- Email: [email protected]
Data transmission strategies for resource monitoring in cloud computing platforms
Accepted Manuscript Title: Data Transmission Strategies for Resource Monitoring in Cloud Computing Platforms Author: Xiaobo Ji Fan Zeng Mingwei Lin PII: DOI: Reference:
S0030-4026(16)30387-4 http://dx.doi.org/doi:10.1016/j.ijleo.2016.04.114 IJLEO 57597
To appear in: Received date: Accepted date:
3-3-2016 20-4-2016
Please cite this article as: Xiaobo Ji, Fan Zeng, Mingwei Lin, Data Transmission Strategies for Resource Monitoring in Cloud Computing Platforms, Optik - International Journal for Light and Electron Optics http://dx.doi.org/10.1016/j.ijleo.2016.04.114 This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
Data Transmission Strategies for Resource Monitoring in Cloud Computing Platforms
Xiaobo Ji1, Fan Zeng2, Mingwei Lin3 1
Department of Information, Research Institute of Field Surgery, Daping Hospital, Third Military Medical University, Chongqing, China
2
Department of Information, Research Institute of Field Surgery, Daping Hospital, Third Military Medical University, Chongqing, China 3
Faculty of Software, Fujian Normal University, Fuzhou 350007, China 1
[email protected], [email protected], [email protected]
Abstract This paper proposes three data transmission strategies for resource monitoring in cloud computing platforms, which are dynamic periodic push strategy, window-based event-driven push strategy, and window-based hybrid push strategy, respectively. The dynamic periodic push strategy periodically pushes resource status information at a dynamic time interval that is dynamically updated by considering the change degree of resource status information. The window-based event-driven push strategy pushes resource status information if the change degree of resource status information is larger than a threshold that is dynamically updated by using the exponentially weighted moving average. The window-based hybrid push strategy combines the dynamic periodic push strategy and the window-based event-driven push strategy and then further reduces the communication overhead and improves the data coherence. Experimental results show that the window-based hybrid push strategy performs better than previous data transmission strategies in terms of the number of data transmissions and the data coherence.
Key words: data transmission strategy, cloud computing platform, resource monitoring
1. Introduction Cloud computing is a new computing model that is based on the Internet network [1]. Compared with traditional IT infrastructure, cloud computing platforms introduce the virtualization technology to build resource pools and can deal with the problem that
the traditional IT infrastructure shows low hardware resource utilization and long application deployment time [2]. Therefore, more and more enterprises choose to move their applications to cloud computing platforms [3]. As applications deployed in cloud computing platforms increase in type and number, scales of cloud computing platforms are increasingly expanding and cloud computing platforms become more and more complex [4], [5], [6]. Moreover, virtual machines in cloud computing platforms share hardware resources [7]. These make cloud computing platforms be prone to error [8]. Resource monitoring can help system administrators understand running status of cloud computing platforms and system administrators can take precautions before cloud computing platforms fail to run [9][10]. The data transmission strategy is an important part of resource monitoring and has a significant influence on the performance of resource monitoring in terms of the communication overhead and the data coherence [11]. Existing data transmission strategies cannot achieve a better trade-off between the low communication overhead and high data coherence. In this case, three mechanisms were proposed, which are the offset-sensitive mechanism (OSM), the time-sensitive mechanism (TSM), and the announcing with change and time consideration (ACTC), respectively [12]. In these three mechanisms, the time interval is dynamically updated as the average of time intervals between resource status information changes and the threshold is dynamically updated as the average amount of changes for resource status information. However, the time interval does not consider the change degree of resource status information and the threshold does not consider the change trend of
resource status information. Therefore, the time interval and the threshold in these three mechanisms may be too small or too large and then performances of these three mechanisms will be affected. In this paper, three data transmission strategies are proposed, which are the dynamic periodic push strategy, window-based event-driven push strategy, window-based hybrid push strategy, respectively. The dynamic periodic push strategy periodically pushes resource status information at a time interval that is dynamically updated by considering the change degree of resource status information. The window-based event-driven push strategy pushes resource status information if the change degree of resource status information is larger than a threshold that is dynamically updated by using the exponentially weighted moving average. The window-based hybrid push strategy combines the advantages of the dynamic periodic push strategy and the window-based event-driven push strategy and then further reduces the communication overhead and improves the data coherence. Experimental results show that the window-based hybrid push strategy performs better than previous data transmission strategies in terms of the number of data transmissions and the data coherence. The remainder of this paper is organized as follows. Section 2 briefly reviews existing works on data transmission strategy. Section 3 shows our proposed data transmission strategies. Section 4 presents experimental results. Finally, conclusions are drawn in Section 5.
2. Existing works on data transmission strategy
In large-scale distributed networks, there are two basic data transmission strategies, which are the push strategy and the pull strategy [13] [14]. According to different trigger conditions, the push strategy can be divided into the periodic push strategy and the event-driven push strategy, respectively. For the periodic push strategy, monitored nodes periodically send their resource status information to the monitoring node at a time interval. For the event-driven push strategy, monitored nodes send their resource status information to the monitoring node if the change degree of resource status information is larger than a threshold. The pull strategy can also be divided into the periodic pull strategy and the event-driven pull strategy, respectively. For the periodic pull strategy, the monitoring node periodically sends requests for obtaining resource status information to monitored nodes at a time interval and monitored nodes send their resource status information to the monitoring node after receiving requests. For the event-driven pull strategy, the monitoring node sends requests for obtaining resource status information to monitored nodes if the change degree of the resource status information is larger than a threshold and then monitored nodes send their resource status information to the monitoring node after receiving requests. The time interval and the threshold have a significant influence on performances of the push strategy and the pull strategy. If the time interval or the threshold is too small, useless resource status information will be collected and then the communication overhead will be increased. If the time interval or the threshold is too large, important resource status information will be lost and then the data coherence will be lowered. In order to reduce the communication overhead and improve the data coherence, a number of
data transmission strategies have been proposed. R. Sundaresan et al. proposed three modified pull strategies, which are the regular polling strategy, the adaptive polling strategy, and the slacker polling strategy [15][16]. In the regular polling strategy, the time interval is a constant and cannot change with the change of resource status information. In the adaptive polling strategy, the time interval can be dynamically adjusted by using a damping factor. However, the damping factor is a constant and then the time interval cannot change with the change degree of resource status information. In the slacker polling strategy, the moving average estimator is introduced to estimate the next time interval. All these three data transmission strategies cannot pull resource status information during the time interval. F. F. Han et al. proposed a periodically and event-driven push (PEP) algorithm, which is based on the push strategy [17]. In the PEP algorithm, monitored nodes periodically send their resource status information to the monitoring node at a time interval of 1 second. If the change degree of resource status information is larger than a predefined threshold, monitored nodes also send their resource status information to the monitoring node. The PEP algorithm could reduce the communication overhead and improve the data coherence. However, the time interval and the threshold in the PEP algorithm are constants and cannot change with the change degree resource status information. Therefore, the time interval and the threshold in the PEP algorithm may be too small or too large. W. C. Chung and R. S. Chang proposed three data transmission mechanisms for grid
computing, which are the offset-sensitive mechanism (OSM), the time-sensitive mechanism (TSM), and the announcing with change and time consideration mechanism (ACTC) [12]. All these three mechanisms are based on the push strategy. The OSM is an event-driven push strategy. In the OSM, monitored nodes send their resource status information to the monitoring when the change degree of resource status information is larger than a threshold. The threshold is initialized to 0 and then the threshold is dynamically updated as the average amount of changes for resource status information. In the TSM, monitored nodes periodically send their resource status information to the monitoring node at a time interval. The time interval is calculated as the average of time intervals between resource status information changes. The ACTC combines the advantages of the OSM and TSM strategies and effectively improves the performance of data transmission. However, the time interval does not consider the change degree of resource status information and the threshold does not consider the change trend of resource status information. Therefore, the time interval and the threshold in these three mechanisms may be too small or too large. H. Huang and L. Wang proposed a combined push-pull model named P&P for cloud computing environment [18]. The P&P model consists of the P&P-Push algorithm and the P&P-Pull algorithm, respectively. The P&P-Push algorithm runs on monitored nodes, while the P&P-Pull algorithm runs on the monitoring node. The P&P model could intelligently switch between the P&P-Push algorithm and the P&P-Pull algorithm according to users’ requirements and resource status. It can reduce the communication overhead and improve the data coherence.
