Databases Reference
In-Depth Information
cycles to do the same work. Instruction sets differ across architectures and as a result
some CPUs require multiple instructions to perform tasks that another CPU can per-
form in a single instruction. These combined effects result in CPUs in different archi-
tectures being comparable by clock speed within a factor of two to four times—not very
telling.
Another major consideration in CPU performance is the size and efficiency of the
CPU cache, which can dramatically reduce the time lag (latency) of accessing memory
that is in RAM. RAM memory is relatively slow to access compared to memory on the
CPU, usually by a few orders of magnitude.
With all the variables, how can the performance of CPUs be evaluated? In fact the
only reasonable way is through direct comparison, and an industry standard has grown
up around this. The Standard Performance Evaluation Corporation (SPEC) is a consor-
tium of companies that defines benchmarking guidelines for CPUs. Their website pub-
lishes the latest official performance data on the newest CPUs (see www.spec.org).
Modern CPUs are generally much faster than the main memory (RAM) they
access (Figure 13.1). As a result every time a CPU has to access memory from RAM it
typically needs to wait (stall) while the memory is being retrieved. This delay is
referred to as “memory access latency,” or memory latency for short. Because database
servers are both instruction and data intensive, the cache architecture used by the
CPU is very important to the performance of the database (Figure 13.2). A fast CPU
with a relatively small cache will result in a database that performs much more poorly
Figure 13.1
Modern CPU feature highlights. (Image courtesy of AMD.)
