Databases Reference
In-Depth Information
way queries can be answered in a fraction of the time. That's the concept of shared-
nothing partitioning, and we'll see it has distinct benefits and design challenges.
Interestingly, shared-nothing partitioning is more like political divide and conquer
than the usual computer science version because it is nonrecursive.
6.1
Understanding Shared-nothing Partitioning
6.1.1 Shared-nothing Architecture
The two extremes of parallel-processing architecture are shared everything and shared
nothing (Figure 6.1). The terms mean pretty much what they imply. With shared every-
thing, a single computer solves complex processing problems using shared memory,
shared disk, and a shared bank of CPUs. This class of machine is typically called a sym-
metric multiprocessor (SMP) or nonuniform memory architecture (NUMA). NUMA
systems are a special subset of SMP systems. On the other extreme, shared-nothing sys-
tems use a set of relatively independent servers to work cooperatively on subsets of a
problem. Occasionally these servers will need to share data, which they will pass
between them over a high-speed interconnect. Shared-nothing systems have become
famous for their ability to scale out to a very large number of processors and servers,
which has earned the high end of this domain the name massively parallel processing
(MPP).
There are three major products that offer shared-nothing architecture today: DB2's
Data Partitioning Facility, which allows users to purchase commercially available com-
Figure 6.1
Shared everything versus shared nothing.
ponents, servers, disks, and network interconnects to construct a shared-nothing sys-
tem; Informix Extended Parallel Server (XPS)
1
is similarly based on commodity compo-
nents; and NCR Teradata, which uses commodity disk servers with proprietary hard-
