Ben's Server and Workstation Primer

Original Article Date: 2004-08-15

Caseology
  Rackmount or Tower?
  What are U talking about?
  Serverboard Form-Factors
Server CPUs
  Intel Xeon vs. AMD Opteron
  Xeon Generations
  Decyphering Opteron Numbering
  Pentium 4 Processors in Servers and Workstations
Server and Workstation Boards
Other Considerations
  Storage Devices
  Server Memory
  Separating PCI-X from PCI-Express

Servers are a different breed to desktop computers, because of the very different purpose they fulfill. From a small corporate network file-server to one of a thousand racks running databases at Google.com, the thing they have in common is that they are there to process other computer's requests. They serve.

They usually demand more memory, more processing power and faster hard-drives. Ok, so a fast desktop computer could use those. But because of the simultaneous work-loads that servers undergo, they usually want more than one processor - so dual processor servers are often the norm, and even quad processor systems are not unknown. Servers usually demand extensive networking capabilities, and this in turn demands a faster peripheral bus and other enhancements.

Workstations are a bit of a halfway-house between desktops and servers. Really, they are just super-high-end desktop PCs, running high-performance CAD or imaging software, or crunching vast quantities of numbers for scientific or financial use. They work, hard. In the world of hardware, however, they are often grouped with servers, as they typically use hardware common to their data-serving cousins.

So you're looking to build a server, or workstation? Well let's start on the outside, since the case size and shape is going to be based upon the intended purpose of the server, and will affect choices on server boards and CPUs available. 

Caseology

Rackmount or Tower?

Servers essentially come in two shapes - Tower and Rackmount.

A tower is just like a desktop tower, an upright box in which the motherboard sits vertically, only bigger. The towers need to be taller and deeper to fit the larger board form-factors required on certain server-boards. Workstations are always towers.

Rackmount servers are thin boxes that are stacked horizontally in a vertical frame, or rack, such that space can be efficiently organized when running many servers in a data-center.

Your choice of rackmount or tower casing is dependent simply on how many server PCs you need. If you have only a single file server to build, you'd go for a tower. If, on the other hand, you want to install a DNS server and 10 web servers, it makes sense to buy a rack and then mount the 11 boxes onto the rack, thus saving on space. It is not uncommon amongst ISPs or large websites such as Google or Amazon, for there to be thousands of racked servers housed in super-air-conditioned data-centers.

What are U talking about?

Rackmount server cases are specified in "U's". I believe it is short for "units" and refers to a standard for measuring the height of the case, when fitted into a rack. A 1U server, for instance, is very thin - only 1.7" high, and so you can stack a large number of individual PCs in a relatively small space. 2U is 3.4" - exactly double the height of 1U, whilst 3U, 4U and 5U are respectively 3, 4 and 5 times the height of 1U.

The choice of which U to go for depends upon what you're putting in it. Generally speaking, the higher the processing power, the bigger the U required. Entry-level servers, typically built using Pentium 4 CPUs, for instance, will operate well in a 1U case, whereas it would be difficult to run a quad-Opteron solution in anything less than a 3U.

The main reason for this is cooling. Because of the restrictions on space, particularly height, traditional means of cooling a CPU (i.e. "stick a fan on top") go out the window. Instead the entire case must act as a duct through which air can pass over the CPU and heatsink. So this needs to be borne in mind - when you go 1U, you can only have a passive heatsink (i.e. no fan), and must rely upon the fans within the case providing sufficient cfms (cubic feet per minute) passing through. Now can you see why these data centers need super air-conditioning?

Serverboard Form Factors

Desktop PC builders are used to just one standard, ATX (ok, two if you include Micro-ATX). But building a server or workstation can bring about a bewildering set of form-factors to consider, when selecting cases and fitting motherboards inside them.

