Information Technology Reference
In-Depth Information
maintaining a high predictive accuracy. AGPM uses ensemble Machine Learning
methods to construct a meta-model comprising several sub-models to achieve
higher quality predictions.
3.1 AGPM Learning Process
AGPM drives an adaptive learning process that comprises 4 stages (see figure 1):
(
i
) execution of workflow tasks, (
ii
) performance data gathering, (
iii
)model
learning, and (
iv
) tasks run-time prediction. In the following paragraphs we
briefly discuss each of them.
Fig. 1.
Learning process carried out by AGPM
Stage 1: Workflow tasks execution.
This stage involves the execution and mon-
itoring of tasks as well as the generation of the corresponding
execution logs
,
which are later used in the following stages. This stage is carried out entirely by
the WMS.
Stage 2: Performance-data Gathering.
Consists in the harvesting of the neces-
sary information for the further learning/refinement of the performance models.
Execution logs are used to extract valuable information of tasks performance
such as the parameters and the data used, provenance information and the char-
acteristics of the resources where the tasks executed. The appliance of AGPM is
not restricted to applications for grid and cloud but also to web services. AGPM
compiles all the information that can be gathered from the running workflow
management system. The collected data is stored in separate databases for each
type of runnable task.
Stage 3: Model Learning.
At this point of the process, the databases contain up-
dated information of the last task execution. AGPM then learns a new model for
each type of task following a two-step procedure consisting on (
i
) data prepro-
cessing, and (
ii
) ensemble model learning. AGPM pre-processes the databases in
order to prepare the data for the ensemble learning strategy. As a second step,
multiple models are learned from the data and combined in order to perform
future running-time predictions. Sections 3.2 and 3.3 provide a deeper insight
on the described process which is the central contribution of this paper.