3. Proposed data transmission strategies This section presents our proposed data transmission strategies, which are the dynamic periodic push strategy, the window-based event-driven push strategy, and the window-based hybrid push strategy. 3.1 Dynamic periodic push strategy In order to address the problem that the time interval in existing periodic push strategies does not considering the change degree of resource status information, a dynamic period push strategy is proposed. In the proposed dynamic period push strategy, a constant denoted as min_variation is defined to determine whether resource status information during the time interval changes greatly. If the resource status information during the time interval changes and its change degree is larger than min_variation, the time interval (TI) is updated as TI TI
min_variation A B
(1)
where A is the value of resource status information after change and B is the value of resource status information before change. When the timer expires, monitored nodes send their resource status information to the monitoring node. If the change between resource status information sent now and resource status information sent last time is less than min_variation, the time interval is updated as
1 TI TI 1 C P e
(2)
where C is the value of resource status information sent now and P is the value of resource status information sent last time. Figure 1 Algorithm of the dynamic periodic push strategy 3.2 Window-based event-driven push strategy In order to deal with the problem that the threshold in existing event-driven push strategies does not consider the change trend of resource status information, a window-based event-driven push strategy is proposed. Definition 1 Resource status information window, a triplet
Lmax , L, d _ threshold , is
used to store resource status information. where Lmax and L are the maximum length and current length of the window. The term d _ threshold is the threshold of the change of resource status information in the window and its initial value is 0. The core idea of the window-based event-driven push strategy is to introduce the exponentially weighted moving average to predict the value of d _ threshold . The exponentially weighted moving average is designed from the moving average and calculates the prediction value by using the following formula: E t O t 1 1 E t 1
(3)
where E t is the prediction value at time period t and O t 1 is the observation value at time period t 1 . The coefficient is a constant smoothing factor between 0 and 1. Assume that the resource status information window has a resource status information data series S1 , S 2 , S3 ,, S L . If a piece of new resource status information denoted as
S L 1 is stored into the resource status information window, the threshold is calculated
as S L 1 S L 1 d _ threshold L d _ threshold L 1 S 2 S1 0
, L 1 , L 1 , L0
(4)
where d _ threshold L 1 is the prediction value of the threshold after storing the new resource status information into the resource status information window. Figure 2 Algorithm of the window-based event-driven push strategy 3.3 Window-based hybrid push strategy The time interval in the dynamic push strategy changes with the change degree of resource status information and then the dynamic push strategy could reduce the transmission of useless resource status information and the loss of important resource status information. Therefore, the dynamic push strategy could reduce the communication overhead with relatively high data coherence. However, the dynamic push strategy cannot push the important resource status information during the time interval. The window-based event-driven push strategy can be aware of the change trend of resource status information by introducing the exponentially weighted moving average to predict the threshold. Resource status information changes significantly recently and then the threshold is very high. If resource status information changes slightly at the next time period, the change of resource status information will be less than the threshold and then resource status information will not be pushed at this time period. In this case, collected resource status information will be insufficient and the
performance of resource monitoring will be lowered. The dynamic periodic push strategy can address the problem that the window-based event-driven push strategy may not push resource status information for a long time, while the window-based event-driven push strategy can address the problem that the dynamic periodic push strategy cannot push important resource status information during the time interval. In order to address problems in the dynamic periodic push strategy and the window-based event-driven push strategy, this paper combines the dynamic periodic push strategy and the window-based event-driven push strategy and proposes a window-based hybrid push strategy. Figure 3 Algorithm of the window-based hybrid push strategy
4. Experimental analysis
In order to evaluate three proposed data transmission strategies, a series of experiment have been conducted on a private cloud computing platform. This section shows the experiment environment and then analyzes the experimental results. 4.1 Experiment environment The experiment platform consists of six physical nodes, which are connected by using the 100M Ethernet. One is selected to be a cloud controller, while the rest are used as compute nodes. A private cloud computing platform is deployed on these six physical nodes by using the Xen [19] and OpenStack open source software [20]. Resource monitoring is performed by collecting resource parameters, which can be obtained from the /proc file in the operating system. In order to simplify experiments, only one
of resource parameters, i.e. the CPU utilization, is used to evaluate the performance of our proposed data transmission strategies [21]. In order to evaluate the performance of our proposed data transmission strategies, the dynamic periodic push (DPP) strategy, the window-based event-driven push (WEDP) strategy, and the window-based hybrid push (WHP) strategy are compared with the slacker polling strategy (SP), the periodically and event-driven push (PEP) algorithm, the announcing with change and time consideration (ACTC) mechanism, and the combined push-pull (P&P) model. Evaluation metrics in our experiments are the number of data transmissions and data coherence. The number of data transmissions is the sum of the number of pull operations and the number of push operations. The data coherence (coh) is calculated as n
coh
C i R i
2
i 1
(5)
n
where C i and R i are the point on the graph of actual measurements and the point on the graph of generated measurements, respectively. The higher the data coherence is, the smaller the value of coh is. 4.2 Experimental results We have conducted six groups of experiments on the private cloud computing platform. The first four experiments are conducted to test whether the performance of the window-based hybrid push strategy depends on the constant min_variation and the maximum length of the window. The rest experiments are performed to compare our proposed data transmission strategies with existing data transmission strategies.
