Table and Index Organization - Relational Database Index Design and the Optimizers - page 21

In addition to the I/O time estimates, the cost of disk space and memory
influences index design.
Local disk drives provide physical data storage without the additional func-
tion provided by disk servers (such as fault tolerance, read and write cache,
striping, and so forth), for a very low price.
Disk servers are computers with several processors and a large amount of
memory. The most advanced disk servers are fault tolerant: All essential com-
ponents are duplicated, and the software supports a fast transfer of operations to
a spare unit. A high-performance fault tolerant disk server with a few terabytes
may cost $2 million. The cost per gigabyte, then, is in the order of U.S.$500
(purchase price) or U.S.$50 per month (outsourced hardware).
Both local disks and disk servers employ industry-standard disk drives. The
largest drives lead to the lowest cost per gigabyte; for example, a 145-GB drive
costs much less than eight 18-GB drives. Unfortunately, they also imply much
longer queuing times than smaller drives with a given access density (I/Os per
gigabyte per second).
The cost of memory has been reduced dramatically over the last few years as
well. A gigabyte of random access memory (RAM) for Intel servers (Windows
and Linux) now costs about $500 while the price for RISC (proprietary UNIX and
Linux) and mainframe servers (z/OS and Linux) is on the order of U.S.$10,000
per gigabyte. With 32-bit addressing, the maximum size of a database buffer pool
might be a gigabyte (with Windows servers, for example), and a few gigabytes
for mainframes that have several address spaces for multiple buffer pools. Over
the next few years, 64-bit addressing, which permits much larger buffer pools,
will probably become the norm. If the price for memory (RAM) keeps falling,
database buffer pools of 100 gigabytes or more will then be common.
The price for the read cache of disk servers is comparable to that of RISC
server memory. The main reason for buying a 64-GB read cache instead of
64 GB of server memory is the inability of 32-bit software to exploit 64 GB for
buffer pools.
Throughout this topic, we will use the following cost assumptions:
CPU time $1000 per hour, based on 250 mips per processor
Memory $1000 per gigabyte per month
Disk space $50 per gigabyte per month
These are the possible current values for outsourced mainframe installations.
Each designer should, of course, ascertain his or her own values, which may be
very much lower than the above.
DBMS SPECIFICS
Pages
The size of the table pages sets an upper limit to the length of table rows.
Normally, a table row must fit in one table page; an index row must fit in one