The other major hardware area to look at when upgrading a network is storage, starting with disks attached direct to file servers. SCSI is seen as the technology to go for here, with the latest Ultra320 products, for example, able to feed a server with up to 320MB of data per second.

SCSI also supports more devices per controller, which is important when it comes to storage arrays (discussed shortly), plus it's a very robust and reliable technology.

However, recent advances in the capacity and speed of ATA drives and the introduction of Serial ATA (Sata) products has seen some of these advantages eroded, especially at the low end, where the cost differential of SCSI can be a major factor.

Upgrading from ATA disks to SCSI is an expensive option, but a switch to Sata can be a cost-effective way of boosting performance. A new storage controller and replacement drives are both required, but Sata disks are available with capacities beyond 100GB and with spin speeds of 10,000rpm.

And that's important, as spin speed is one of the biggest determinants of performance. Of course, you'll need to take a full backup and recover data to the new disks, which takes time, and the disruption involved needs to be factored into the cost.

The same applies if you're considering upgrading to a storage array, which is another way of enhancing both the performance and availability of storage subsystems.

Array for reliability
With a storage array data is 'striped' across multiple disk drives; these can be written to and read from in parallel, leading to faster throughput compared to serial access when the data is all on the one drive.

There are several ways of configuring a disk array, some giving extra performance, some greater redundancy, and some both. The different configurations are known as Raid (Redundant Array of Inexpensive Disks) levels, with the most common implementations being Raid 0 and 5.

With a Raid 0 array the data is simply striped across the disks as just described. No special hardware is required, with software support provided in most network operating systems.

Both SCSI and ATA disks can be used (even mixing them together), making this a good choice if you're looking for increased storage performance at low cost. However, there's no redundancy in a Raid 0 array and if a disk in the stripe set fails all the data on the logical volume involved will be lost.

With a Raid 5 array extra error correction information is interleaved with the data, and the information held on a failed disk recreated on the fly by the controlling Raid software until a replacement is fitted.

On the downside, Raid 5 doesn't necessarily improve performance. Our results of tests run on a file sharing server with a hardware Raid controller fitted showed that, when configured with a Raid 0 stripe set there was a small (7.5 per cent) increase in throughput overall, compared to accessing the disks individually (a setup referred to as JBOD - Just a Bunch of Disks).

When the same system was reconfigured for Raid 5 throughput was often worse than a JBOD setup, particularly at low loads.

Raid 5 support is also available in some operating systems, including Windows, although if it's performance you're after a specialised Raid controller is best. You'll definitely need one to hot-swap disks without powering down the server, and a Raid controller will reduce the drain on the server CPU, provide extra caching facilities and, in some cases, provide battery backup to maintain cache contents should power be lost.

The cost and effort required to add Raid support varies. Some vendors, for example, integrate the necessary circuitry onto their server motherboards, requiring just a small hardware key to make it available. Most, though, require a plug-in adaptor and a special drive cage in the server, or an external drive housing, if you want to be able to hot-swap disks without powering down.

In terms of technology, it's possible to create a storage array using both ATA and SCSI disks. However, even with Sata drives you're limited to the number of disks that can be used and they have to be located inside the server, not externally.

For anything other than a small network SCSI is the best option, but you can spend a lot of money for a small performance boost. For companies serious about improving storage throughput the answer is a San.

With a San, storage is attached not to individual servers but to a separate storage-only network, to which the servers are also connected. Expensive Fibre Channel hardware has traditionally been used for this, but in the past year or so IP-based Sans using ordinary Ethernet switches have been introduced, bringing down the costs significantly.

Software matters too
The OS and other software used on servers has an impact on performance. Upgrading to the latest versions is tempting but it can also be problematic and time consuming. It may be worth doing if the result is significantly faster file sharing or more responsive applications.

The new Windows Server 2003 release is a good case in point as, along with enhancements in terms of security and functionality, it's widely hailed as the best-performing version of Windows ever and Microsoft is understandably keen for customers to upgrade, especially those still running NT (around 40 per cent of the installed server base).

Of course there's a cost associated with this, but the good news is that Windows Server 2003 really is a lot quicker than both NT and 2000. You don't necessarily have to upgrade server hardware to get the benefit, either.

There are minimum requirements for the different editions of the new software; and we tested the kind of effect migration to Windows Server 2003 can have on a modest and aging host system.

At low client loads there was little difference, so on a small network the upgrade might not be worth it. But as the number of users rises the difference in throughput is real and appreciable. Of course, you don't just get better throughput; there are numerous other file and print sharing enhancements to tempt existing customers to migrate to the new operating system.

Availability, for instance, is enhanced with a new Automated System Recovery (ASR) option, together with a hugely useful Shadow Copy facility to take point-in-time snapshots of important data volumes. Third-party backup programs have long provided this service, but it's now built into the operating system, enabling open file backups to be taken without the need for special backup agents.

The Shadow Copy feature also allows users to find previous versions of files themselves, using Windows Explorer.