The parameter Push_interval in the P&P model was set to 1s. The parameters PULL_INIT_INTERVAL, PULL_INTERVAL_MIN, and PULL_INTERVAL_MAX were set to 5s, 3s, and 12s, respectively. (1) Relations between the number of data transmissions and min_variation Figure 4 Relations between the number of data transmissions and min_variation For the window-based hybrid push strategy, the number of data transmissions reduces with the increase of the value of the min_variation when the maximum length of the window is fixed. The number of data transmissions reduces most significantly when the min_variation is equal to 25%. If the min_variation is larger than 50%, the number of data transmissions will tend to be stable. (2) Relations between the number of data transmissions and the maximum length of the window Figure 5 Relations between the number of data transmissions and maximum length of the window As shown in Figure 5, the number of data transmissions depends on the maximum length of the windows when the constant min_variation is fixed. As the maximum length of the window increases, the number of data transmissions decreases. The number of data transmissions tends to be stable when the maximum length of the window is larger than 50. (3) Relations between the data coherence and min_variation Figure 6 Relations between the data coherence and min_variation As shown in Figure 6, the data coherence depends on the parameter min_variation. As
the parameter min_variation increases, the data coherence decreases. The data coherence tends to be stable when the parameter min_variation is larger than 45%. (4) Relations between the data coherence and the maximum length of the window Figure 7 Relations between the data coherence and maximum length of the window As shown in Figure 7, as the maximum length of the window increases, the data coherence decreases when the parameter min_varition is fixed. When the maximum length of the window increases, the threshold increases and then some important resource status information are lost during the time interval. Therefore, the data coherence decreases. When the maximum length of the window is larger than 35, the data coherence tends to be stable. (5) Comparison of the number of data transmissions As shown in Figure 8, the number of data transmissions of the proposed DPP and WEDP strategies is more than the PEP algorithm, the ACTC mechanism, and the P&P model. This is because that the time interval in the proposed DPP strategy and the threshold in the proposed WEDP strategy may be too small. The number of data transmissions of the proposed WHP strategy is less than other data transmission strategies. This is because that its time interval can be dynamically updated by considering the change degree of resource status information and its threshold can be dynamically updated by using the exponentially weighted moving average. Therefore, the time interval and the threshold can change with resource status information and will not be too small or too large. (6) Comparison of the data coherence
As shown in Figure 9, the data coherence of the proposed DPP and WEDP strategies is lower than other data transmission strategies, but lower than 10% and can satisfy the basic requirement of resource monitoring. The proposed WHP strategy shows the highest data coherence. This is because the WHP strategy can push important resource status information during the time interval. Moreover, its time interval and threshold can change with resource status information and will not be too small or too large. The time interval and the threshold in the PEP algorithm are constants, so they will be too large when resource status information changes frequently and significantly. The time interval in the ACTC mechanism is simply calculated as the average of time intervals between resource status information changes and its threshold is simply calculated as the average of change degree for resource status information. The time interval in the ACTC mechanism cannot be aware of the change degree of resource status information and its threshold cannot be aware of the change trend of resource status information, so the time interval and the threshold in the ACTC mechanism may be too large. The parameters in the P&P model are constants, so they cannot change with resource status information and they may be too large. Large time interval and threshold will result in missing important resource status information and reduce data coherence.
V. Conclusions
Three data transmission strategies are proposed in this paper, which are the dynamic periodic push strategy, the window-based event-driven push strategy, and the
window-based hybrid push strategy. The dynamic periodic push strategy periodically pushes resource status information at a time interval that is dynamically updated by considering the change degree of resource status information. The window-based event-driven push strategy pushes resource status information when the change degree of resource status information is larger than a threshold that is predicted by using the exponentially weighted moving average. The window-based hybrid push strategy combines the advantages of the dynamic periodic push strategy and the window-based event-driven push strategy. A series of experiments have been conducted on a private cloud computing platform built by using the Xen and OpenStack open source software. Experimental results show that our proposed window-based hybrid push strategy outperforms previous data transmission strategies.
Acknowledgments
The work of this paper is supported by the Natural Science Foundation Project of CQ CSTC under Grant No. cstc2013jcyjA40059, National Natural Science Foundation of China under Grant No. 61502102, No. 61370078, and No. 61402109, Fujian Province Education Scientific Research Projects for Young and Middle-aged Teachers under Grant No. JA15122, National Undergraduate Training Programs for Innovation and Entrepreneurship under Grant No. 201510394021, and Fujian Normal University Undergraduate Training Programs for Innovation and Entrepreneurship under Grant No. cxxl-2015163.
References
[1] Ronnie D. Caytiles, Sunguk Lee, Byungjoo Park. Cloud computing: The next computing paradigm. International of Journal of Multimedia and Ubiquitous Engineering, vol. 7, no. 2, pp. 297-302, 2012. [2] Chi Yuan, Honglei Yu, Deli Yang. Cost comparison between cloud computing and traditional data center. WIT Transactions on Modelling and Simulation, vol. 60, pp. 733-740, 2014. [3] Leire Orue-Echevarria, Juncal Alonso, Hugo Bruneliere, Andreas Menychtas, Philip Langer, Manuel Wimmer. Cloudifying applications with ARTIST a global modernization approach to move applications to cloud. Proceedings of the 4th International Conference on Cloud Computing and Services Science, pp. 737-745, 2014. [4] H. S. Pannu, J. Liu, Q. Guan, S. Fu. AFD: Adaptive failure detection system for cloud computing infrastructures. 2012 IEEE 31st International Performance Computing and Communications Conference, pp. 71-80, 2012. [5] H. S. Pannu, J. Liu, S. Fu. AAD: Adaptive anomaly detection system for cloud computing infrastructures. Proceedings of 2012 31st International Symposium on Reliable Distributed Systems, pp. 396-397, 2012. [6] Q. Guan, S. Fu, N. D. Bardeleben, S. Blanchard. Exploring time and frequency domains for accurate and automated anomaly detection in cloud computing systems. Proceedings of 2013 IEEE 19th Pacific Rim International Symposium on Dependable Computing, pp. 196-205, 2013.
