Information Technology Reference
In-Depth Information
(a) Execution time in SPARQL
(b) Execution time in SQWRL
Fig. 7.
Performance of Pattern Matching
Results.
After running the queries (patterns) over the ATM model we found that there
are four applicable patterns (e.g., SP1, SP8, DP2.1, DP6). In Fig. 7, we present the
performance in milliseconds, of our implementation for both: SPARQL and SQWRL.
In general, there are significant differences between the two query representations.
In particular, the worst performance in SPARQL (11.7ms) is much faster then the best
performance in SQWRL (713ms). The main reason is that in the case of SPARQL
queries, the inferred model is computed only once before the matching starts and used
throughout all the queries. Thus, the inference time is not taken into account in the
SPARQL execution time. In the SQWRL case, execution time is highly affected by the
inference time.
Considering Fig. 7(b), it is an almost linear correlation between the size of the prob-
lem model and execution time. However, Fig. 7(a) indicates that the SPARQL perfor-
mance is constant after a certain model size (model size
2378)
5
. Even though the
SPARQL case outperforms the SQWRL case, designers need to be aware on the fact
that SPARQL engine does not exploit the semantics of OWL-
≥
DL
. Moreover, SPARQL
is meant to be used for querying RDF and OWL-
DL
needs to be serialized before it
can be queried. This serialization of OWL-
DL
to RDF is vendor specific therefore it
could be the case that the same OWL-
has several representations in RDF and conse-
quently different SPARQL queries. However, this dilemma does not hold for SQWRL.
DL
6
Related Work
The growing size of pattern libraries has spawned the following challenges: 1) finding
a relevant pattern in the pattern library, 2) selecting a pattern that is suited with the
problem-at-hand, and 3) applying a pattern. For the first challenge, though there is no
central index as mentioned in [3] several initiatives are trying to collect software design
5
We acknowledge some irregularities on the execution time for the 3 smallest models.
