Scaling Out

---

Since the 1980s, computing environments have tended to become more distributed. Increasingly agile, powerful, and affordable low-end and mid-range servers became available. Easy to deploy and configure, such servers are now an established part of the IT landscape. As servers become overloaded or databases grow, additional servers can be added, or multiple functions residing on one server can be distributed over several others. This is known as scaling out or horizontal scaling. The purpose of scaling out is to improve performance (e.g., Web server farms) and availability (e.g., high-availability clusters) or to provide an immediate fix to a pressing IT situation. When it comes to increasing storage capacity, for example, scaling out offers a rapid and often cost-effective solution. Increasingly flexible, powerful, and affordable low-end servers have resulted in a scaling out in the data center to a more distributed computing model.

On the downside, though, an overabundance of small servers can create too much of an administrative burden, especially when hundreds of servers are present. Every server added to the data center requires space, power, telecommunications, and networking interfaces, storage capacity and connections, backup and restore systems, and an update to asset management inventories and supplier contracts. When taken to extremes, scaling out can result in fragmented databases and a series of disconnected, difficult-to-manage servers. Some servers end up underutilized while others are badly overloaded. In many cases, the proliferation of distributed servers within large enterprise organizations has proved to be an administrative money pit as well as a nightmare. As more systems are added to the network, the costs of administration and maintenance increase.

Horizontally scaled systems can be further characterized by multiple servers connected by relatively slow communication links. Although networks and cluster interconnects continue to improve performance in terms of high bandwidth/low latency, their speed is still dwarfed by the capabilities of bus and switch interconnects. As scaled-out systems communicate by message passing, the elapsed time to complete a task can sometimes increase markedly.

Additionally, each horizontally scaled server runs a copy of the operating system. This can result in synchronization challenges, not to mention trouble with the division of labor among servers. While the mantra in most enterprises is "one application, one server," distributed environments still tend to offer much coarser granularity when it comes to server management and workload planning. Costs, too, can mount up. Licensing costs, for example, for running hundreds or even thousands of small servers can be prohibitive, and managing these assets can turn into an administrative nightmare (see Chapters 16 and 17).

---