Xiao Liu PhD Thesis.pdf - Faculty of Information and Communication ...

More documents

Recommendations

Info

emphasises the importance of effective turning points discovery and handling in the forecasting of scientific workflow activity durations. In this chapter, to address the above issues, we propose a statistical time-series pattern based forecasting strategy which can achieve better and maintain consistent performance in the interval forecasting of activity durations in cloud workflow systems. 4.3 Statistical Time-Series Pattern Based Forecasting Strategy In this chapter, we utilise time-series based forecasting models. Time-series models are probably the most widely used statistical ways for formulating and forecasting the dynamic behaviour of complex objects [13, 57]. A time series is a set of observations made sequentially through time. Some representative time series, including marketing time series, temperature time series and quality control time series, are effectively applied in various scientific and business domains [45]. Similarly, a cloud workflow activity duration time series, or duration series for short, is composed of ordered duration samples obtained from cloud workflow system logs or other forms of historical data. Therefore, duration series is a specific type of time series in the cloud workflow domain. In this thesis, the term “time series” and “duration series” can be used interchangeably in most cases. Current forecasting strategies for computation tasks mainly reside on the prediction of CPU load [3, 103]. However, this is quite different from the prediction of cloud workflow activity durations due to the reasons mentioned earlier. In this chapter, instead of applying traditional multivariate models, we conduct univariate time-series analysis which analyses the behaviour of duration series itself to build a model for the correlation between its neighbouring samples, or in other words, forecasting the future activity durations only based on the past activity durations [13]. Specifically, a novel non-linear time-series segmentation algorithm named K-MaxSDev is proposed to facilitate a statistical time-series pattern based forecasting strategy. Two problems of the duration series which seriously hinder the effectiveness of conventional time-series forecasting strategies are limited sample size and frequent turning points. Limited sample size impedes the fitting of timeseries models and frequent turning points where dramatic deviations take place 46
significantly deteriorate the overall accuracy of time-series forecasting. To address such two problems, we utilise a statistical time-series pattern based forecasting strategy. First, an effective periodical sampling plan is conducted to build representative duration series. Second, a pattern recognition process employs our K- MaxSDev time-series segmentation algorithm to discover the minimum number of potential patterns which are further validated and associated with specified turning points. Third, given the latest duration sequences, pattern matching and interval forecasting are performed to make predictions based on the statistical features of the best-matched patterns. Meanwhile, concerning the occurrences of turning points, three types of duration sequences are identified and then handled with different pattern matching results. The simulation experiments conducted in our SwinDeW-C cloud workflow system demonstrate that our time-series segmentation algorithm is capable of discovering the smallest potential pattern set compared with three generic algorithms, viz. Sliding Windows, Top-Down and Bottom-Up [48]. The comparison results further demonstrate that our statistical time-series pattern based strategy behaves better than several representative time-series forecasting strategies, viz. MEAN [31], LAST [33], Exponential Smoothing (ES) [33], Moving Average (MA) [13], and Auto Regression (AR) [13] and Network Weather Service (NWS) [93], in the prediction of high confidence duration intervals and the handling of turning points. 4.3.1 Statistical Time-Series Patterns The motivation for pattern based time-series forecasting comes from the observation that for those duration-series segments where the number of concurrent activity instances is similar, these activity durations reveal comparable statistical features. It is obvious that when the number of concurrent activities increases, the average resources that can be scheduled for each activity decrease while the overall workflow overheads increase, and vice versa. Similar observation is reported in [32] between the task execution time and the resource load in a global-scale (grid) environment. Accordingly, the duration series behaves up and down though it may not necessarily follow a linear relationship. This phenomenon, for example, is especially evident when workflow systems adopt market-driven or trust-driven 47
Page 1 and 2:
A Novel Probabilistic Temporal Fram
Page 3 and 4:
Declaration This thesis contains no
Page 5 and 6:
Abstract Cloud computing is a lates
Page 7 and 8:
Component 2, the state of scientifi
Page 9 and 10:
http://dx.doi.org/10.1016/j.jpdc.20
