Building Scalable Analysis Services Applications - Microsoft SQL Server 2008 Analysis Services - page 449

. No single point of failure —Failure of a single computer in a distributed architecture

doesn't necessarily mean that the entire application fails. The other computers in
the system can take on the work of the failed computer.

. True scalability —You scale the application by adding new computers. With an

appropriate distributed network infrastructure, you should be able to scale your
application far beyond any scale-up solution.

OLAP Farm

A scale-out approach to increasing the capacity of online analytical processing (OLAP)
applications is known as an OLAP farm. Analysis Services provides a number of features
that enable you to build an OLAP farm, a collection of computers that work together to
support a multidimensional model. Figure 25.1 shows a diagram of the architecture of an
OLAP farm.

You build your OLAP farm around the master server, which contains metadata and
manages the whole system. The front-end servers communicate with the client computers
and execute client requests. Front-end servers are typically configured to run under a
network load-balancing cluster. A client application establishes a connection to a virtual IP
address. That connection is redirected to one of the front-end servers. The front-end server
parses and resolves the queries from the client application, and executes all the calcula-
tions defined in your cube.

Data Storage

There are two approaches to data storage in an OLAP farm:

. You can use a synchronization mechanism to have your data statically replicated

across all the front-end servers. (For information about synchronization, see Chapter
40, “Deployment Strategies.”) In this mode, each computer holds an exact replica of
the database. Any change to the database requires an explicit operation to update
the database on each server with the new data. The data on a server is not updated
until some external operation distributes the updates.

. You can use link dimensions and link measure groups to load data from the master

server to the front-end servers dynamically. In this scenario, each front-end server
checks for updates and automatically refreshes its state if changes on the master server
require an update. For a discussion of linked dimensions and linked measure groups,
see the “Linked Dimensions and Measure Groups” section later in this chapter.

To execute a query, the front-end server first determines whether the answer is already
available in its local cache. If not, with the static replication model, it retrieves data from
the local storage.

NOTE
A front-end server needs a fast CPU for making calculations and a lot of memory for
caching the results.