Journal of Computational Information Systems 9: 18 (2013) 7389 7396 Available at http://www.jofcis.com A Heuristic Location Selection Strategy of Virtual Machine Based on the Residual Load Factor Gaochao XU 1,2, Yan DING 1, Xu XU 1, Yushuang DONG 1, Jia ZHAO 1, Yunmeng DONG 1, 1 College of Computer Science and Technology, Jilin University, Changchun 130012, China 2 Symbol Computation and Knowledge Engineer of Ministry of Education, Jilin University, Changchun 130012, China Abstract Live VM (virtual machine) migration strategy has become a research hotspot in the field of green cloud data center based on virtualization technology. In view of the fact that the physical hosts of the current most data centers are under a heavy load and thus the performance of data centers declines, this paper has done some research work on location selection strategy of live VM migration. A heuristic approach HB-LR based on residual load factor is proposed in this paper. Its main idea consists of two parts: one is that it combines a heuristic idea with live VM migration to achieve a live VM migration strategy with global search ability. The other is that it has simulated the proposed problem as a bin packing and aimed to search out the most suitable target host for each VM. The final experimental results show that: compared with random migration, HB-LR has better load balancing effect in cloud date centers, significantly reduces the total incremental energy consumption while optimizing the data centers service performance as well as make it have more green and high-efficient data center operations. Keywords: Data Center; Virtualization; Heuristic; Service Performance; Residual Load Factor 1 Introduction Cloud computing [1] is the most promising and valuable research direction in the field of distributed computing currently. Cloud computing provides users with the platform of infrastructure and software service as well as provides service to users demand via the Internet. The infrastructure of cloud service is the cloud data center and most hosts of the cloud data center are having an overweight load, resulting in a decline of computing efficiency of data centers. At present, energy saving and high performance computing of cloud data centers are hotspots. In order to achieve high performance computing and energy saving of cloud data centers, it is essential to select a best location for VM in the process of live migration [2]. Selecting the best location for VM can improve the capacity of a physical host, decrease energy consumption of data centers, reduce Corresponding author. Email address: zhaiyj049@sina.com (Yunmeng DONG). 1553 9105 / Copyright 2013 Binary Information Press DOI: 10.12733/jcisP0166 September 15, 2013
7390 G. Xu et al. /Journal of Computational Information Systems 9: 18 (2013) 7389 7396 the maintenance costs of cloud data centers and improve the efficiency of the cloud data centers. As a result, its service quality and speed with which cloud computing provides will be improved by increasing computational efficiency of the cloud data centers. However, the consideration on various methods of location selection of live VMs migration is not quite perfect in the current research. They can t be sure that every VM is migrated to the best target host. In order to further optimize the location selection process of live VMs migration, this paper has proposed a heuristic approach HB-LR to seek for the physical host with the largest residual load factor. The weight values are to be obtained through the heuristic idea, i.e. the physical host with the biggest residual load factor. Then by comparing the weight values, according to the bin packing model, the migrant VM is migrated to the physical host with larger weights to achieve the balancing of load and improve the resource utilization and computing performance of the cloud data center as well as make the energy consumption more efficient. The rest parts of this paper are as follows. In the second part, we introduce the related work of VM migration and location selection approaches in brief. In the third part, the prerequisite of the algorithm is pointed out at first, and then the design and implementation of our algorithm are introduced in detail. In the fourth part, the experiment and its result are given to evaluate the proposed HB-LR algorithm. In the fifth part, we summarize the full paper and future work is put forward. 2 Related Works The VM migration strategy is the most widely used energy-saving strategy of the cloud computing data centers currently [1]. In [3] Jing Tai Piao et al. put forward an optimal placement and migration approach of VM through network, which reduces the data overhead during VM migration and optimizes the performance of the cloud data centers from an overall perspective. However, this strategy may lead to a low resource utilization of physical host and increase the operation cost of the cloud data center. In [4] Rahman. M et al. put forward a hybrid approach which uses the live migration technology and combines static and dynamic configuration so as to adapt to an excellent initial placement in the changing load characteristics. This will reduce the energy consumption of the cloud center data and speed up the computation efficiency. In [5] Corentin Dupont et al. put forward