New Data Warehouse Technologies - Data Warehouse Systems: Design and Implementation - page 507

called caching . Note, however, that caching only speeds up database reads,
while updates or writes must still be written through the cache to disk.
Therefore, the performance benefit only applies to a subset of database tasks.
In addition, managing the cache is itself a process that requires substantial
memory and CPU resources. An IMDBS reduces to a minimum these data
transfers, since data are mainly in memory. It follows that the optimization
objectives of disk-based database systems are opposed to those of an IMDBS.
Traditional DBMSs try to minimize input/output (I/O) using the cache,
consuming CPU cycles to maintain this cache. In addition, as we have seen,
they keep redundant data, for example, in index structures, to enable direct
access to records without the need to go down to the actual data. On the
contrary, an IMDBS is designed with the optimization goal of reducing both
memory consumption and CPU cycles.
Like traditional DBMSs, typical IMDBSs support the ACID properties ,
namely, atomicity, consistency, isolation, and durability. The first three ones
are supported as in traditional DBMSs. Since the main memory is volatile,
durability is supported by transaction logging, in which snapshots of the
database are called periodically at certain time instants (called savepoints or
checkpoints , depending on the technology and the vendor) and are written
to nonvolatile media. If the system fails and must be restarted, the database
either rolls back to the last completed transaction or rolls forward to complete
any transaction that was in progress when the system failed. IMDBSs also
support durability by maintaining one or more copies of the database, which,
as in traditional systems, is called replication . Nonvolatile RAM provides
another means of in-memory database persistence.
Finally, disk-based storage can be applied selectively in an IMDBS. For
example, certain record types can be written to disk, while others are
managed entirely in memory. Functions specific for disk-based databases, such
as cache management, are applied only to records stored on disk, minimizing
the impact of these activities over performance and CPU demands.
Figure 13.4 depicts the typical data storage architecture of an IMDBS. 1
The database is stored in main memory, and it is composed of three main
parts. The main store contains data stored in a column-oriented fashion.
For query optimization reasons, some products also store together groups
of columns that are usually accessed together. These are called combined
columns. The buffer store is a write-optimized data structure that holds
data that have not yet been moved to the main store. That means that a
query can need data from both the main store and the buffer. The special
data structure of the buffer normally requires more space per record than
the main store. Thus, data are periodically moved from the buffer to the
main store, a process that requires a merge operation. There are also data
structures used to support special features. Examples are inverted indexes
1 This figure is inspired by the SAP HANA architecture (described later in the
chapter), although most IMDBSs follow a similar architecture.