Building up your company servers

Servers are generally categorised as belonging to one of three broad groups: entry-level, departmental/workgroup and enterprise. What one business may call a workgroup server, another will use to run the entire business. Similarly, entry-level systems can be used to support a departmental network in a large enterprise or run everything in a small business.

Generally, the smallest single-processor servers tend to fall firmly into the entry-level category, with prices to match - in some cases selling for little more than most desktop PCs. However, the common features found in larger servers, such as built-in management and fast storage, are often included.

Most servers are shared resources, holding files and running applications used by hundreds, sometimes thousands, of people in an organisation. The hardware needs to be powerful enough for these demands and also has to be highly available.

Distinguishing between types of server
There is a lot of overlap between server types and different manufacturers interpret the terms in different ways, which does not help customers when it comes to making comparisons. Another confusing factor is the emergence of the specialised server appliance.

The server appliance is designed to support just one or two specific applications, such as the running of an email or web server, enabling what are relatively simple platforms to support large numbers of users. Some vendors are also supplying storage through building network attached storage appliances.

As a result, server categories are defined as much on the levels of fault tolerance as they are on performance and capacity - even with entry-level servers which can be very hard to distinguish from highly configured desktops.

Entry-level servers
The processors of entry-level servers are likely to be the same as those of high-end desktop PCs with dual-processor configurations just as likely to be found in the workstations as among smaller server systems.

Memory configurations are also similar, although virtually all servers now have some kind of error checking and correction support, and can be fitted with more RAM than the average desktop.

Entry-level systems are designed to support only small networks. To vendors that typically means less than a dozen users. File and printer sharing is the main requirement, and the disk subsystem is likely to be only lightly stressed, with simple SCSI interfaces providing the necessary performance and expansion capabilities. Internal storage will normally be fixed rather than hot-pluggable, with duplexing rather than Raid to provide redundancy.

Look for management features similar to those found on high-end desktops, such as temperature, voltage and other environmental sensors, although with a server more sophisticated alerting software should be included. But for anything other than a small Lan, most entry-level servers are of limited utility, particularly if you need to support shared applications.

Workgroup servers
'Workgroup' and 'departmental' are terms that are often used interchangeably and cover a wide range of potential uses. Distinctions between these types of server are often blurred. Individual workgroup servers can be found running small firms, while multiple servers can be found supporting single workgroups or departments inside larger businesses.

However, workgroup servers are more powerful than entry-level systems, offer greater levels of redundancy and, depending on the model, should be configured to scale to support increasingly large numbers of users.

Enterprise servers
At the highest level of performance the term 'enterprise server' is used, primarily to identify systems aimed at large companies and ISPs looking for performance and high availability.

Enterprise server performance is mainly the result of more and faster processors. Highly scalable Risc-based enterprise servers have long been available with eight, 12 or more processors, but with the Profusion chipset, even Intel vendors can build eight-way systems.

Raid is standard on this type of high-end server, boosting both performance and availability, with a choice of Fibre Channel or SCSI as the interface. Interestingly, the current implementations of Fibre Channel do not offer much in terms of throughput gains, at least not with conventional server-bound storage.

Scaling workgroup server performance
Increases in the performance of workgroup servers come in two varieties. First, there are more powerful processors designed specifically for server use, such as Risc processors and Intel's Xeon processor for servers, which has several advantages over its desktop counterpart.

The basic computing configuration is much the same, but there are large differences in terms of memory support, with up to 2Mb of second-level cache available on some Xeon processors. This can greatly increase throughput for transaction processing and other common server applications. The amount of addressable main memory is also extended, with support for 4Gb and above on motherboards with the necessary hardware to hold that amount.

The second way of boosting performance is to have more than one processor. In this respect, server processors, such as the Xeon, can also be optimised for use in symmetrical multi-processing (SMP) machines.

Server-specific chipsets can better co-ordinate the activity of the processors and enhance shared access to memory, with specialised controllers to handle the larger address spaces. Such systems can then be equipped with two, four or eight processors - more on certain Risc platforms - with processing tasks divided up by the operating system and shared equally across all the chips to provide scalable performance beyond that possible with just a single processor.

Match the processor to the application
Select a processor fast enough to support your particular network. If the server is only for basic file and printer sharing, the server processor is unlikely to be a limiting factor on performance, and you can get away with what might otherwise appear to be a relatively slow processor.