a kind of framework with flexibility and consciousness of energy conservation to redistribute the VMs in the cloud data center. it computes and formulates the optimal location and the purpose is to reduce the energy consumption and improve the performance of the cloud data center. At the present, many heuristic ideas are proposed to optimize VM migration algorithm for location selection. To achieve energy saving and high efficient computing, a heuristic searching idea is applied to reallocate and integrate the VM in the cloud data center [6, 7, 8, 9]. Lawler. E [10] considers it be an NP problem. That is, no precise solution can be given in polynomial time. Since it has a large solution space, it is generally solved by heuristic ideas. The heuristic idea finds the most suitable physical host for each VM and minimizes the energy consumption as well as achieves the goal of energy saving [11]. The idea of bin packing is also a kind of heuristic ideas. The idea that simultaneously optimizes the VM migration and placement [12] in virtualization heterogeneous systems is proposed by Li Bo. And it uses a heuristic approach which has used a variable box and cost in bin packing. Improved Best Fit Decreasing (BFD) algorithm is another approach to solve the bin packing. In [13] Anton. B et al. have proposed VM management policy based on energy saving and efficiency
G. Xu et al. /Journal of Computational Information Systems 9: 18 (2013) 7389 7396 7391 in cloud computing data centers, using the improved Best Fit Decreasing algorithm to optimize placement of VMs in real time. The placement is ensured to be the best at any time. The resource utilization of physical hosts is improved and the goal of energy saving of cloud data centers is achieved. In [14], Holland J et al. have used the genetic algorithm model to simulate the problem of VM placement. STILLWELL M et al. consider that the hybridization process of combination is operated by exchanging the gene fragments of two chromosomes, while the mutation process is completed by random exchange of VMs on two physical hosts [15]. 3 HB-LR Algorithm Design In the algorithm part of this paper, the main idea of HB-LR is to find the best target host of the VM migration for load balancing. To address this problem, we have presented and utilized the residual load factor of physical host to measure the load of each host. Assume that the migration is under the same network environment, and we don t know the factor of the physical host s residual load. Under this circumstances, we should fully take into consideration the residual load factor of the physical host to design the location selection strategy of live VM migration. As for selecting the host for live VM migration, it can be abstracted as a classic packing problem. In many physical host, looking for the best target host of live VM migration is necessary since not every available physical host is the best location selection. If the load of a VM is very large and the residual load factor of physical host is very small, it can make the load of the physical host overweight, increase energy consumption and result in that the goal of energy conservation will not be reached. Therefore, as for the situation that we cannot be sure which physical hosts residual load factors are highest, we have combined the algorithm of dijkstra to the bin packing, using the ideas of dijkstra algorithm to find the physical host with the biggest residual load factor, which is a weight value that we need to find and then simulate the proposed problem as a bin packing, the VMs are expressed as the balls of different sizes and the physical hosts as the boxes. Putting the balls into the boxes is the process of migrating the VMs into physical hosts. 3.1 Implementation of HB-LR algorithm The implementation of the HB-LR algorithm: The first step: The primary task of migration strategy is to select the best location, the idea of this paper is first to find a weight for each physical host. As the migration does not involve the length of the path within the LAN, the network connection s influence on the migration is eliminated naturally. In this case, we have defined a weight as the residual load factor of the physical host. The second step: Assume that there are m VMs and n physical hosts in the data center, the VM set of a cloud data center is expressed as V = {v 1, v 2,..., v m }, the set of physical hosts is denoted as S = {s 1, s 2,..., s n }, in which m n. Given two sets: C = { } and B = { }, C = { } is expressed as the set which is used for storing hosts temporarily and B = { } means the set of X ij values obtained in each round of processes. The third step: First of all we take out a VM v i from the set V. Since our goal is to find the physical host with the largest residual load factor, we must first obtain the remaining memory L m and the remaining CPU L c of the physical host. And the current state L i (CPU and memory