Page 11 and 12:
17. X. Liu, J. Chen, and Y. Yang, A
Page 13 and 14: 4.2 RELATED WORK AND PROBLEM ANALYS
Page 15 and 16: STRATEGY ..........................
Page 17 and 18: FIGURE 7.4 EXPERIMENTAL RESULTS (NO
Page 19 and 20: Chapter 1 Introduction This thesis
Page 21 and 22: Clearly, if the execution time of t
Page 23 and 24: to seek all the candidates. Further
Page 25 and 26: e unnecessary. Hence, an effective
Page 27 and 28: temporal violations, viz. recoverab
Page 29 and 30: In Chapter 2, we introduce the rela
Page 31 and 32: a two-stage searching process with
Page 33 and 34: QoS requirements. Generally speakin
Page 35 and 36: handling strategies employed in a s
Page 37 and 38: activity points to conduct temporal
Page 39 and 40: compensation, is believed as a suit
Page 41 and 42: successful completion. Specifically
Page 43 and 44: Chapter 4 and Chapter 5 respectivel
Page 45 and 46: etween time deficit compensation an
Page 47 and 48: service user’s requirements and t
Page 49 and 50: conducted for three times, i.e. sta
Page 51 and 52: PTDA+ACOWR as an example to introdu
Page 53 and 54: currently running. It is built on t
Page 55 and 56: workflow instance to determine whet
Page 57 and 58: framework for cost-effective delive
Page 59 and 60: pattern based forecasting strategy.
Page 61 and 62: 4.2 Related Work and Problem Analys
Page 63: 4.2.2 Problem Analysis In cloud wor
Page 67 and 68: activity durations: characteristics
Page 69 and 70: specified with different means and
Page 71 and 72: Table 4.1 Notations Used in K-MaxSD
Page 73 and 74: 1) Duration series building As ment
Page 75 and 76: Table 4.5 Algorithm 4: Pattern Matc
Page 77 and 78: to search for those activities whic
Page 79 and 80: cycle. Here, we also trace back the
Page 81 and 82: ecognition. Sliding Window ranks in
Page 83 and 84: predicted value will be the one pre
Page 85 and 86: segmentation algorithm is capable o
Page 87 and 88: distributed soft real-time system,
Page 89 and 90: Table 5.1 Overview on the Support o
Page 91 and 92: 2 2 can be denoted as N ( µ , σ )
Page 93 and 94: mean iteration times or with some p
Page 95 and 96: Figure 5.4 Choice Building Block No
Page 97 and 98: activities may be omitted for the e
Page 99 and 100: 5.3.1 Calculating Weighted Joint Di
Page 101 and 102: constraints until the constraint is
Page 103 and 104: Therefore, it can be expressed as n
Page 105 and 106: Table 5.3 Specification of the Work
Page 107 and 108: and the constraint for activity X 1
Page 109 and 110: ease of discussion and to avoid the
Page 111 and 112: Unfortunately, conventional discret
Page 113 and 114: The probability consistency range w
Page 115 and 116:
6.2.3 Temporal Checkpoint Selection
Page 117 and 118:
100 times each. All the activity du
Page 119 and 120:
normal distribution and correlated
Page 121 and 122:
Figure 6.3 Checkpoint Selection wit
Page 123 and 124:
work and problem analysis. Section
Page 125 and 126:
expected time redundancy of subsequ
Page 127 and 128:
the probability for self-recovery i
Page 129 and 130:
skipped if self-recovery applies. A
Page 131 and 132:
To evaluate the average performance
Page 133 and 134:
activities and around 300 violation
Page 135 and 136:
Figure 7.3 Experimental Results (No
Page 137 and 138:
Figure 7.5 Cost Reduction Rate vs.
Page 139 and 140:
8.1 Related Work and Problem Analys
Page 141 and 142:
scientific processes, human interve
Page 143 and 144:
alone does not involve any time def
Page 145 and 146:
swap slower machines in the active
Page 147 and 148:
TD( a p ) L{ ( ai , R j ) | i = p +
Page 149 and 150:
For example, in Figure 8.2, local w
Page 151 and 152:
value to sample all of the solution
Page 153 and 154:
mapping task a i to resource R j fr
Page 155 and 156:
have done some simulation experimen
Page 157 and 158:
violations can still be recovered b
Page 159 and 160:
PTDA+ACOWR for Level III Violations
Page 161 and 162:
an integer from 1 to 5 where the ex
Page 163 and 164:
D: Definition for Fitness Value Fit
Page 165 and 166:
strategy; and End(Rescheduling) is
Page 167 and 168:
(a) Optimisation Ratio on Total Cos
Page 169 and 170:
makespan. In our two stage workflow
Page 171 and 172:
espectively. It is clearly that ACO
Page 173 and 174:
comparison results on the violation
Page 175 and 176:
percentiles. The average global vio
Page 177 and 178:
equivalent times of ACOWR are calcu
Page 179 and 180:
Chapter 9 Conclusions and Future Wo
Page 181 and 182:
the real world, simulation experime
Page 183 and 184:
ased temporal consistency model. Ac
Page 185 and 186:
the whole lifecycle support for hig
Page 187 and 188:
which can be further investigated i
Page 189 and 190:
Bibliography [1] W. M. P. van der A
Page 191 and 192:
Systems", ACM Trans. on Software En
Page 193 and 194:
conjunction with 23 rd Parallel and
Page 195 and 196:
Elsevier, vol. 84, no. 3, pp. 354-3
Page 197 and 198:
Distributed Computing Systems", Jou
Page 199 and 200:
Appendix: Notation Index Symbols α
Page 201 and 202:
PTR ( U ( SW ), ( a p + , a p + 1 m
show all

Xiao Liu PhD Thesis.pdf - Faculty of Information and Communication ...

Create successful ePaper yourself

Delete template?

Save as template?