Databases Reference
In-Depth Information
false sharing,
7
because each data node will use its CPUs independently of the
others.
6.9
Where the Money Goes
Shared-nothing partitioning remains most useful for complex query processing, which
is a hallmark of data warehousing and decision support systems. These systems are usu-
ally expensive. After reading this chapter you may think the reason for the expense is the
cost of all the components. In fact the dominant expense in most warehousing systems
is storage, which is always the most expensive component in a warehouse. In fact the
cost of storage is itself very cheap. In Fall 2006 multiple companies began selling ter-
abyte-sized drives for only $500 U.S. However, these high-volume inexpensive drives
are slow and not suitable for data warehousing applications. While raw storage is cheap,
performance and reliability are expensive. High-performing parallel access RAID stor-
age remains expensive per GB, which drives the cost of storage for enterprise database
systems. Chapter 13 discusses storage requirements, striping, RAID, SAN, and NAS in
more detail.
6.10
Grid Computing
Another form of shared-nothing architecture focuses on a widely distributed model for
data analysis, known as Grid computing. Advances in Grid computing are offering
alternatives for specific classes of problems. Grid computing is emerging as another
model for massive computational processing. It is not yet a well-established industrial
platform. Grid computing exploits “function shipping,” whereby distinct problems are
shipped to any number of participating computers on a network. Each node on the
Grid (a participating computer) computes a solution for the problem it receives. This
allows a massive number of computers to collectively solve pieces of a problem with lit-
tle communication between the systems. The most well-known example of Grid com-
puting is the Internet-connected computers in the Search for Extraterrestrial Intelli-
gence (SETI).
With SETI, Internet users can participate by running a program that downloads
and analyzes radio telescope data. While participants are not using their computer for
personal use the computer's available processing cycles can be used to help in the search
for otherworldly life. While shared-nothing architectures scale to hundreds of nodes,
Grid computing can scale to tens of thousands if not millions of nodes. Industrial Grid-
based solutions are starting to emerge. IBM reported that the Institute of Plant Genetics
7
False sharing is described in Chapter 13.
