Cost-Based Optimization of Integration Flows - Datenbanken ...

3.4 Optimization Techniques
[30ms,
P(A)]
@type='MATMAS04'
Translation (o3)
[in: msg1, out: msg2]
Receive (o1)
[service: s3, out: msg1]
Switch (o2)
[in: msg1]
[30ms,
P(B)]
@type='MATMAS05'
Translation (o5)
[in: msg1, out: msg2]
Assign (o4)
[in: msg2, out: msg3]
Assign (o6)
[in: msg2, out: msg3]
Invoke (o7)
[service s1, in: msg3]
Assign (o8)
[in: msg2, out: msg4]
Invoke (o9)
[service s2, in: msg4]
(a) Plan P 1 (b) Reordering Switch Paths (c) Merging Switch Paths
Figure 3.15: Example Reordering and Merging of Switch Paths
paths (assuming non-disjoint expressions, e.g., A : var1 < x and B : var1 < y), where the
total costs are independent of the path probabilities because the XPath expression is only
evaluated once. Therefore, we benefit from merging if P (A) < 1.
Finally, note that the monitored path probabilities are conditional probabilities due to
the ordered if-elseif-else semantics of the Switch operator. For example, we monitor the
relative frequency of P (path 1 ) but the conditional frequency of P (path 2 |path 1 ). Please,
refer to Subsection 3.3.4 on how to estimate conditional probabilities in this context.
Selection Reordering
Similar to traditional query processing, reordering of selective operators such as Selection,
Projection (distinct), Groupby, Join, and Setoperation (distinct) is important in order
to find the optimal plan that reduces the amount of processed data as early as possible.
In contrast to existing approaches, in the context of integration flows, the control-flow
semantics must be taken into account when evaluating selective operators. Essentially,
this control-flow awareness applies to all selective data-flow-oriented operators. However,
we use the technique WD4: Early Selection Application in order to explain this controlflow-awareness.
The core idea of selection reordering is to reduce the amount of processed data by
reordering Selection operators by their selectivity f oi = |ds out |/|ds in |, where f oi ∈ [0, 1].
The costs of a single Selection operator is given by |ds in |. Thus, the costs of a sequence
of Selection operators are determined by
C(P ) =
m∑
|ds in (o i )| =
i=1
⎛
⎞
m∑ ∏i−1
⎝ f oj · |ds in (o 1 )| ⎠ . (3.22)
i=1
This implies that the order of Selection operators is optimal if f oi ≤ f oi+1 . Due to
the problem of data correlation, the first optimization of a plan orders the Selection
operators according to this optimality condition, while all subsequent optimization steps
use the introduced correlation table for correlation-aware incremental re-ordering.
j=1
67