ATX - 12" x 9.5" - is also used in this arena, mainly in the entry-level server and workstation market. ATX server boards differ from their desktop cousins in that the CPU and DIMM slots have been shuffled about to improve on airflow, particularly suitable in rackmount servers. ATX can be somewhat limited, however, if you want more gear on your serverboard, especially when incorporating more than one CPU. Hence, the need for larger form-factors...

EATX (Extended ATX) - 12" x 13" - provides additional depth to the board, to allow more slots and chips to be mounted onboard when using two processors, enabling more serious server specifications.

SSI-EEB - 12" x 13" - same size as EATX but with different mounting holes. Cases will very often be designed to fit either EATX or SSI-EEB motherboards.

SSI-MEB - 16" x 13" - Together with the proprietory Serverworks SWTX form-factor (16.3" x 13"), this form-factor is for Quad-CPU solutions. Special cases are required for these boards. 

The following diagram helps get the different form factors into perspective:

Server CPUs

Intel Xeon vs. AMD Opteron

Servers and workstations are powered by CPUs of a different breed to desktop computers. Both Intel and AMD have developed specialist chips designed specifically for server and workstation use.

Intel released the Xeon™ processor a number of years ago. Initially running at a front-side bus of 200MHz, this was later increased to 400, 533 and now 800MHz. The key difference between the Pentium desktop CPUs and the Xeons was the ability to run them in pairs or even quads, so you could power a machine with 2 or more processors. There were other differences between the Pentium and Xeon, including cache sizes (which were usually bigger on the Xeon), but the key was enabling multi-processor usage.

Last year AMD launched the Opteron™ processor range for workstations and CPUs. The main difference between the Opteron and Xeon is that the Opteron can operate in a 64-bit environment as well as the standard 32-bit. This can have an advantage in very large data systems, such as large databases and CAD programs. The Opteron also has it's MCH (Memory Controller Hub or "Northbridge") chipset on the chip itself, which can allegedly make for more efficient RAM management.

The choice of going for Xeon or Opteron is often simply down to personal choice, since the cost-to-performance ratios are about the same. There is a wider choice of boards available for the Xeon (although this choice can be bewildering, especially when you consider that certain Xeons will only work with certain boards), whilst the Opteron can run 64-bit applications. My exclusive CPU Matrix should help you orient yourself in your choice.

Xeon Generations

The Xeon has a relatively long history now, and so one should be aware of the different generations of Xeons currently on sale, since particular generations of Xeons will only work on certain motherboards.

The new 800MHz 604-pin Xeon "Nocona" CPU incorporates many of the advances made in the desktop market, such as the 90nm process ("Prescott" in the desktop world), an 800 MHz FSB, and ability to run DDR-2 memory and support the PCI-Express transport. The chip is available in clock speeds of between 2.8GHz and 3.6GHz, and with a standard 1MB cache. It is priced competitively and is specifically designed for workstations. The chip will only operate on boards featuring the new Intel E7525 "Tumwater" chipset.

The 533MHz 604-pin Xeon is the current main-selling chip targetted at servers, and is supported by the widest range of motherboards currently on sale. Boards featuring the Intel E7505 and E7501 chipsets will run this chip. The CPU is available in speeds of between 2.4GHz and 3.2GHz, with cache sizes between 512K and 2MB.

The 400MHz 603-pin Xeon is still available new in speeds between 2.4GHz and 3.0GHz, and is designed to run on boards using the Intel E7500 and 860 chipsets.

Decyphering Opteron Numbering

AMD Opterons all use the same 940-pin configuration, so any Opteron will run on any Opteron board. However, if you're looking to use two or more CPUs on the same board, then not all Opterons are up to the task. You need to know about Opteron numbering.

AMD market the Opteron CPU using a three digit number - e.g. Opteron 248. The last two numbers relate to the overall performance of the chip (higher is faster), and is an index of a number of factors including clock speed and cache size.

