID). One writes the data on each discs (in an equivalent way), but one reads only on one disc. In the event of crash landing, one continues on the second disc. In the event of crash landing complete of a server, one can permanently couple 2 servers in the same way as the discs above. All these systems are known as RAID. For the power supply, one can use 2 coupled power supply (duplicated or redundant). In the event of breakdown of one, The other continues. Each one must thus be does not measure to feed the autonomous whole of manner. For the electronic boards, the PCI 64 makes it possible to remove or insert a board without switching off computer, provided that the server is installed with a compatible operating system (for example Win2000). The unfolding of each part of a data-processing installation to guarantee that if one break down, the second part takes the place immediately calls the redundancy of the equipment. In short, there is job.

On the level of the peripherals, we know of them already some. The backup tapes are practically always used, but not really as small safeguard if with speeds of transfer which have nothing to do with the bands of the types QIC of first year. Moreover, one finds into external cases containing the hard disks for example, always in RAID.

7.4. Basic configuration of a server.

Before returning in purely technical solutions, let us see a little the use of a network server. By definition, a server is not a workstation. Result, the graphics card, the CD-Rom reader and the floppy drive of the computer are not paramount components. The screen should not either be a multi-media model of high range. The screen of the server is generally one 15"(even one 14 black and white) which turns" in neutral ". One is interested in posting only in the critical cases. The CD-Rom reader is generally not either of type SCSI but well E-IDE, considering his weak use. According to the operating system, one can (or one must) configure it via a station.

The memory must when it to be sufficient, the discs of double capacities, even triple compared to the maximum capacity which you will use on this machine. I speak here about the capacity effective, usable, without returning in technology RAID

7.5. Memory server.

First comparisons with the traditional memories, the memories used by servers are the same uses by traditional PC. The current servers use memories ECC (Error Checking and Correcting or Error Correction Codes). This technology uses several bits of control (parity) for the checking of the data in memory. These memories are car-corrective. This memory can detect 4 errors and correct one without stopping of them the system. A new type of memories with correction of error AECC (Advanced Error Correcting Codes) can detect 4 errors and correct 4 without stopping of them the system. Inevitably, these mechanisms slow down a little the system.

7.6. Internal bus. 

In office automation, the current ports are PCI 32 bits and AGP. These ports have two problems. The first, they are not hot plug. Board replacement requires to stop the server. In the small servers, this does not pose problems in practices. Indeed, as each function is carried out by only one chart, the server does not ensure in any event plus its function in the event of breakdown of chart. On the other hand, in the servers of high range, all the charts are redundant. A chart network is duplicated. In the event of breakdown of a chart, the function continues on the second equivalent chart. This makes it possible "to repair the server" without stopping it.

The second problem of these ports are related to speed. For recall, an office automation port PCI is able to transmit 132 maximum MB/s on the whole of ports PCI (divided band-width). Let us check for example a  network board bases 1000 (Ethernet Gigabits). As the connections network are of series type, one can divide the rate of transfer by 10, which makes us 8 per only one chart network a rate of transfer of 100 MB/s. Perhaps you think that there remains walk, but... hard disks SCSI are also connected via an interface on port PCI. Let us take for example, a connection hard disk Wide Ultra 3 SCSI (Ultra 160/m) which are able to transfer 160 MB/s: 100 MB/s + 160 MB/s = 260 MB/s, largely higher than the 132 MB/s of port PCI 32 bits.

All the servers worthy of this name must thus use internal faster bus: the PCI-X. Developed jointly by principal the actor of the data-processing hardware, the servers use connections PCI on 64 bits (the rate of transfer is thus doubled). Moreover, common port PCI uses a speed of 33 MHz. Ports PCI-X go up to 533 MHz. This gives us a rate of transfer of 533 MB/s * 8 (port 64 bits) = 4256 MB/s for the whole of bus PCI-X. Generally, a server also accepts 1 or 2 port PCI 32 bits (chart screen for example or Ethernet 100 of reserve). The ports 64 bits accept generally only the charts 32 bits functioning in 3,3 V to recognize the ports 64 bits which accept charts PC 32 bits, it is enough to check if there are 2 notches (only charts 3,3 V) or 1 notch (accepts charts 3.3 and 5V) in part 32 bit of bus PCI 64 bits.

With ports PCI-X, we find the awaited characteristics: speed and Hot plug (if the driver board allows it). A last precision, these boards and the installation of these bus are expensive. Each server does not include an office a PCI-X to 533 MHz. There are charts with 33, 66, 100 and 133 MHz. Moreover, many server do not include one, but 2 or three separate ports PCI-X. This also makes it possible to remove the necks.