Often entry-level servers with just a single 200MHz Pentium processor can easily support quite large workgroup Lans. However, when running applications, such as a database engine, or an email or web server, buy the fastest processor possible.

Going down the SMP route
For servers running one or more applications, you may want to consider a SMP system with two or more processors on board. Network operating systems, including Netware, Windows NT and most Unix implementations (including Linux), support multi-processing as standard, allowing server tasks to be shared across processors to improve performance.

Be careful when deciding how many processors are needed, since the gain from adding any more may be negligible. If the server is only used for sharing files and printers, there's likely to be very little load on the processor, so adding chips will have no effect. In this case the only advantage may come from having a spare processor, with some servers able to re-boot on just the one in the event of chip failure.

Server clustering for workgroups
Server clustering is not required by every workgroup. Clustering links servers so that, theoretically, processing loads can be shared across servers as well as processors, making for massively scalable solutions.

But very specialised software is needed to support this kind of clustering, with the most widely available implementation, Microsoft Cluster Server only providing for failover of a two-server cluster at present. However, clustering is an expensive solution, and would only be needed to support large mission-critical applications.

Consider availability features
Some server features help minimise problems, tell you when they're about to occur and deal with them when they do. These can include equipment such as redundant power supplies and backup cooling fans - some big enterprise servers ship with multiple power units and huge arrays of fans.

Some vendors also offer hard disk drives that let you know when they're about to fail, an option that is called either predictive failure analysis or Self-Monitoring Analysis and Reporting Technology.

At the very least, a server should have built-in monitors of internal temperature, voltage and fan activity, and software to communicate this information to the system administrator. These features are built into the latest motherboards on low-end machines, but only enabled on higher spec machines.

The ability to both monitor and manage the server remotely over the Lan is another essential. Also, out-of-band management enables you to dial into the server and manage it when the rest of the Lan is down, with the option of having the server dial out to issue pager or email messages when a failure occurs.

Additional processors can also improve server availability, and some systems can either tolerate the loss of a processor or automatically reboot with the failed processor disabled. Moreover, server-specific processors will have their own monitoring circuitry built-in. For example, a separate system management bus on the Xeon helps to maintain availability.

Storage and redundancy features
Storage is also beefed up with workgroup servers - more room for disks and hot-swap backplanes are almost standard features. SCSI remains the preferred interface, with the necessary controllers integrated onto the motherboard. In many cases, support for more than one SCSI channel is offered, making for easy duplexing.

Raid support is also common. It is standard equipment in most four-way systems and is increasingly offered as an option on the smaller dual-processor platforms, with some manufacturers now supporting Raid as well as SCSI on the motherboard. At the top of the range, support for 64bit PCI and hot-swapping of adapters is the norm.

Connecting servers to the Lan
A network interface card (Nic) will be needed connect your server to the Lan. Some manufacturers integrate this onto the motherboard, but usually on entry-level systems only, and most use plug-in Lan adapters. Always opt for the best quality branded cards: 3Com and Intel are leaders in this area, with Adaptec and Accton-SMC also worth considering.

The type of Nic you need depends on the network. On a Token Ring Lan you need a 16Mbps adapter, although you can get Fast Token Ring Nics which support 100Mbps connectivity. This is well worth having as it provides a much larger pipe to the network and it?s much the same on more popular Ethernet Lans where 100Mbps Fast Ethernet has been available for some time. Even if you don't need 100Mbps, you should opt for a dual-speed 10/100Mbps adapter (for Token Ring, 16/100Mbps), because they are relatively cheap and make upgrading simple.

Choosing a server OS
There are three major products to choose between: Netware from Novell, Microsoft Windows NT and Unix - the latter coming in a number of variants, including Linux. Of these, Netware is the frontrunner when it comes to file and print servers. However, Netware doesn't make a very good application platform. Applications have been ported as Netware Loadable Module, but few developers have bothered as this remains difficult to code. Instead, Unix has become the predominant application platform, with Linux widespread as an internet operating system.

Microsoft has made it a lot easier to develop for the NT platform, and there's no shortage of expertise with so many Windows desktops and Windows applications around.

Administration and management of servers
A Windows server with its user-friendly desktop front end is a lot easier to administer and manage than either Unix or Netware, particularly now NT Server is the most popular operating system for new server implementations.

Netware 5.x offers both a Windows-like graphical user interface and Java application support. The Novell software also supports shared-processing clusters, but Microsoft released Windows 2000 last month with better security and improved file and print services.