[7] Fei Xu, Fangming Liu, Hai Jin, Athanasios V. Vasilakos. Managing performance overhead of virtual machines in cloud computing: A survey, state of the art, and future directions. Proceedings of the IEEE, vol. 102, no. 1, pp. 11-31, 2014. [8] M. Gupta, A. B. Sharma, H. Chen, G. Jiang. Context-aware time series anomaly detection for complex systems. 2013 Workshop on Data Mining for Service and Maintenance, pp. 14-22, 2013. [9] Junwei Ge, Bo Zhang, Yiqiu Fang. Research on the resource monitoring model under cloud computing environment. Lecture Notes in Computer Science, vol. 6318 LNCS, pp. 111-118, 2010. [10] Chao-Tung Yang, Wen-Chung Shih, Chih-Lin Huang. Implementation of a distributed data storage system with resource monitoring on cloud computing. Lecture Notes in Computer Science, vol. 7296 LNCS, pp. 64-73, 2012. [11] Kai Lin, Weiqin Tong, Xiaodong Liu, Liping Zhang. A self-adaptive mechanism for resource monitoring in cloud computing. IET International Conference on Smart and Sustainable, pp. 178-182, 2013. [12] Wu-Chun Chung, Ruay-Shiung Chang. A new mechanism for resource monitoring in Grid computing. Future Generation Computer Systems, vol. 25, no.1, pp. 1-7, 2009. [13] M. L. Massie, B. N. Chun, D. E. Culler. The ganglia distributed monitoring system: Design, implementation, and experience. Parallel Computing, vol. 30, no. 7, pp. 817-840, 2004. [14] E. Imamagic, D. Dobrenic. Grid infrastructure monitoring system based on
Nagios. Proceedings of the 2007 Workshop on Grid Monitoring, pp. 23-28, 2007. [15] R. Sundaresan, M. Lauria, T. Kurc, S. Parthasarathy, J. Saltz. Adaptive polling of grid resource monitors using a slacker coherence model. Proceedings of IEEE International Symposium on High Performance Distributed Computing, pp. 260-269, 2003. [16] R. Sundaresan, T. Kurc, M. Lauria, S. Parthasarathy, J. Saltz. A slacker coherence protocol for pull-based monitoring of on-line data sources. Proceedings of 3rd IEEE/ACM International Symposium on Cluster Computing and the Grid, pp. 250-257, 2003. [17] F. F. Han, J. J. Peng, W. Zhang, Q. Li, J. D. Li, Q. L. Jiang, Q. Yuan. Virtual resouce monitoring in cloud computing. Journal of Shanghai University, vol. 15, no. 5, pp. 381-385, 2011. [18] H. Huang, L. Wang. P&P: A combined push-pull model for resource monitoring in cloud computing environment. Proceedings of 2010 IEEE 3rd International Conference on Cloud Computing, pp. 260-267, 2010. [19] P. Barham, B. Dragovic, K. Fraser, S. Hand, T. Harris, A. Ho, R. Neugebauer, L. Pratt, A. Warfield. Xen and the art of virtualization. Operating Systems Review (ACM), vol. 37, no. 5, pp. 164-177, 2003. [20] A. Corradi, M. Fanelli, L. Foschini. VM consolidation: A real case based on OpenStack Cloud. Future Generation Computer Systems, vol. 32, no. 1, pp. 118-127, 2014. [21] Mingwei Lin, Zhiqiang Yao, Tianqiang Huang. A hybrid push protocol for
resource monitoring in cloud computing platforms. Optik - International Journal for Light and Electron Optics, vol. 127, no. 4, pp. 2007-2011, 2016.
Figure 1 shows the algorithm of the proposed dynamic periodic push strategy. Algorithm 1: Dynamic Periodic Push Strategy Input: min_variation; N = 0; Tc presents the current time; Output: None 1: Push current resource status information to the monitoring node; 2: Timer = ∞; 3: T0 = Tc; 4: WHILE(TRUE) 5: IF The Timer expires 6: Push current resource status information to the monitoring node; 7: IF |C-P| < min_variation 8: Calculate the new TI using the equation (2); 9: ENDIF 10: Reset Timer to TI; 11: ENDIF 12: IF The resource status information changes 13: IF N = 0 14: Push current resource status information to the monitoring node; 15: TI = Tc – T0; 16: N = N + 1; 17: Reset Timer to TI; 18: ELSE IF |A-B| > min_variation 19: Calculate the new TI using the equation (1); 20: ENDIF 21: ENDIF 22:ENDWHILE
Figure 2 shows the algorithm of the window-based event-driven push strategy. Algorithm: Window-based Event-driven Push Strategy Input: Lmax; L = 0; Output: None 1: Push resource status information to the monitoring node; 2: Store one copy of pushed resource status information in the window; 3: Update the threshold of the window by using the equation (4); 4: L = L + 1; 5: WHILE(TRUE) 6: IF Resource status information changes and |A-B| > d_thresholdL 7: Push current resource status information to the monitoring node; 8: IF L = Lmax 9: Remove S1 from the window; 10: L = L – 1; 11: ENDIF 12: Store one copy of pushed resource status information in the window; 13: Update the threshold of the window by using the equation (4); 14: L = L + 1; 15: ENDIF 16: ENDWHILE
Figure 3 shows the algorithm of the window-based hybrid push strategy. Algorithm: Window-based Hybrid Push Strategy Input: min_variation; N = 0; Tc presents the current time; Lmax; L = 0; Output: None 1: Push current resource status information to the monitoring node; 2: Timer = ∞; 3: T0 = Tc; 4: Store one copy of pushed resource status information in the window; 5: Update the threshold of the window by using the equation (4); 6: L = L + 1; 7: WHILE(TRUE) 8: IF The Timer expires 9: IF |C-P| > min_variation 10: Push current resource status information to the monitoring node; 11: IF L = Lmax 12: Remove S1 from the window; 13: L = L -1; 14: ENDIF 15: Store one copy of pushed resource status information in the window; 16: Update the threshold of the window by using the equation (4); 17: ELSE 18: Calculate the new TI using the equation (2); 19: ENDEIF 20: Reset Timer to TI; 21: ENDIF 22: IF resource status information changes 23: IF N = 0 24: Push current resource status information to the monitoring node; 25: TI = Tc – T0; 26: N = N + 1; 27: Store one copy of pushed resource status information in the window; 28: Update the threshold of the window by using the equation (4); 29: L = L + 1; 30: Reset Timer to TI; 31: ELSE IF |A-B| > d_thresholdL 32: Push current resource status information to the monitoring node; 33: Calculate the new TI using the equation (1); 34: IF L = Lmax 35: Remove S1 from the window; 36: L = L - 1; 37: ENDIF 38: Store one copy of pushed resource status information in the window; 39: Update the threshold of the window by using the equation (4); 40: L = L + 1; 41: ENDIF 42: ENDIF 43:ENDWHILE
Figure 4 illustrates the relations between the number of data transmissions and the
The number of data transmissions
min_variation when the maximum length of the window is fixed to 25. 4500 4000 3500 3000 2500 2000 0
10
20
30
40
50
60
min_variation(%)
Figure 5 shows the relations between the number of data transmissions and the
The number of data transmissions
maximum length of the window when the constant min_variation is fixed to 25%. 5500 5000 4500 4000 3500 3000 2500 2000 0
5
15
20
25
30
35
Maximum length of window
40
50
100
Figure 6 shows the relations between the data coherence and the constant min_variation when the maximum length of the window is fixed to 25.