PCI-X 1.0 left in 1999, with a supply voltage of 3,3 V:

Version PCX-2.0, left in 2002, is also fed in 1,5 V according to the versions. The boards are hot Plug.

With these characteristics, an office PC cannot be used as server network.

7.7. Microprocessor server.

    For the effective characteristics of the processors dedicated to the servers networks, you can refer in the page microprocessor server. This part takes again only the general cases.

7.7.1. Introduction

The processor of a server is not office an animal of competition. A server dos not create multi-media-applications. Except for the servers of programs, the processors are generally "weak". A server of Web can at ease be satisfied with Pentium III, even of a CELERON.

On the other hand, in the heavy applications, the manufacturers of processors moved towards two directions: specialized processors and the multiprocessor. Both are partly dependent.

The current processors are 32 bits. This means that the instructions out of assembler that they read are coded on 32 bits. With the roadhogs of data processing, to increase the performances of a processor, you can either increase speed, or to double the number of instructions per cycle of clock. This solution already used, but the processors 64 bits use this possibility differently. Indeed, like the current processors, the programs are written in 32 bits. A processor 64 bits cannot thus read instructions 32 bits and screw poured. INTEL with its processor 64 bits ITANIUM left in July 2001 circumvented the problem by not taking the old instructions 32 bits (that which we know). This required to rewrite the programs and operating systems or rather recompiler, i.e. to reconvert the program assembler 32 bits in 64 bits. Windows 64 bits exists for these processors, but few programs are really on the market. This reduces Intel Itanium to computer servers or very high range stations. AMD chose the opposite way. While creating a processor 64 bits kept compatibilities 32 bits. The AMD 64 bits thus carry out as much the current applications that the applications 64 bits.

A last thing, the use out of bi-processor and superior requires an operating system adapted. Windows NT, 2000 and XP Pro are sold in manner specific. Novell obliges an additional option. UNIX - Linux is native multiprocessors, if the function is established according to the mother chart/OS. The versions "home" of the operating systems Microsoft (Win95,98 Me and XP Home) do not manage the multiprocessor.

7.7.2. INTEL

At INTEL, the processors specialized 32 BITS are of type XEON (more old Pentium Pro). Compared to Pentium normal (Pentium III, Pentium IV), INTEL generally inserts masks more important L1 and L2. To perfect work, the socket and the chipsets are different. Nevertheless, Pentium III could be used out of bi-processor. INTEL has to remove this possibility in an internal way, but not in the majority of the CELERON. 

Itanium and Itanium II are reserved for the demanding networks since they work in 64 bits.

Notice, to add a second processor (on a mother chart which accepts it) requires a of the same processor speed and (in practice) of the same series of manufacture. This is not always easy to obtain.

A last remark, with Pentium IV to 3.06 Ghz, INTEL includes from now on, L'hypertreading. This technique makes it possible to emulate two software processors in only one Pentium. The advantage would be related to speed but the various tests are mitigated enough, in particular because the application must be dedicated to this process in the case of workstations. On the other hand, this function is largely established in the ITANIUM and XEON.

At the beginning of 2004, INTEL announces architecture NOCOMA for the XEON. This modified processor is a processor 32 bits able to carry out certain applications 64 bits. It is thus comparable with Opteron of AMD and positions between the normal XEON and the ITANIUM Full 64 bits.

7.7.3. AMD

AMD produced since September 2001 of the specific Athlon processors able to work out of bi-processors, with such a specific chipset him. They are Athlons of the type MP (multiprocessors). The opteron (version server of Athlon 64 bits) allowing to use to 8 processors simultaneously.

AMD develops 2 versions of its processor 64 bits: Opteron and Athlon 64 bits. Opteron, left in April 2003, is the version server - station data-processing of high range, as well as Itanium and its successor Itanium II. Athlon 64 bits for stations left in September 2003. For recall, these processors are also compatible 32 bits and can thus be used with operating systems 32 usual bits.

The principal modifications compared to K7 architecture comes from the number and the size of the registers (the working memories intern) which must support at the same time new instructions AMD64 of AMD and SSE of INTEL.

Opteron are engraved in 0,13 µ, just like Athlons current, and use a specific socket of type 940. The L2 cache passes from 512K to 1 MB. Management memory does not deal any more with the chipset, but well directly with the processor which manages 2 benches (32 bits) of DDR333.

Opteron uses 3 Hypertransports bus which can be connected directly to another processor (3,2 GB/s into bidirectional) that is to say with a chip managing the PCI-X or AGP.

