Cost-Based Optimization of Integration Flows - Datenbanken ...
Cost-Based Optimization of Integration Flows - Datenbanken ...
Cost-Based Optimization of Integration Flows - Datenbanken ...
- No tags were found...
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
6.5 Experimental Evaluation<br />
(a) Selectivity Variations<br />
(b) Execution Time<br />
(c) <strong>Optimization</strong> Time<br />
(d) Cumulative Execution Time<br />
Figure 6.19: Simple-Flow Correlation Comparison Scenario<br />
times with and without the use <strong>of</strong> our correlation table. We observe that without the use<br />
<strong>of</strong> the correlation table, the optimization technique selection reordering assumes statistical<br />
independence and thus, changed the plan back and forth, even in case <strong>of</strong> constant<br />
workload characteristics. Due to the direct triggering <strong>of</strong> re-optimization by on-demand reoptimization,<br />
overall 1,345 re-optimization steps have been executed. In addition to this<br />
re-optimization overhead, the permanent change between suboptimal and optimal plans<br />
led to a degradation <strong>of</strong> the execution time because non-optimal plans were used for long<br />
time horizons. In contrast, when using our correlation table, the required re-optimization<br />
steps were reduced to nine. It is important to note that, in case <strong>of</strong> using the correlation table,<br />
a single workload change triggers several re-optimization steps because we allow only<br />
one operator reordering per optimization step in order to learn the conditional selectivities<br />
stepwise, which reduce the risk for cyclic re-optimizations. However, over time, the conditional<br />
selectivity estimates converge to the real selectivities. Finally, as a combination <strong>of</strong><br />
the reduced number <strong>of</strong> re-optimization steps and the prevention <strong>of</strong> using suboptimal plans<br />
due to unknown correlation, we achieved a 7% overall improvement with regard to the<br />
cumulative execution time (see Figure 6.19(d)). In conclusion, the use <strong>of</strong> our correlation<br />
table ensures robustness in the presence <strong>of</strong> correlated data or conditional probabilities,<br />
which has high importance when using on-demand re-optimization.<br />
197