There's also a San boot feature and enhancements to the Distributed File System (DFS), the software which allows a single logical file system to be created across multiple servers.

There are benefits, too, when it comes to running applications on the server, starting with better scaling of performance as additional processors are added. There's also support for more memory. Additionally, the IIS 6.0 web server, which is included as part of the new operating system, features a revised request processing architecture designed to make it both more robust and powerful than before.

Quantifying the effect of these and other enhancements depends on the applications involved but, to give you an idea of the kind of improvement you might expect, we ran some tests on a two-way SMP server running a simple web-based application.

Upgrading from Windows 2000 Advanced Server to Windows Server 2003 provides an immediate increase in the number of requests that can be handled, both with one and two processors installed. When going from one to two processors the performance gain is significantly greater when using the Windows Server 2003 software.

Another key change with IIS 6.0 is the ability to run websites and applications using their own self-contained worker processes, preventing them from interfering or stopping other sites and programs. The main web server code and associated web administration service, similarly, have their own dedicated processes, making it harder for rogue applications to bring the server down.

The Active Directory Service is also enhanced in Windows Server 2003, and there are numerous security features, including the ability to lock down the configuration to reduce the vulnerability of servers connected to the internet.

Such improvements, together with the file sharing and application performance enhancements, can be used to make a strong case for upgrading to Windows Server 2003.

They also make it easier to do what many companies now want, which is to consolidate operations onto a smaller number of platforms. That's even easier when migration to the new software is combined with the kind of hardware upgrades covered in this feature.

Balancing the load
So much, then, for what you can do to improve network performance and availability with upgrades to server hardware and software. However, there is one other way of achieving those same goals - by spreading the processing load over two or more servers in a cluster.

Unfortunately special clustering software (and hardware) is needed where file sharing and traditional client/server applications are concerned, making this a complex and expensive option. But that doesn't apply to web server clustering, which can be both simple and affordable.

All that's required for this kind of load balancing clustering is a means of intercepting client requests and splitting them between the individual servers (nodes) that make up the cluster.

The cluster as a whole can then be given a single virtual name and IP address, just as though it were one system, and client requests redirected to member nodes using a variety of algorithms. The simplest of these, known as round-robin, balances loads by sending each request to the next server in turn.

Alternatively, it's possible to direct traffic to the least busy server, to the one with the least number of active sessions, and so on. Performance can be scaled simply by adding more nodes to the cluster, while availability is enhanced because, should a node fail for any reason, the others can carry on regardless.

There are two ways of implementing this, one of which is through software on the servers themselves. Windows 2000 Server and Windows Server 2003 both ship with load-balancing utilities, as do most Linux implementations, and there are several third-party products around too.

The other approach is to build a load-balancing cluster using an external switch (such as the one pictured above) - not a regular Layer 2 product, but one capable of making routing decisions at Layer 4 and above. The switch is positioned between the cluster and its connection to the local network or outside world and routes request at wire speed for maximum throughput.

As with all hardware add-ons, load-balancing switches can be expensive, but they do have their advantages over the software approach. For example, a switch can handle a mix of platforms running a variety of applications, with no need for additional software on the servers involved.

Many also include mechanisms to handle the replication of data, although most companies deal with this by having a single database behind the front-end web cluster.

Another advantage is the ability to build in extra redundancy - anything from spare power supplies to complete duplication of the switch. It's also possible to do a lot more than just balance on IP address. Some products can switch on URL, for example, and type of application, to more intelligently balance client loads based on the content requested.

However, this kind of balancing can be complex to configure, with no guarantee that the end result will be worth it. It's important also to look for 'persistence' features, to make sure existing sessions aren't switched to a new server mid-way through a transaction, for example.

Adding strings to bows
Finally, it's important to emphasise the need for a multi-pronged approach to any network expansion or upgrade. As we've tried to illustrate throughout this feature, changing just one component will only bring limited benefits unless other parts of the server and supporting network are also upgraded.

To gain the benefit of extra Lan adaptors, for example, faster or more processors are also needed. More than that, you will often find that improving one part of an infrastructure does little more than highlight weaknesses elsewhere, making careful planning and preparatory testing vital.

By the same token you shouldn't expect cumulative improvements when several components are upgraded at the same time - the results can vary considerably. In our tests we made several of our suggested changes to an old PIII-based server.

Among the upgrades we included were a second processor, memory expansion from 128MB to 1GB, swapping of the standard 10/100Mbps Lan adaptor for a Gigabit Ethernet card and a second disk with the two drives configured as a Raid 0 array.

We expected and achieved good results with each of these changes on their own, but when they were all applied together file sharing performance only improved by just over 30 per cent overall.

That was something of a disappointment, but understandable given that we couldn't change things like the type of processor, bus speeds and chipset.

That said, the overall improvement was still worth it and could be improved upon using higher specification hardware as a starting point. Indeed similar gains, and more, can be achieved by anyone looking to improve and expand their network by following the procedures discussed in this feature.