As Opteron is dedicated server, it is declined under 3 versions: 100, 200 and 800 which respectively have 0, 1 and 3 buses processors which can be used for machines using 1,2 and 8 opteron. Each version is declined in various speeds. Version 200 is currently the only available one with speeds of 1,4, 1,6 and 1,8 Ghz.

7.8. The techniques multiprocessors.

To work with several processors simultaneously (in the same machine) inevitably requires a mother chart which accepts it. The principle must make it possible to share the memory, the accesses discs and in general all the internal buses. 

Two techniques are currently used: the SMP with commutated bus (Symetric multiprocessing) and the multi-processing Numa. The difference between the two start to be reduced, quite simply because the manufacturers start to mix both, even if the principle of operation is different.

The SMP is especially used for a small number of processors, Numa is better for a great number of processors.

7.8.1. SMP, standard UMA (Uniform Memory Access)

Architecture SMP consists in using several processors sharing the same memory and the same internal peripherals. Only one operating system makes turn the sets of the processors. Following several technological projections, the limits of this principle were pushed back. Indeed, to divide does not want to say use at the same time.

Structure standard SMP (UMA)

The system bus was a long time the weak point of the SMP. Thus, the first multiprocessors made communicate the processors between-them via shared systems buses. Those quickly became saturated beyond some processors. The increase in the memory hiding place and the increase in the work frequency of this bus have made it possible to improve the performances of a server. Nevertheless, the upgrading capabilities of these buses are weak, the band-width remaining in all the cases constant.

To work out evolutionary platforms, the manufacturers of processors worked on architectures with commutated buses. This A made it possible to create infrastructures of interconnection whose band-width could be increased by stages, thanks to additional switches. This type of connection is at the base of modular systems. The elementary components are not any more the processors, but boards girls Bi or quadri processors inserted in connectors on a basic central board. It is Sun which used this technique the first with a machine able to exploit until 64 microprocessors simultaneously. The board accommodating the girl board allows a flow of 12,8 GB/s and makes it possible to plug in until 16 boards four-processors. Each addition of boards four-processor sees the opening of channels of additional interconnection and  thus an increase in the band-width. In system SUN, the memory is localised on each board girl. She thus seems held by board. In fact, all the accesses report are made by by the central bus, whether the access is on the same board girl or another. By this principle, technique SUN uses a technique SMP. Each manufacturer currently uses a technique if not identical, at least equivalent. Certain firms have nevertheless to insert a local controller on each board girl.

In the PC world, it is INTEL (via repurchases) which is leader. INTEL uses a commutation controller of 2 buses of access report, of 2 buses to access to modules four-processors (the total is thus limited to two boards, that is to say 8 processors) and 1 bus of inputs/outputs. The whole is supplemented by cache accelerators.

Another solution, currently deployed by Unisys consists in using an operating system per processor (architecture NUMA below). This currently makes it possible to use to 32 processors simultaneously, in Windows 2000-NT, Netware or UnixWare.

7.8.2. Multi-processing of the type Numa (No Uniform memory access)

Structure NUMA

Architecture NUMA makes it possible to use more processors. Technology makes it possible to gather groups of processors, using their own local memory, and to connect them between-them by buses able to deliver several giga Octets a second. By no uniform access to the memory, it should be understood here that a processor will not reach within the same times a data in memory if this one formed part of a local or distant memory. This difference in times is reduced nevertheless, thus gathering architectures UMA and NUMA. The memory is by the whole of the processors. This implies that system NUMA exploits a management of coherences of the memory hiding place able to take into account the whole of the processors attached to the platform.

Technology multiprocessor is not based nevertheless solely on the bus management of connection. The communications on the buses of interconnection must also allow to maximize the treatment of the tasks between the processors.

A last remark, and of size, architecture NUMA obliges that each processor makes turn its own operating system, whereas in case SMP, only one operating system turns for the whole of the processors. This thus dedicates NUMA for systems UNIX multiprocessors or owners and SMP for the world of servers INTEL - Windows, even AMD Opteron uses NUMA (Memory controller is included in processor).

1. Hardware course II - 2. Introduction to communication - 3. Base of transmission -  4. Ethernet - 5. Switch, router, ... Ethernet - 6. Internet connections - 7. Networks servers - 8. SCSI, SAS, RAID - 9. Backup Tape drives, NAS, ... - 10. Hardware Security - 11. Wireless and Wifi - 12. UPS, Inverter, Power Supply - 13 . Corporate network - 14. Alternative technologies - 15. Touch Screen, video projector 17. Exercise: hardware firewall

The Hardware 1 course: PC and peripherals, the Hardware 2 course: Network, servers and communication.

The hardware training.

© YBET data processing 2006