Information Technology Reference
In-Depth Information
select a subset of checkpoints for violation handling from the set of necessary and
sufficient checkpoints for temporal verification. Traditionally, scientific workflow
temporal verification adopts the philosophy that to maintain satisfactory temporal
QoS, similar to the handling of functional exceptions, temporal violation handling
should be triggered on every necessary and sufficient temporal checkpoint [33]. How-
ever, there is a common but overlooked phenomenon that the execution delay may
often be small enough so that the saved execution time of the subsequent workflow
activities could automatically compensate for it. In [28], an adaptive violation han-
dling point selection strategy is proposed to fully utilize this kind of “
self-recovery
”
phenomenon to significantly reduce the number of violation handling points, so as to
reduce the temporal total violation handling cost.
3
Open Challenges
As discussed in Section 2, a clear research roadmap for scientific cloud workflow
temporal verification is to follow the generic temporal verification framework and
investigate new strategies for each component. Although temporal verification for
scientific workflows has been extensively studied in the last few years, there are still
many open challenges especially brought by the shift of the computing paradigm from
conventional cluster or grid to the cloud. However, since not every researcher needs
to follow the generic framework and they may only be interested in the studies of one
of the components, we try to focus on three major and high-level open challenges in
this paper instead of discussing the very specific tasks for each component in this
section. For each open challenge, we will discuss its problems and try to point out
some potential research directions with preliminary results if available. Specifically,
there are three major open challenges as follows.
3.1
Open Challenge #1: The Forecasting Strategy for Scientific Cloud
Workflow Activity Durations
The accurate modeling and estimation of activity durations is very important to the
effectiveness of temporal verification. However, this is not a trivial issue in a cloud
environment. Traditional scientific workflow systems are often deployed in communi-
ty based computing environment such as clusters and grids where resources are usual-
ly reserved in advance and “best-effort” based QoS strategy are adopted [40]. In such
a case, the performance of the underlying computing resources is relatively stable and
thus the activity durations can be modeled with reasonable accuracy with simple sta-
tistic models [22]. Some advanced forecasting strategies based on time-series patterns
and CPU trend analysis have been proved to be very effective for scientific
workflows[2, 58]. In contrast, cloud is a multitenant computing environment and
dynamic resource provisioning is often required to guarantee the target service quality
purchased by the users. Therefore, the underlying resources are dynamically changing
and hence their performance is hard to predict. In addition, traditional forecasting
strategies mainly concern with the computing time of the workflow activities.
