Databases Reference
In-Depth Information
Memory
Processor
Processor
Processor
Processor
Processor
Disk
Disk
Disk
Disk
Disk
Disk
Disk
Disk
Disk
Disk
Disk
Disk
Disk
Disk
Disk
Disk
Disk
Disk
FIGURE 3.3
Shared-everything architecture.
Shared-everything architecture
Shared-everything architecture refers to system architecture where all resources are shared including
storage, memory, and the processer (
Figure 3.3
). The biggest disadvantage of this architecture is the
limited scalability. Two variations of shared-everything architecture are symmetric multiprocessing
(SMP) and distributed shared memory (DSM).
In the SMP architecture, all the processors share a single pool of memory for read-write access
concurrently and uniformly without latency. Sometimes this is referred to as uniform memory access
(UMA) architecture. The drawback of SMP architecture is when multiple processors are present
and share a single system bus, which results in choking of the bandwidth for simultaneous memory
access, therefore, the scalability of such system is very limited.
The DSM architecture addresses the scalability problem by providing multiple pools of memory
for processors to use. In the DSM architecture, the latency to access memory depends on the relative
distances of the processors and their dedicated memory pools. This architecture is also referred to as
nonuniform memory access (NUMA) architecture.
Both SMP and DSM architectures have been deployed for many transaction processing systems,
where the transactional data is small in size and has a short burst cycle of resource requirements. Data
warehouses have been deployed on the shared-everything architecture for many years, and due to
the intrinsic architecture limitations, the direct impact has been on cost and performance. Analytical
applications and Big Data cannot be processed on a shared-everything architecture.
Shared-nothing architecture
Shared-nothing architecture is a distributed computing architecture where multiple systems (called
nodes) are networked to form a scalable system (
Figure 3.4
). Each node has its own private memory,