Data coherence(%)
10 9 8 7 6 5 0
10
20
30
40
50
60
min_variation(%)
Figure 7 shows the relations between the data coherence and the maximum length of the window when the parameter min_variation is fixed to 25%. 11
Data cohrehence(%)
10 9 8 7 6 5 4 0
5
15
20
25
30
35
Maximum length of window
40
50
100
Figure 8 The number of data transmissions for various data transmission strategies
The number of data transmissions
5700 5400 5100 4800 4500 4200 3900 3600 3300 3000 DPP
WEDP
WHP
SP
PEP
ACTC
P&P
Figure 9 The data coherence for various data transmission strategies 9
Data cohrehence(%)
8 7 6 5 4 3 2 1 0 DPP
WEDP
WHP
SP
PEP
ACTC
P&P
S0030-4026(16)30387-4 http://dx.doi.org/doi:10.1016/j.ijleo.2016.04.114 IJLEO 57597
To appear in: Received date: Accepted date:
3-3-2016 20-4-2016
Please cite this article as: Xiaobo Ji, Fan Zeng, Mingwei Lin, Data Transmission Strategies for Resource Monitoring in Cloud Computing Platforms, Optik - International Journal for Light and Electron Optics http://dx.doi.org/10.1016/j.ijleo.2016.04.114 This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
Data Transmission Strategies for Resource Monitoring in Cloud Computing Platforms
Xiaobo Ji1, Fan Zeng2, Mingwei Lin3 1
Department of Information, Research Institute of Field Surgery, Daping Hospital, Third Military Medical University, Chongqing, China
2
Department of Information, Research Institute of Field Surgery, Daping Hospital, Third Military Medical University, Chongqing, China 3
Faculty of Software, Fujian Normal University, Fuzhou 350007, China 1
[email protected], [email protected], [email protected]
Abstract This paper proposes three data transmission strategies for resource monitoring in cloud computing platforms, which are dynamic periodic push strategy, window-based event-driven push strategy, and window-based hybrid push strategy, respectively. The dynamic periodic push strategy periodically pushes resource status information at a dynamic time interval that is dynamically updated by considering the change degree of resource status information. The window-based event-driven push strategy pushes resource status information if the change degree of resource status information is larger than a threshold that is dynamically updated by using the exponentially weighted moving average. The window-based hybrid push strategy combines the dynamic periodic push strategy and the window-based event-driven push strategy and then further reduces the communication overhead and improves the data coherence. Experimental results show that the window-based hybrid push strategy performs better than previous data transmission strategies in terms of the number of data transmissions and the data coherence.
Key words: data transmission strategy, cloud computing platform, resource monitoring
1. Introduction Cloud computing is a new computing model that is based on the Internet network [1]. Compared with traditional IT infrastructure, cloud computing platforms introduce the virtualization technology to build resource pools and can deal with the problem that
the traditional IT infrastructure shows low hardware resource utilization and long application deployment time [2]. Therefore, more and more enterprises choose to move their applications to cloud computing platforms [3]. As applications deployed in cloud computing platforms increase in type and number, scales of cloud computing platforms are increasingly expanding and cloud computing platforms become more and more complex [4], [5], [6]. Moreover, virtual machines in cloud computing platforms share hardware resources [7]. These make cloud computing platforms be prone to error [8]. Resource monitoring can help system administrators understand running status of cloud computing platforms and system administrators can take precautions before cloud computing platforms fail to run [9][10]. The data transmission strategy is an important part of resource monitoring and has a significant influence on the performance of resource monitoring in terms of the communication overhead and the data coherence [11]. Existing data transmission strategies cannot achieve a better trade-off between the low communication overhead and high data coherence. In this case, three mechanisms were proposed, which are the offset-sensitive mechanism (OSM), the time-sensitive mechanism (TSM), and the announcing with change and time consideration (ACTC), respectively [12]. In these three mechanisms, the time interval is dynamically updated as the average of time intervals between resource status information changes and the threshold is dynamically updated as the average amount of changes for resource status information. However, the time interval does not consider the change degree of resource status information and the threshold does not consider the change trend of
resource status information. Therefore, the time interval and the threshold in these three mechanisms may be too small or too large and then performances of these three mechanisms will be affected. In this paper, three data transmission strategies are proposed, which are the dynamic periodic push strategy, window-based event-driven push strategy, window-based hybrid push strategy, respectively. The dynamic periodic push strategy periodically pushes resource status information at a time interval that is dynamically updated by considering the change degree of resource status information. The window-based event-driven push strategy pushes resource status information if the change degree of resource status information is larger than a threshold that is dynamically updated by using the exponentially weighted moving average. The window-based hybrid push strategy combines the advantages of the dynamic periodic push strategy and the window-based event-driven push strategy and then further reduces the communication overhead and improves the data coherence. Experimental results show that the window-based hybrid push strategy performs better than previous data transmission strategies in terms of the number of data transmissions and the data coherence. The remainder of this paper is organized as follows. Section 2 briefly reviews existing works on data transmission strategy. Section 3 shows our proposed data transmission strategies. Section 4 presents experimental results. Finally, conclusions are drawn in Section 5.
2. Existing works on data transmission strategy
In large-scale distributed networks, there are two basic data transmission strategies, which are the push strategy and the pull strategy [13] [14]. According to different trigger conditions, the push strategy can be divided into the periodic push strategy and the event-driven push strategy, respectively. For the periodic push strategy, monitored nodes periodically send their resource status information to the monitoring node at a time interval. For the event-driven push strategy, monitored nodes send their resource status information to the monitoring node if the change degree of resource status information is larger than a threshold. The pull strategy can also be divided into the periodic pull strategy and the event-driven pull strategy, respectively. For the periodic pull strategy, the monitoring node periodically sends requests for obtaining resource status information to monitored nodes at a time interval and monitored nodes send their resource status information to the monitoring node after receiving requests. For the event-driven pull strategy, the monitoring node sends requests for obtaining resource status information to monitored nodes if the change degree of the resource status information is larger than a threshold and then monitored nodes send their resource status information to the monitoring node after receiving requests. The time interval and the threshold have a significant influence on performances of the push strategy and the pull strategy. If the time interval or the threshold is too small, useless resource status information will be collected and then the communication overhead will be increased. If the time interval or the threshold is too large, important resource status information will be lost and then the data coherence will be lowered. In order to reduce the communication overhead and improve the data coherence, a number of
data transmission strategies have been proposed. R. Sundaresan et al. proposed three modified pull strategies, which are the regular polling strategy, the adaptive polling strategy, and the slacker polling strategy [15][16]. In the regular polling strategy, the time interval is a constant and cannot change with the change of resource status information. In the adaptive polling strategy, the time interval can be dynamically adjusted by using a damping factor. However, the damping factor is a constant and then the time interval cannot change with the change degree of resource status information. In the slacker polling strategy, the moving average estimator is introduced to estimate the next time interval. All these three data transmission strategies cannot pull resource status information during the time interval. F. F. Han et al. proposed a periodically and event-driven push (PEP) algorithm, which is based on the push strategy [17]. In the PEP algorithm, monitored nodes periodically send their resource status information to the monitoring node at a time interval of 1 second. If the change degree of resource status information is larger than a predefined threshold, monitored nodes also send their resource status information to the monitoring node. The PEP algorithm could reduce the communication overhead and improve the data coherence. However, the time interval and the threshold in the PEP algorithm are constants and cannot change with the change degree resource status information. Therefore, the time interval and the threshold in the PEP algorithm may be too small or too large. W. C. Chung and R. S. Chang proposed three data transmission mechanisms for grid
computing, which are the offset-sensitive mechanism (OSM), the time-sensitive mechanism (TSM), and the announcing with change and time consideration mechanism (ACTC) [12]. All these three mechanisms are based on the push strategy. The OSM is an event-driven push strategy. In the OSM, monitored nodes send their resource status information to the monitoring when the change degree of resource status information is larger than a threshold. The threshold is initialized to 0 and then the threshold is dynamically updated as the average amount of changes for resource status information. In the TSM, monitored nodes periodically send their resource status information to the monitoring node at a time interval. The time interval is calculated as the average of time intervals between resource status information changes. The ACTC combines the advantages of the OSM and TSM strategies and effectively improves the performance of data transmission. However, the time interval does not consider the change degree of resource status information and the threshold does not consider the change trend of resource status information. Therefore, the time interval and the threshold in these three mechanisms may be too small or too large. H. Huang and L. Wang proposed a combined push-pull model named P&P for cloud computing environment [18]. The P&P model consists of the P&P-Push algorithm and the P&P-Pull algorithm, respectively. The P&P-Push algorithm runs on monitored nodes, while the P&P-Pull algorithm runs on the monitoring node. The P&P model could intelligently switch between the P&P-Push algorithm and the P&P-Pull algorithm according to users’ requirements and resource status. It can reduce the communication overhead and improve the data coherence.
