Parallel database systems are typically based on multiple standard microcessors interconnected by a local high-speed network. Effective support for inter- and intra-transaction parallelism requires both adequate use of I/O parallelism and processing parallelism.
I/O parallelism must he supported by an allocation of the database across multiple disks (declustering), either within eonveritional disIc farms or disk arrays.
Declustering supports intra-query parallelism by reading and writing large amounts of data processed by a single query in parallel from or to multiple disks.
Inter-transaction parallelism is supported because independent I/O requests n different disks can be served in parallel.
Major Architectures for Parallel Database Systems
With respect to processing parallelism, there are three major architectures for parallel database systems:
1) Shared-Memory Multiple CPU Parallel Database Architecture:
Shared-Memory Mulliple CPU (figure 3.1) refers to the use of multipocessors for database processing. In this case, one has a tightly coupled system where all processors share a common main memory as well as peripheral devices (terminals, disks). There is only a single copy of the DBMS code that can be executed in multiple processes to utilize all processors. This approach is also referred to as Symmetric Multi-Processing (SMP).
2) Shared-Disk Multiple CPU Parallel Database Architecture:
Shared Disk (SD, Figure 3.2) systems consist of multiple loosely coupled PE. However, the database is not partitioned but shared among the PE so that each DBMS instance has direct access to any data object. This assumes that each node can access any disk.
3) Shared-Nothing Multiple CPU Parallel Database Architecture:
Shared Nothing (SN, figure 3.3) systems consists of multiplye autonomous Processing Elements (PE) each owning a private main memory and running separate copies of the operating system, DBMS and other software.
Inter-processor communication takes place by means of message passing (loose coupling). A PE can consist of 1 or more processors, i.e., each node in a SN system may be a multi-processor. The database is partitioned among the PEs so that each DBMS instance can directly access only data from the local partition. Access to non-local data requires a distributed query and transaction execution.
