Databases Reference
In-Depth Information
Figure 4: Diagram of for all statement
2.
A dynamic performance optimization mechanism
Unlike array computation, the load information cannot be estimated at
compilation time. The information must be collected at run time, and the
necessary actions must be taken when the load imbalance is excessive.
Location-Independent Object Reference Mechanism
The location-independent object reference mechanism is implemented within a
heap management object
(HMO) allocated in each PE
10)
. An HMO can translate
an abstract OID to its address when the object referred by the OID is local, or can
obtain the ID of the PE holding the object when the object is remote.
An HMO manages all incoming messages and outgoing messages. The HMO
looks up its
object reference table
(ORT) whenever an OID reference is found in
an operation, and determines whether the OID is located within the PE. If the
target object is not registered in the ORT, the object is remote and the HMO sends
an inter-processor communication to delegate the operation to another PE.
When the HMO cannot find the entry for the given OID, the HMO looks up its
object allocation table
(OAT) and checks whether the object is registered. If the
object is registered, the HMO sends a 1-to-1 message to the PE. If not, the HMO
broadcasts a request message and registers the PE ID after receiving the reply from
the remote PE holding the object.
This indirect reference mechanism is sufficiently effective even for complex
operations such as traverses
10)
.
Since the individual OIDs contain no location information, objects can be
relocated without destroying the relationships among them. This feature is essential
for dynamic performance optimization mechanisms.