3. Proposed data transmission strategies This section presents our proposed data transmission strategies, which are the dynamic periodic push strategy, the window-based event-driven push strategy, and the window-based hybrid push strategy. 3.1 Dynamic periodic push strategy In order to address the problem that the time interval in existing periodic push strategies does not considering the change degree of resource status information, a dynamic period push strategy is proposed. In the proposed dynamic period push strategy, a constant denoted as min_variation is defined to determine whether resource status information during the time interval changes greatly. If the resource status information during the time interval changes and its change degree is larger than min_variation, the time interval (TI) is updated as TI TI
min_variation A B
(1)
where A is the value of resource status information after change and B is the value of resource status information before change. When the timer expires, monitored nodes send their resource status information to the monitoring node. If the change between resource status information sent now and resource status information sent last time is less than min_variation, the time interval is updated as
1 TI TI 1 C P e
(2)
where C is the value of resource status information sent now and P is the value of resource status information sent last time. Figure 1 Algorithm of the dynamic periodic push strategy 3.2 Window-based event-driven push strategy In order to deal with the problem that the threshold in existing event-driven push strategies does not consider the change trend of resource status information, a window-based event-driven push strategy is proposed. Definition 1 Resource status information window, a triplet
Lmax , L, d _ threshold , is
used to store resource status information. where Lmax and L are the maximum length and current length of the window. The term d _ threshold is the threshold of the change of resource status information in the window and its initial value is 0. The core idea of the window-based event-driven push strategy is to introduce the exponentially weighted moving average to predict the value of d _ threshold . The exponentially weighted moving average is designed from the moving average and calculates the prediction value by using the following formula: E t O t 1 1 E t 1
(3)
where E t is the prediction value at time period t and O t 1 is the observation value at time period t 1 . The coefficient is a constant smoothing factor between 0 and 1. Assume that the resource status information window has a resource status information data series S1 , S 2 , S3 ,, S L . If a piece of new resource status information denoted as
S L 1 is stored into the resource status information window, the threshold is calculated
as S L 1 S L 1 d _ threshold L d _ threshold L 1 S 2 S1 0
, L 1 , L 1 , L0
(4)
where d _ threshold L 1 is the prediction value of the threshold after storing the new resource status information into the resource status information window. Figure 2 Algorithm of the window-based event-driven push strategy 3.3 Window-based hybrid push strategy The time interval in the dynamic push strategy changes with the change degree of resource status information and then the dynamic push strategy could reduce the transmission of useless resource status information and the loss of important resource status information. Therefore, the dynamic push strategy could reduce the communication overhead with relatively high data coherence. However, the dynamic push strategy cannot push the important resource status information during the time interval. The window-based event-driven push strategy can be aware of the change trend of resource status information by introducing the exponentially weighted moving average to predict the threshold. Resource status information changes significantly recently and then the threshold is very high. If resource status information changes slightly at the next time period, the change of resource status information will be less than the threshold and then resource status information will not be pushed at this time period. In this case, collected resource status information will be insufficient and the
performance of resource monitoring will be lowered. The dynamic periodic push strategy can address the problem that the window-based event-driven push strategy may not push resource status information for a long time, while the window-based event-driven push strategy can address the problem that the dynamic periodic push strategy cannot push important resource status information during the time interval. In order to address problems in the dynamic periodic push strategy and the window-based event-driven push strategy, this paper combines the dynamic periodic push strategy and the window-based event-driven push strategy and proposes a window-based hybrid push strategy. Figure 3 Algorithm of the window-based hybrid push strategy
4. Experimental analysis
In order to evaluate three proposed data transmission strategies, a series of experiment have been conducted on a private cloud computing platform. This section shows the experiment environment and then analyzes the experimental results. 4.1 Experiment environment The experiment platform consists of six physical nodes, which are connected by using the 100M Ethernet. One is selected to be a cloud controller, while the rest are used as compute nodes. A private cloud computing platform is deployed on these six physical nodes by using the Xen [19] and OpenStack open source software [20]. Resource monitoring is performed by collecting resource parameters, which can be obtained from the /proc file in the operating system. In order to simplify experiments, only one
of resource parameters, i.e. the CPU utilization, is used to evaluate the performance of our proposed data transmission strategies [21]. In order to evaluate the performance of our proposed data transmission strategies, the dynamic periodic push (DPP) strategy, the window-based event-driven push (WEDP) strategy, and the window-based hybrid push (WHP) strategy are compared with the slacker polling strategy (SP), the periodically and event-driven push (PEP) algorithm, the announcing with change and time consideration (ACTC) mechanism, and the combined push-pull (P&P) model. Evaluation metrics in our experiments are the number of data transmissions and data coherence. The number of data transmissions is the sum of the number of pull operations and the number of push operations. The data coherence (coh) is calculated as n
coh
C i R i
2
i 1
(5)
n
where C i and R i are the point on the graph of actual measurements and the point on the graph of generated measurements, respectively. The higher the data coherence is, the smaller the value of coh is. 4.2 Experimental results We have conducted six groups of experiments on the private cloud computing platform. The first four experiments are conducted to test whether the performance of the window-based hybrid push strategy depends on the constant min_variation and the maximum length of the window. The rest experiments are performed to compare our proposed data transmission strategies with existing data transmission strategies.
