Databases Reference
In-Depth Information
small degree of business logic (the degree varies widely by application) and coupling
with the application servers. One of the well-known architectures used for three-tier cli-
ent server processing is CORBA, or Common Object Request Broker Architecture.
13.3
Symmetric Multiprocessors and NUMA
13.3.1 Symmetric Multiprocessors and NUMA
The mid-1990s saw the rise of servers with multiple CPUs, known as multiprocessors or
symmetric multiprocessors (SMP). These servers included between two and eight
CPUs. Typically only one of these CPUs can access memory at the same time. To avoid
the serialization of memory access, the CPU caches become increasingly important for
multiprocessors and increasingly sophisticated algorithms are used to avoid caches
misses on the CPU caches as access to local RAM (memory not in the CPU cache),
because the weak link in the performance characteristic becomes increasingly the access
to RAM. The serialization of memory access on a symmetric multiprocessor increases
this problem significantly. As a result, SMPs rarely have more than eight CPUs. As a
result the next generation of multiprocessors was developed in the mid- to late 1990s
called nonuniform memory access (NUMA).
NUMA processors divide CPUs into groups called “quads,” each with their own
local bus and RAM. Because each quad has its own local RAM and bus, the memory
access and bus traffic are no longer necessarily balanced, hence the term “nonuniform
memory access.” NUMA architectures can scale to much larger numbers of CPUs,
often 32 or 64 units. However, maintaining memory consistency (cache coherence)
between the quads becomes complex, and NUMA systems remain an expensive option.
While NUMA systems can scale to a larger number of CPUs than regular SMPs, they
cannot scale as large as shared-nothing or Grid systems (discussed in Chapter 6). Even
so, there is one tremendous advantage of SMP and NUMA system-based databases not
offered by Grid and shared-nothing database architectures: lower administrative costs.
While the capital expense to purchase a large NUMA system is significantly higher than
purchasing multiple smaller single CPU systems and connecting them in a Grid or
shared-nothing MPP, the result of Grid and MPP architectures is a network with multi-
ple physical systems that each need to be administered and attended to. In contrast, a
large NUMA system is usually more expensive to purchase, but the multiple CPUs it
contains are housed within a single server—one box to manage.
Dmitri Besedin published a study of SMP and NUMA memory access bandwidth
for a few different architectures and concurrency models on his website. His results
from his August 9, 2005, study are repeated here in Table 13.1.