7392 G. Xu et al. /Journal of Computational Information Systems 9: 18 (2013) 7389 7396 usage) of remaining resource on each physical host is represented as follows, 1 i n. L i = ωl c + φl m (1) ω + φ = 1 (2) The weight values of Memory and CPU are determined by the learning of BP neural network, we set the weight value of Memory φ = 0.4 and the weight value of CPU ω = 0.6. At this point, the residual load of the physical host, L i = 0.6L c + 0.4L m. The sum of the residual load of the data center is: n Q i = L i (3) Where Q i represents the total residual load of the data center, the ratio of the i-th physical host s residual load in total residual load is the residual load factor: i=1 E i = L i /Q i (4) Then retrieve a single physical host s j from the set S of physical hosts and put s j into set C, while obtaining the residual load factor E of the physical host s j. If a physical host s j is turned off, it will be removed from the set C; if works well, the physical host s j+1 will be continue to retrieve from physical host set S and then put into set C. At the same time the residual load factor E j+1 of s j+1 is obtained. Thus, the set C = {s j, s j+1 } get the residual load factor E of the physical host from the formula (2)(3)(4)(5). If E j > E j+1, viz, the residual load factor of the physical host s j is higher than that of physical host s j+1, we will remove s j+1 from the set C, otherwise remove s j. The fourth step: HB-LR continues to take out physical hosts from set S and puts them into set C. Repeat the above process until all the elements in the set S is fetched. When there is only one element in the C = {s k }, the s k is the physical host with the highest residual load factor and it is the best location for the VM v i, i.e., E k > E 1 > E 2... > E k 1 > E k+1... > E n 1 > E n. When we make sure s k is the best location, the VM v i is expressed as a ball and the physical host s j is expressed as a box, then put v i into the box s j. The process is the migration of the VM. At this moment, X ij = 1 and it is put into the set B. The fifth step: After live migration of a VM is completed, we continue to take the next VM from the set V. Repeat the above process. When there are m elements in set B, the migration events of all the VMs in the set V are completed. 3.2 Model of HB-LR algorithm From the macro of view, our algorithm is conform to the dijkstra algorithm since mainly dijkstra algorithm is to calculate the shortest path from a node to all other nodes, just like migration from VM to many physical hosts, and the weight value is the residual load of physical host in our algorithm. First of all, Look for from the starting until find a set of nodes from which to the starting point the path is shortest, after finding a shortest path, the nodes will be added to the set, when all the nodes are added to the set, the shortest path between the starting point each node can be calculated. And this process is that we looking for weight values, we take out a physical host from the set of physical hosts, first get its remaining CPU and memory, and then calculated the residual load factor of the physical host according to formula (2)(3)(4)(5), which
G. Xu et al. /Journal of Computational Information Systems 9: 18 (2013) 7389 7396 7393 we call weight value, numerous physical hosts to be put in the set and then make a comparison. To find a physical host with the highest residual load factor is equal to find a node with minimum weight value. From this view HB-LR is conform to dijkstra algorithm. There are three main algorithms to solve this kind of problem: adaptive algorithm for the First time (First-Fit), optimal adaptive algorithm (Best-Fit) and the Next adaptive algorithm (Next- Fit). Only the best adaptation algorithm has the function of saving the space, using resources with high efficiency, avoiding waste, achieving the goal of energy-saving. The reason for that our algorithm is adapted in the best adaptation algorithm in bin packing is mainly that the best adaptation algorithm takes the resource utilization of box into consideration. Each box is open. The ball is not randomly put into the boxes. One can use the iteration comparison to find the most suitable box and thus improve the space utilization of the box. An efficient placement is completed. First of all, the bin packing is to put the balls into the boxes of different sizes while our migration strategy is to migrate the migrant VMs to physical hosts. In essence there is no difference between the two. When we abstract VMs into balls of different sizes and the physical hosts are abstracted into boxes, we can say that the VM migration process is abstracted as the process of a bin packing. This is the reason for that our proposed problem can be seen as a bin packing. 4 Evaluation In this paper, the CloudSim is applied to stimulate the HB-LR approach. In the simulation experiments, we randomly select 50 physical hosts with the same configuration and 100 VMs with the same performance. The residual CPU and the residual memory of the current physical hosts are obtained. The residual load factor of each physical host are calculated. Then, the migration locations can be decided by the residual load factor of each physical host. We will verify HB-LR strategy by experiments to reflect energy-saving goal through comparing the HB- LR approach and random migration approach. It includes three aspects: Firstly, the degrees of the loading balance of the target hosts are compared after HB-LR migration strategy and after random migration strategy; Secondly, the external service performance of the cloud date center is compared after the two kinds of migration