Databases Reference
In-Depth Information
Figure 6: Execution trace(200k 0.5% 5 nodes)
In candidate generation, occured during interval from 3.5s to 12s and also
shown in figure as phase #1, heavy probe operation during join in the first half of
this phase resulting in 100% CPU load and low disk reading throughput. In later
part of this phase disk I/O bound occured when the result of the join stored in
disk. CPU bound is also observed in other phase involving hash table probing
such as candidate matching (phase #6, 21.5s-23.5s) and during global support
gathering (phase #2, 12s-16.5s) which employs hash table updating for aggregation.
However significant network throughput dominates the global support gathering
(phase #3 and #4, 16.5s -21.5s) when processing nodes exchanging their local
support counts.
6
Performance Evaluation using Commercial Parallel RDBMS
Since the size of databases and the amount of required processing power has
increased incredibly, parallel processing ability has become a must for commercial
RDBMS. It is interesting to know whether currently available technology can
achieve sufficient performance when handling complex query such as association
rule mining.
6.1
Parallel Execution Environment
In our experiment we employed commercial Parallel RDBMS: IBM DB2 UDB
EEE version 6.1 on IBM UNIX Parallel Server System: IBM RS/6000 SP. 12 nodes