The parameter Push_interval in the P&P model was set to 1s. The parameters PULL_INIT_INTERVAL, PULL_INTERVAL_MIN, and PULL_INTERVAL_MAX were set to 5s, 3s, and 12s, respectively. (1) Relations between the number of data transmissions and min_variation Figure 4 Relations between the number of data transmissions and min_variation For the window-based hybrid push strategy, the number of data transmissions reduces with the increase of the value of the min_variation when the maximum length of the window is fixed. The number of data transmissions reduces most significantly when the min_variation is equal to 25%. If the min_variation is larger than 50%, the number of data transmissions will tend to be stable. (2) Relations between the number of data transmissions and the maximum length of the window Figure 5 Relations between the number of data transmissions and maximum length of the window As shown in Figure 5, the number of data transmissions depends on the maximum length of the windows when the constant min_variation is fixed. As the maximum length of the window increases, the number of data transmissions decreases. The number of data transmissions tends to be stable when the maximum length of the window is larger than 50. (3) Relations between the data coherence and min_variation Figure 6 Relations between the data coherence and min_variation As shown in Figure 6, the data coherence depends on the parameter min_variation. As
the parameter min_variation increases, the data coherence decreases. The data coherence tends to be stable when the parameter min_variation is larger than 45%. (4) Relations between the data coherence and the maximum length of the window Figure 7 Relations between the data coherence and maximum length of the window As shown in Figure 7, as the maximum length of the window increases, the data coherence decreases when the parameter min_varition is fixed. When the maximum length of the window increases, the threshold increases and then some important resource status information are lost during the time interval. Therefore, the data coherence decreases. When the maximum length of the window is larger than 35, the data coherence tends to be stable. (5) Comparison of the number of data transmissions As shown in Figure 8, the number of data transmissions of the proposed DPP and WEDP strategies is more than the PEP algorithm, the ACTC mechanism, and the P&P model. This is because that the time interval in the proposed DPP strategy and the threshold in the proposed WEDP strategy may be too small. The number of data transmissions of the proposed WHP strategy is less than other data transmission strategies. This is because that its time interval can be dynamically updated by considering the change degree of resource status information and its threshold can be dynamically updated by using the exponentially weighted moving average. Therefore, the time interval and the threshold can change with resource status information and will not be too small or too large. (6) Comparison of the data coherence
As shown in Figure 9, the data coherence of the proposed DPP and WEDP strategies is lower than other data transmission strategies, but lower than 10% and can satisfy the basic requirement of resource monitoring. The proposed WHP strategy shows the highest data coherence. This is because the WHP strategy can push important resource status information during the time interval. Moreover, its time interval and threshold can change with resource status information and will not be too small or too large. The time interval and the threshold in the PEP algorithm are constants, so they will be too large when resource status information changes frequently and significantly. The time interval in the ACTC mechanism is simply calculated as the average of time intervals between resource status information changes and its threshold is simply calculated as the average of change degree for resource status information. The time interval in the ACTC mechanism cannot be aware of the change degree of resource status information and its threshold cannot be aware of the change trend of resource status information, so the time interval and the threshold in the ACTC mechanism may be too large. The parameters in the P&P model are constants, so they cannot change with resource status information and they may be too large. Large time interval and threshold will result in missing important resource status information and reduce data coherence.
V. Conclusions
Three data transmission strategies are proposed in this paper, which are the dynamic periodic push strategy, the window-based event-driven push strategy, and the
window-based hybrid push strategy. The dynamic periodic push strategy periodically pushes resource status information at a time interval that is dynamically updated by considering the change degree of resource status information. The window-based event-driven push strategy pushes resource status information when the change degree of resource status information is larger than a threshold that is predicted by using the exponentially weighted moving average. The window-based hybrid push strategy combines the advantages of the dynamic periodic push strategy and the window-based event-driven push strategy. A series of experiments have been conducted on a private cloud computing platform built by using the Xen and OpenStack open source software. Experimental results show that our proposed window-based hybrid push strategy outperforms previous data transmission strategies.
Acknowledgments
The work of this paper is supported by the Natural Science Foundation Project of CQ CSTC under Grant No. cstc2013jcyjA40059, National Natural Science Foundation of China under Grant No. 61502102, No. 61370078, and No. 61402109, Fujian Province Education Scientific Research Projects for Young and Middle-aged Teachers under Grant No. JA15122, National Undergraduate Training Programs for Innovation and Entrepreneurship under Grant No. 201510394021, and Fujian Normal University Undergraduate Training Programs for Innovation and Entrepreneurship under Grant No. cxxl-2015163.
References
[1] Ronnie D. Caytiles, Sunguk Lee, Byungjoo Park. Cloud computing: The next computing paradigm. International of Journal of Multimedia and Ubiquitous Engineering, vol. 7, no. 2, pp. 297-302, 2012. [2] Chi Yuan, Honglei Yu, Deli Yang. Cost comparison between cloud computing and traditional data center. WIT Transactions on Modelling and Simulation, vol. 60, pp. 733-740, 2014. [3] Leire Orue-Echevarria, Juncal Alonso, Hugo Bruneliere, Andreas Menychtas, Philip Langer, Manuel Wimmer. Cloudifying applications with ARTIST a global modernization approach to move applications to cloud. Proceedings of the 4th International Conference on Cloud Computing and Services Science, pp. 737-745, 2014. [4] H. S. Pannu, J. Liu, Q. Guan, S. Fu. AFD: Adaptive failure detection system for cloud computing infrastructures. 2012 IEEE 31st International Performance Computing and Communications Conference, pp. 71-80, 2012. [5] H. S. Pannu, J. Liu, S. Fu. AAD: Adaptive anomaly detection system for cloud computing infrastructures. Proceedings of 2012 31st International Symposium on Reliable Distributed Systems, pp. 396-397, 2012. [6] Q. Guan, S. Fu, N. D. Bardeleben, S. Blanchard. Exploring time and frequency domains for accurate and automated anomaly detection in cloud computing systems. Proceedings of 2013 IEEE 19th Pacific Rim International Symposium on Dependable Computing, pp. 196-205, 2013.