strategy; Thirdly, the energy consumption of the target hosts is compared after the two kinds of migration strategy. The final experimental results have demonstrated that our proposed HB-LR approach is a high-efficient heuristic location selection strategy of live VM migration for load balancing and energy saving. In the first set of experiments, we have verified the feasibility HB-LR from the perspective of system load balancing after VM migration completed. As shown in Figure 1, we can find that when VM is migrated to the physical host, the load on physical host changes. As the migrating time increases, both load balancing degrees decrease and the degree of random migration is larger than that of HB-LR. According to the experimental result, HB-LR has better load balance effect, therefore it indirectly reduces the energy consumption of the cloud date center and enhances the computing power of the cloud data center. In the second set of experiments, we have compared energy consumption in the cloud data center by the two kinds of migration strategies. We analyze the proposed HB-LR approach by statistics of energy consumption of each period in the cloud data center. Figure 2 shows the comparison of random migration and HB-LR in energy consumption. The result indicates that, compared with the random migration strategy, HB-LR has reduced the energy consumption by
7394 G. Xu et al. /Journal of Computational Information Systems 9: 18 (2013) 7389 7396 0.35 0.30 RM HB-LR Load balancing degree (B) 0.25 0.20 0.15 0.10 0.05 0.00 0 200 400 600 800 1000 1200 1400 Time (s) Fig. 1: Comparison of load balancing degree 6%, 10%, 15%, 20% and 33% in the five groups experiments. And as the time increases, the reduction increases. Therefore, in this respect HB-LR is not a short-term local optimal solution but an energy-saving program from a long-term perspective. 40 RM HB-LR 35 Power consumption (kw/h) 30 25 20 15 10 5 0 0 2 4 6 8 10 12 Time (h) Fig. 2: Comparison of energy consumption In the third set of experiments, we have verified HB-LR by evaluation of the external service performance of the cloud data center after the two kinds of migration strategies complete. We chooses the throughput as the evaluation criteria as the throughput is usually the overall evaluation of the ability of a system. The result is showed in Figure 3, where the external service performance of the cloud data center is different by using two different migration strategies. After random migration, the cloud data center shows a higher performance when the users have access to the cloud data center. However, with the response time increasing, the external service performance of the cloud data center is in a waving way, which is not stable. After using the HB-LR approach, although the service performance is not as high as the random migration strategy at first, as the response time increases, the external service performance of the cloud data center be-
G. Xu et al. /Journal of Computational Information Systems 9: 18 (2013) 7389 7396 7395 comes gradually stabilized. By the comparison of performance of the cloud data center s external service, it can be seen that HB-LR has better stability and efficiency. 6.0 5.4 RM HB-LR Throughoutput (req/s) 4.8 4.2 3.6 3.0 2.4 1.8 1.2 0.6 0.0 0 200 400 600 800 1000 1200 1400 Time (s) Fig. 3: Comparison of external service performance 5 Conclusion and Future Work On the basis of summarizing the relative work, a new live virtual migration and location selection strategy HB-LR is proposed. The main idea, process achievement and evaluation are given. It uses a heuristic idea that is based on residual load factor. We combined the idea of dijkstra algorithm and the idea of bin packing, use the dijkstra algorithm to find the physical host with the highest residual load factor and stimulate the bin packing to migration VMs to the target host, in order to achieve the HB-LR algorithm and the search for a global optimal solution. This paper shows how the experiment verified the HB-LR algorithm from the degree of load balancing, external service performance of the data center and energy consumption. The results indicate that HB-LR can solve the problem that part of the cloud data centers hosts are overloaded and cause decline in computing performance, and HB-LR can find the best location for VM and ensure that the cloud data center have a load balancing to some extent. HB-LR has achieved a green, efficient cloud data center and stable performance of the data centers external service. It not only improves the external service performance of data center, and also minimizes the power consumption of data centers comparably. It aims to achieve more energy-saving during the longterm operation of a cloud data center. There are some open problems needing further study and some empirical problems need to more experiments to get a better solution. The value of φ and ω in the residual load rate is an empirical issue which needs more experiment to obtain optimal values to fit φ + ω = 1. In order to further improve the performance of HB-LR, we plan to study the robustness of HB-LR in the next step. HB-LR method should have ability in achieving that multiple VMs find the best migration locations at the same time and maintaining the stability of the performance during the VM migration.
