Building a scalable network

Jan 12, 2000
Tim Landgrave

When trying to build a data center that can support hundreds or thousands of users, the first issue most companies face is scalability. ASPs have to deal not only with multiple users, but multiple users from multiple companies.

You want to be sure that if you use the services of an ASP, then your data is secure and your performance won't diminish as the ASP adds additional customers. But even if you don't use an external hosting provider, you need to consider how to re-architect your internal systems to allow you to "self-host" your applications.

In order to be successful, you have to consider how you will accomplish the deployment of a scalable system.

Part two in a four-part series

In part one of this series, "Why Every CIO Should Care About Hosting," we examined how to make informed decisions about outsourcing. Check back to find out more information on hosting in the enterprise.

If bigger is better, then should you be taller or wider?
Simply defined, scalability is the ability to add users to a system without impacting the performance of the existing users. Strictly speaking, you're not building a scalable system by adding different sets of computers for each set of users you want to support. But with the current state of software, hardware, and applications, this is sometimes the most cost-effective choice.

If you're building for the long term, however, you have to decide to manage scalability in one of two fundamental ways—with hardware or software. Either way, you're trying to accommodate the development of "Internet Scale" systems. You should certainly assume that at some point in the future, you're going to open up your systems to your own prospects, customers, and trading partners. This assumption demands that you consider how to handle hundreds or thousands of users on the same system.

Scaling up: Build it with more hardware
With more manufacturers supplying Symmetrical Multiprocessing Systems (SMP) and with products like Microsoft Windows 2000 Data Center Edition to support them, many companies will opt for the hardware scaling approach. Let's look at the advantages and challenges of choosing a hardware scaling approach:

Advantages

There's a single system to program and manage.
Companies can consolidate multiple servers onto a single system.
There are several high-volume, 8-way systems already available, with 32-way systems only months away.

Challenges

Relying on a single system for production makes scaling down to a single system for starter or development systems more expensive.
There's a single point of failure for all applications.
Although systems with more than eight processors will be available soon, current systems may not yet be large enough to support the applications that companies need.

Ultimately, you will reach a hardware limitation to scalability that will require multiple servers. At this point, you'll need to consider the best way to manage these servers with software.

Scaling out: Rely on software to manage multiple servers
Using a software scaling methodology allows us to ride the PC economics curve. As off-the-shelf systems continue to improve in performance and reliability, it's a natural evolution for the software that manages these single systems to become more aware of the systems around them. Let's consider the advantages and challenges of implementing a software scaling approach:

Advantages

The system can be expanded in a simple and modular way by adding more servers to the server farm.
There's no single point of failure in the system.
The hardware limitation to scalability has been removed.
Linear scalability can be accomplished with predictable, incremental costs.
Development or starter systems can be easily implemented at a low cost.

Challenges

Managing multiple systems is difficult to do today.
Developers are still learning how to write systems that perform well when distributed across multiple machines. (See the side-bar on "Architecting a three-tier data center on Windows 2000" for more details.)

In order for software scaling methodology to become the norm, four major goals have to be accomplished by operating system manufacturers:

The software that manages the server farm must scale linearly with the number of machines with minimal per-server overhead from the operating system.
Server failures must be transparent to users of the system, with automatic system software and application software recovery.
Load balancing between the servers must be dynamic. That is to say, as servers are dropped or added, the existing load is spread evenly between the remaining servers.
This needs to be accomplished using standard, off-the-shelf hardware in order for the system to be affordable.

Which is the right choice?
The fact is that your solution will involve some combination of the two methods. As the cost of MIPS continues to drop and demands on performance continue to rise, you'll continue to deploy better performing hardware. But software to manage and maintain a distributed system will be essential to keeping your data center operating over the long haul. And just like the proprietary, hardware-based word processor and PBX before them, the mainframes of the future will become software-based distributed systems.

Now it's bigger, but is it any better?

Just because you can make it bigger, doesn't mean you can add users and applications and keep the whole thing running. Next week in part 3, we'll look at the issues surrounding managing the platform itself and the applications on the platform.

How have you solved your scalability issues?
Have you used a combination of software and hardware to produce a scalable network capable of handling your enterprise's growth? Share your experiences by dropping us a note.

Tim Landgrave is the founder and CEO of eAdvantage. eAdvantage assists small to medium enterprises in the development of their electronic business strategy.