[7] Fei Xu, Fangming Liu, Hai Jin, Athanasios V. Vasilakos. Managing performance overhead of virtual machines in cloud computing: A survey, state of the art, and future directions. Proceedings of the IEEE, vol. 102, no. 1, pp. 11-31, 2014. [8] M. Gupta, A. B. Sharma, H. Chen, G. Jiang. Context-aware time series anomaly detection for complex systems. 2013 Workshop on Data Mining for Service and Maintenance, pp. 14-22, 2013. [9] Junwei Ge, Bo Zhang, Yiqiu Fang. Research on the resource monitoring model under cloud computing environment. Lecture Notes in Computer Science, vol. 6318 LNCS, pp. 111-118, 2010. [10] Chao-Tung Yang, Wen-Chung Shih, Chih-Lin Huang. Implementation of a distributed data storage system with resource monitoring on cloud computing. Lecture Notes in Computer Science, vol. 7296 LNCS, pp. 64-73, 2012. [11] Kai Lin, Weiqin Tong, Xiaodong Liu, Liping Zhang. A self-adaptive mechanism for resource monitoring in cloud computing. IET International Conference on Smart and Sustainable, pp. 178-182, 2013. [12] Wu-Chun Chung, Ruay-Shiung Chang. A new mechanism for resource monitoring in Grid computing. Future Generation Computer Systems, vol. 25, no.1, pp. 1-7, 2009. [13] M. L. Massie, B. N. Chun, D. E. Culler. The ganglia distributed monitoring system: Design, implementation, and experience. Parallel Computing, vol. 30, no. 7, pp. 817-840, 2004. [14] E. Imamagic, D. Dobrenic. Grid infrastructure monitoring system based on
Nagios. Proceedings of the 2007 Workshop on Grid Monitoring, pp. 23-28, 2007. [15] R. Sundaresan, M. Lauria, T. Kurc, S. Parthasarathy, J. Saltz. Adaptive polling of grid resource monitors using a slacker coherence model. Proceedings of IEEE International Symposium on High Performance Distributed Computing, pp. 260-269, 2003. [16] R. Sundaresan, T. Kurc, M. Lauria, S. Parthasarathy, J. Saltz. A slacker coherence protocol for pull-based monitoring of on-line data sources. Proceedings of 3rd IEEE/ACM International Symposium on Cluster Computing and the Grid, pp. 250-257, 2003. [17] F. F. Han, J. J. Peng, W. Zhang, Q. Li, J. D. Li, Q. L. Jiang, Q. Yuan. Virtual resouce monitoring in cloud computing. Journal of Shanghai University, vol. 15, no. 5, pp. 381-385, 2011. [18] H. Huang, L. Wang. P&P: A combined push-pull model for resource monitoring in cloud computing environment. Proceedings of 2010 IEEE 3rd International Conference on Cloud Computing, pp. 260-267, 2010. [19] P. Barham, B. Dragovic, K. Fraser, S. Hand, T. Harris, A. Ho, R. Neugebauer, L. Pratt, A. Warfield. Xen and the art of virtualization. Operating Systems Review (ACM), vol. 37, no. 5, pp. 164-177, 2003. [20] A. Corradi, M. Fanelli, L. Foschini. VM consolidation: A real case based on OpenStack Cloud. Future Generation Computer Systems, vol. 32, no. 1, pp. 118-127, 2014. [21] Mingwei Lin, Zhiqiang Yao, Tianqiang Huang. A hybrid push protocol for
resource monitoring in cloud computing platforms. Optik - International Journal for Light and Electron Optics, vol. 127, no. 4, pp. 2007-2011, 2016.
Figure 1 shows the algorithm of the proposed dynamic periodic push strategy. Algorithm 1: Dynamic Periodic Push Strategy Input: min_variation; N = 0; Tc presents the current time; Output: None 1: Push current resource status information to the monitoring node; 2: Timer = ∞; 3: T0 = Tc; 4: WHILE(TRUE) 5: IF The Timer expires 6: Push current resource status information to the monitoring node; 7: IF |C-P| < min_variation 8: Calculate the new TI using the equation (2); 9: ENDIF 10: Reset Timer to TI; 11: ENDIF 12: IF The resource status information changes 13: IF N = 0 14: Push current resource status information to the monitoring node; 15: TI = Tc – T0; 16: N = N + 1; 17: Reset Timer to TI; 18: ELSE IF |A-B| > min_variation 19: Calculate the new TI using the equation (1); 20: ENDIF 21: ENDIF 22:ENDWHILE
Figure 2 shows the algorithm of the window-based event-driven push strategy. Algorithm: Window-based Event-driven Push Strategy Input: Lmax; L = 0; Output: None 1: Push resource status information to the monitoring node; 2: Store one copy of pushed resource status information in the window; 3: Update the threshold of the window by using the equation (4); 4: L = L + 1; 5: WHILE(TRUE) 6: IF Resource status information changes and |A-B| > d_thresholdL 7: Push current resource status information to the monitoring node; 8: IF L = Lmax 9: Remove S1 from the window; 10: L = L – 1; 11: ENDIF 12: Store one copy of pushed resource status information in the window; 13: Update the threshold of the window by using the equation (4); 14: L = L + 1; 15: ENDIF 16: ENDWHILE
Figure 3 shows the algorithm of the window-based hybrid push strategy. Algorithm: Window-based Hybrid Push Strategy Input: min_variation; N = 0; Tc presents the current time; Lmax; L = 0; Output: None 1: Push current resource status information to the monitoring node; 2: Timer = ∞; 3: T0 = Tc; 4: Store one copy of pushed resource status information in the window; 5: Update the threshold of the window by using the equation (4); 6: L = L + 1; 7: WHILE(TRUE) 8: IF The Timer expires 9: IF |C-P| > min_variation 10: Push current resource status information to the monitoring node; 11: IF L = Lmax 12: Remove S1 from the window; 13: L = L -1; 14: ENDIF 15: Store one copy of pushed resource status information in the window; 16: Update the threshold of the window by using the equation (4); 17: ELSE 18: Calculate the new TI using the equation (2); 19: ENDEIF 20: Reset Timer to TI; 21: ENDIF 22: IF resource status information changes 23: IF N = 0 24: Push current resource status information to the monitoring node; 25: TI = Tc – T0; 26: N = N + 1; 27: Store one copy of pushed resource status information in the window; 28: Update the threshold of the window by using the equation (4); 29: L = L + 1; 30: Reset Timer to TI; 31: ELSE IF |A-B| > d_thresholdL 32: Push current resource status information to the monitoring node; 33: Calculate the new TI using the equation (1); 34: IF L = Lmax 35: Remove S1 from the window; 36: L = L - 1; 37: ENDIF 38: Store one copy of pushed resource status information in the window; 39: Update the threshold of the window by using the equation (4); 40: L = L + 1; 41: ENDIF 42: ENDIF 43:ENDWHILE
Figure 4 illustrates the relations between the number of data transmissions and the
The number of data transmissions
min_variation when the maximum length of the window is fixed to 25. 4500 4000 3500 3000 2500 2000 0
10
20
30
40
50
60
min_variation(%)
Figure 5 shows the relations between the number of data transmissions and the
The number of data transmissions
maximum length of the window when the constant min_variation is fixed to 25%. 5500 5000 4500 4000 3500 3000 2500 2000 0
5
15
20
25
30
35
Maximum length of window
40
50
100
Figure 6 shows the relations between the data coherence and the constant min_variation when the maximum length of the window is fixed to 25.
Data coherence(%)
10 9 8 7 6 5 0
10
20
30
40
50
60
min_variation(%)
Figure 7 shows the relations between the data coherence and the maximum length of the window when the parameter min_variation is fixed to 25%. 11
Data cohrehence(%)
10 9 8 7 6 5 4 0
5
15
20
25
30
35
Maximum length of window
40
50
100
Figure 8 The number of data transmissions for various data transmission strategies
The number of data transmissions
5700 5400 5100 4800 4500 4200 3900 3600 3300 3000 DPP
WEDP
WHP
SP
PEP
ACTC
P&P
Figure 9 The data coherence for various data transmission strategies 9
Data cohrehence(%)
8 7 6 5 4 3 2 1 0 DPP
WEDP
WHP
SP
PEP
ACTC
P&P