Information Technology Reference
In-Depth Information
4
Implementations and Experiments
A prototype has been developed to implement the proposed approach and offers
4userdrivensteps:
Step1-UploadingEventLogData
The format considered in this paper to represent service interaction log files is
the Comma Separated Values (CSV). The user can upload a CSV file using a
Web interface. For our experiments, we used a CSV log file generated using HP
SOA Manager (SOAM) which is a monitoring tool for Web services. SOAM cap-
tures all the messages that are exchanged to/from the set of registered services.
The resulting log represents a scenario generated based on WS-I (Web Service
Interoperability Organization) for a set of services in the supply chain (Retailer
services). Figure 1 illustrates a short sample of data generated by SOAM.
Step 2 - Indexing Messages and Selection of Correlator Attributes
This step allows to build indexes for each attribute in the service interaction
log. Such indexes allow to speedup the algorithm for building the ACG. We use
information about the size of each index to suggest which attributes are to be
considered as relevant for message correlation.
Step 3 - Aggregating Messages to Build the ACG Graph
This step allows to build the ACG by executing algorithm 1. The inverted in-
dex described in section 3.4 is also built incrementally during this step. Once
the ACG and the inverted index are created, they are stored on the back-end
database.
Step 4 - Filtering and Visualizing the ACG Graph
This is the final step and it allows, optionally, the filtering of the ACG, and the
visualization of the ACG graph highlighting the relevant sequences of correlation
conditions.
An evaluation of the tool has been conducted using service interaction logs
of various sizes, based on the Retailer services. Four logs containing respectively
1000, 2000, 3000 and 4000 messages have been processed to generate their asso-
ciated ACG. A first analysis concerns the impact of the number of messages in
the log on the size of the filtered ACG as illustrated in figure 4 (a). It shows the
number of nodes in the resulting filtered ACG (with 50% threshold) for service
interaction logs of various sizes. This experiment shows that the size of the fil-
tered ACG is relatively the same regardless of the size of the service interaction
log. This is due to the fact that filtered ACG shows the most frequent processes
and therefore do not include all the ways messages are correlated.
A second analysis concerns the impact of the number of messages in the log on
the size of the ACG (unfiltered) as illustrated in figure 4 (b). It shows the number
of nodes in the resulting unfiltered ACG for service interaction logs of various
sizes. The number of nodes in the unfiltered ACG starts stabilizing to a certain
level at around 3000 messages in the log. This is due to the fact that process
instances are usually repetitive in the log and the proposed algorithm aggregates
similar sequences of correlation conditions. Therefore, when a large number of
messages are placed in the ACG, most of the sequences of correlation conditions
