HARDWARE: Server farms
The idea of using multiple low-cost servers with a load-sharing front end has caught on for Web farms and application servers, but can it be extended to other areas?
Kevin McIsaac, program director of server infrastructure strategies at META Group believes it can.
This approach is relatively easy to apply to Web servers as they are essentially stateless--a browser requests some data, the server provides it, and the task is over. As soon as transactions are involved, states are maintained in the back-end database.
Typically, the database runs on high-end symmetrical multiprocessing (SMP) servers that are clustered in pairs. This is expensive, and the second server is generally a hot standby.
But the arrival of Oracle Real Application Clusters (RAC) means "in principle . . . we'll be able to have multiple nodes for the back-end database server running in a cluster with a shared disk, and it will look like a single machine," says McIsaac.
The total bill takes a double whammy with this arrangement. Firstly, cheaper hardware can be used--a pair of eight-way SMP servers is cheaper than one 16-way server.
Secondly, less backup hardware is needed: three eight-way servers could replace two 16-way units for example, and this will give better performance as the spare can be active at all times. Clusters of three to five servers with shared storage will strike "a good balance between the additional availability and complexity," says McIsaac. This type of clustering can reduce costs by 25-40 percent and "I think [that] will get some real interest from people".
McIsaac says this analysis assumes RAC is ready for "prime time"--he thinks it will work in real-life implementations, but is waiting for proof. The concept has been around for years in the mainframe and minicomputer worlds, but Oracle is now working with Red Hat and Dell. He suggests this relatively cheap approach could have a major impact in 12-18 months.
Where server farms are used, it can still be important to get a failed unit back in commission quickly to maintain good overall performance and as a precaution against an outage caused by another unit failing. Blade servers are becoming popular as they offer high rack densities. When an individual blade fails in an HP server, the "rip and replace" feature means its replacement is automatically reconfigured, says Lynch.
