|
|
As a distributed DBMS
|
- System configuration
In order to manage large amount of data, we must examine
communications in loosely coupled parallel processing.
Closely related data should be located in the same cluster and
query processing plans are responsible for decreasing communications.
Kappa is composed of local DBMSs(LDBMSs), all of which manage
one database. Each LDBMS has a full set of functions of DBMS.
Functions of distributed transactions are implemented based on
two phase commitment protocol to handle queries concerning
multiple LDBMS.
Global information such as table names of a database is accessed by
multiple LDBMSs, which can cause concentration of access to a server
DBMS(SDBMS). Replicants of a SDBMS are created to prevent the
congestion.
LDBMSs on each cluster do parallel processing suited for tightly
coupled multi processors.
- Data placement and parallel processing
Load of each cluster and communications among them must be
balanced for efficient parallel processing. In case of DBMS, large
amount of data are stored in secondary memories. Therefore,
data placement is closely related to methods of parallel processing.
- Distribution
Distribution of relations or tables is the simplest way of
exploiting computational power of multiple processors.
Tables must be distribued taking communications and load
balance into account.
- Horizontal partition
Horizontal partitioning is a method of bringing out
parallelism. A relation is horizontally partitioned
in records and distributed to different LDBMSs. Basically,
the same operation is requested to each partitioned
relation and the operational results are collected at last.
This method is effective when a relation is too large to
be operated in a cluster or when CPU bound operations
such as data searching are important.
- Replication
Replicated relations contribute availability and prevent
database access from congestion.
|