The first number determines how many Opterons of this type may be used on the same board. Currently, Opterons come in three series, 100s, 200s and 800s. 100 Series Opterons may only be used in solo, 200s can be used in solo or dual processor scenarios, whilst 800s can be used in solo, dual, quad or eight-way processor scenarios. Because the higher series Opterons are more expensive for the same performance, it makes sense if specifying a uni-processor system to use a 146 series Opteron instead of a 246, and if specifying a dual-processor system to use two 244s instead of two 844s.

Pentium 4 Processors in Servers and Workstations

Whilst the mainstream server and workstation markets features the specialized Xeon and Opteron CPUs, the good old Pentium 4 finds a home in the entry-level market, where racks of 1U cheap uni-processor P4 systems can prove cost-effective, particularly for hosting companies. You can find more on the Pentium 4 processor on my guide to processors. 

Server and Workstation Boards

The board is where it all comes together, so having an idea of what case and which CPU you intend to go for is important before determining the motherboard.

To help in the decision-making process, the famous and exclusive Electronics Nexus motherboard-matrix has returned, this time for server and workstation boards. The matrix is first divided into the three processors used - Xeon, Opteron and Pentium 4. Across each table is the board form-factor, and what defining features come with the board. Down each table is the chipset used.

The main board manufacturers that we recommend and sell are Intel, Tyan and Supermicro. Whilst Intel are respected for solid reliability, Tyan and Supermicro offer equal performance for an often better price, and are equally respected as they specialize in boards for servers and workstations.

Other Considerations

Storage Devices

The high availability and high-performance demands of servers require additional levels of storage specification. SCSI (Small Computer System Interface) has been a higher-spec equivalent of ATA for many years now, providing high-speed disk access, and at a high price.

With the advent of U320 SCSI ("Ultra" 320MB/s), SCSI continues to be nearly twice as fast as the very latest incarnation of ATA - Serial ATA (SATA), which clocks in at 150MB/s. This performance does come at a price, however, as not only are the SCSI hard drives pricey - to operate such drives you need a SCSI controller. This is the principal defining feature that board manufacturers use, when offering variants of the same board. One is a board with a SATA controller, the other comes with an onboard SCSI controller.

On the high-availability side, virtually all server boards on the market today come with on-board RAID controllers. RAID is short for Redundant Array of Independent Disks, and is a controller that allows two or more physical disks to appear as a single drive. RAID can configured to "stripe" across two disks (i.e. write across both disks at once) to improve performance (RAID Configuration "0" or simply "RAID 0"), or can be used to "mirror" the contents of one drive simultaneously onto the other ("RAID 1"). Mirroring is often of more concern in a server, since when a drive is mirrored, if one drive should fail, the other drive will continue with the back-up image of the first drive, as if nothing had happened. Some controllers also allow "RAID 10" or "RAID 5" (both mirroring and striping) for maximum performance together with back-up availability. Such a configuration would, of course, need four disks (two striped, with two mirroring the first two).

Server Memory

My guide to memory will tell you everything you need to know about RAM. The only point to mention here is that, because of the need to maintain integrity of data, server memory is typically ECC (Error Checking and Control) and Registered. Always check with the motherboard to ensure that the memory specification is compatible.

Separating PCI-X from PCI-Express

Until looking into this further, you'd be pretty confused, like I once was, by the terms PCI-X and PCI-Express, and may even thought they were the same thing. Whilst similar in concept, they are in fact different evolutions of the now rather aged PCI (Peripheral Component Interconnect) transport architecture.

PCI-X is short for PCI eXtended, and was developed to extend the 133Mbps limitation of PCI to up 2Gbps, for server applications that needed high-bandwidth peripheral connections, such as Gigabit LAN cards. There are two version of PCI-X - 1.0 (which tops out at 133MHz) and 2.0 (533MHz maximum).

PCI-Express is the next generation PCI architecture principally intended to replace AGP (the Advanced Graphics Port) in both desktop and server PCs, but whose application is limitless to improving the speeds of interconnection across all devices. PCI-Express has a theoretical bandwidth limit of around 128Gbps!

There is a great article on PCI, PCI-X and PCI-Express to be found at processor.com.