7396 G. Xu et al. /Journal of Computational Information Systems 9: 18 (2013) 7389 7396 References [1] M. Armbrust, A. Fox, R. Griffith, A. Joseph, R. Katz, A. Konwinski, G. Lee, D. Patterson, A. Rabkin, I. Stoica and M. Zaharia, Above the Clouds: A View of Cloud Computing, Technical Report EECS-2009-28, 2009. [2] P. Barham, B. Dragovic, K. Fraser, S. Hand, T. Harris, A. Ho, R. Neugebauer, I. Pratt and A. Warfield, Xen and the Art of Virtualization, in Proc. of the 19th ACM Symposium on Operating Systems Principles, pp. 164-177, 2003. [3] Jing Tai Piao, Jun Yan. A network-aware virtual machine placement and migration approach in cloud computing [C], ICGCC9th. NanJing: GCC, 2010: 87-92. [4] Rahman, Mahfuzur, Graham, Peter. Hybrid resource provisioning for clouds. 2012 J. Phys.: Conf. Ser. 385 012004. http://iopscience.iop.org/1742-6596/385/1/012004. [5] Corentin Dupont, Giovanni Giuliani, Fabien Hermenier, Thomas Schulze, Andrey Somov. 2012. An Energy Aware Framework for virtual machine Placement in Cloud Federated Data Centres. Proceedings of the 3rd International Conference on Future Energy Systems. [6] Quan, D.-M., Basmadjian, R., De Meer, H., Lent, R., Mahmoodi, T., Sannelli, D., Mezza, F., Dupont, C. 2011. Energy efficient resource allocation strategy for cloud data centres. In Proceedings of the 26th International Symposium on Computer and information Sciences (London, UK, September 26-28, 2011). ISCIS 11. Springer, 133-141. [7] Ajiro Y, Tanaka A. Improving packing algorithms for server consolidation. Proceedings of the 33rd International Computer Measurement Group Conference. San Diego, 2007: 399-406. [8] Gupta R,Bose S.K,Sundarrajan S et al. A two stage heuristic algorithm for solving server consolidation problem with item-item and bin-item incompatibility constraints. Proceedings of the 2008 IEEE International Conference on Services Computing (SCC 08). Hawaii, 2008: 39-46. [9] Agrawal S, Bose S K, Sundarrajan S.K, Sundarrajan S. Grouping genetic algorithm for solving the server consolidation with conflicts. Proceedings of the 1st ACM/SIGEVO Summit Genetic and Evolutionary Computation. New York, 2009: 1-8. [10] Lawler, E. 1983. Recent results in the theory of machine scheduling. In Mathematical Programming: The State of the Art. Springer-Verlag, Berlin, Germany. [11] Coffman J, Garey M R, Johnson D S, Approximation algorithms for bin packing: A survey. Approximation algorithms for NP-Hard problem. PWS Publishing, 1997: 46-93. [12] Li Bo, Li Jianxin, Huai Jinpeng, et al. Enacloud: An Energy-saving Application Live Placement Approach for Cloud Computing Environments[C], Proc. of the 2009 IEEE International Conf. on Cloud Computing. Bangalore, India: IEEE Computer Society, 2009. [13] Anton B, Rajkumar B. Energy Efficient Resource Management in Virtualized Cloud Data Centers [C], Proc. of IEEE/ACM International Conference on Cluster, Cloud and Grid Computing. Melbourne, Australia: IEEE Computer Society, 2010. [14] Holland J. Adaption in Natural and Artificial Systems [M]. Cambridge, MA: mit Press, 1992. [15] STILLWELL M;SCHANZENBACH D;VIVIEN F, et al. Resource allocation algorithms for virtualized service hosting platforms [J]. Journal of Parallel and Distributed Computing, 2010, 70(9): 962-974.