Xiao Liu PhD Thesis.pdf - Faculty of Information and Communication ...
Xiao Liu PhD Thesis.pdf - Faculty of Information and Communication ...
Xiao Liu PhD Thesis.pdf - Faculty of Information and Communication ...
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
significantly deteriorate the overall accuracy <strong>of</strong> time-series forecasting. To address<br />
such two problems, we utilise a statistical time-series pattern based forecasting<br />
strategy. First, an effective periodical sampling plan is conducted to build<br />
representative duration series. Second, a pattern recognition process employs our K-<br />
MaxSDev time-series segmentation algorithm to discover the minimum number <strong>of</strong><br />
potential patterns which are further validated <strong>and</strong> associated with specified turning<br />
points. Third, given the latest duration sequences, pattern matching <strong>and</strong> interval<br />
forecasting are performed to make predictions based on the statistical features <strong>of</strong> the<br />
best-matched patterns. Meanwhile, concerning the occurrences <strong>of</strong> turning points,<br />
three types <strong>of</strong> duration sequences are identified <strong>and</strong> then h<strong>and</strong>led with different<br />
pattern matching results.<br />
The simulation experiments conducted in our SwinDeW-C cloud workflow<br />
system demonstrate that our time-series segmentation algorithm is capable <strong>of</strong><br />
discovering the smallest potential pattern set compared with three generic algorithms,<br />
viz. Sliding Windows, Top-Down <strong>and</strong> Bottom-Up [48]. The comparison results<br />
further demonstrate that our statistical time-series pattern based strategy behaves<br />
better than several representative time-series forecasting strategies, viz. MEAN [31],<br />
LAST [33], Exponential Smoothing (ES) [33], Moving Average (MA) [13], <strong>and</strong><br />
Auto Regression (AR) [13] <strong>and</strong> Network Weather Service (NWS) [93], in the<br />
prediction <strong>of</strong> high confidence duration intervals <strong>and</strong> the h<strong>and</strong>ling <strong>of</strong> turning points.<br />
4.3.1 Statistical Time-Series Patterns<br />
The motivation for pattern based time-series forecasting comes from the observation<br />
that for those duration-series segments where the number <strong>of</strong> concurrent activity<br />
instances is similar, these activity durations reveal comparable statistical features. It<br />
is obvious that when the number <strong>of</strong> concurrent activities increases, the average<br />
resources that can be scheduled for each activity decrease while the overall<br />
workflow overheads increase, <strong>and</strong> vice versa. Similar observation is reported in [32]<br />
between the task execution time <strong>and</strong> the resource load in a global-scale (grid)<br />
environment. Accordingly, the duration series behaves up <strong>and</strong> down though it may<br />
not necessarily follow a linear relationship. This phenomenon, for example, is<br />
especially evident when workflow systems adopt market-driven or trust-driven<br />
47