Energy-efficient Virtual Machine Provision Algorithms for Cloud ...

Energy-efficient Virtual Machine Provision Algorithms for Cloud Systems Ching-Chi Lin Pangfeng Liu Institute of Information Science, Department of Computer Science Academia Sinica and Information Engineering, Taipei, Taiwan Graduate Institute of Networking and Multimedia, [email protected] National Taiwan University Taipei, Taiwan [email protected]

Abstract—Power consumption is one of the most critical problems in data centers. One effective way to reduce power consumption is to consolidate the hosting workloads and shut down physical machines which become idle after consolidation. Server consolidation is a NP-hard problem. In this paper, we propose two new algorithms, Dynamic Round-Robin (DRR) and Hybrid, which combines DRR and First-Fit, for energyaware virtual machine scheduling and consolidation. We also propose an accurate power model to estimate the power consumption resulted from each algorithm. Strategies we proposed are compared with GREEDY, ROUNDROBIN and POWERSAVE scheduling strategies implemented in the Eucalyptus Cloud system. Our experiment results show that our DRR and Hybrid algorithms reduce power consumption by 56.4% and 55.9% respectively, compared with the ROUNDROBIN scheduling strategy in Eucalyptus. DDR and Hybrid also result in 3% less power consumption on average, compared with the POWERSAVE scheduling strategy in Eucalyptus. Keywords-Cloud Computing, Data Center, Power Saving, Virtual Machine Consolidation.

I. I NTRODUCTION Power conservation is an important issue for data centers. To reduce power consumption of data centers, one can consolidate computation workloads onto a subset of servers, and power off servers that become idle after consolidation. The key idea is to reduce “idle power”, i.e., the power consumed by idle servers that do not have workload, so as to reduce the number of servers needed to power on. In this paper we propose two new strategies of deploying virtual machine to servers and migrating virtual machines among servers: Dynamic Round-Robin (DRR) and Hybrid. DRR extends the traditional Round-Robin scheduling by incorporating the notion of retirement and retirement time threshold. Hybrid combines DRR and the First-Fit scheduling strategy. In addition, VM migration is used to consolidate server utilization. The migration may be triggered every time a virtual machine finishes. We also propose a power model and a migration model to estimate the power consumption by workload execution and VM migration. For the power model, power consumption is assumed to be a linear function of CPU load, and the

Jan-Jan Wu Institute of Information Science, Academia Sinica Taipei, Taiwan [email protected]

power consumption of an idle CPU is half of the peak power consumed by a fully loaded CPU. This power model has been experimentally verified with real VM workload execution on an eight-node cluster and with actual power consumption measured by power meters. For the migration model, live migration is used in our work. Live migration allows a running virtual machine or application to move between two physical servers without disconnecting the client or application. The power overhead incurred by live migration is very small, which also has been verified by our experiments. These new strategies we proposed are compared with GREEDY, ROUNDROBIN, and POWERSAVE scheduling strategies implemented in the Eucalyptus Cloud system. Our experiment results show that our DRR and Hybrid reduce 56.4% and 55.9% power consumption respectively when compared with the power consumption required by the ROUNDROBIN policy in Eucalyptus. DDR and Hybrid also save 3% more power than the POWERSAVE strategy in Eucalyptus. The rest of this paper is organized as follows. Section II introduces the related works. Section III describes power and migration model used in our simulation. Section IV presents our scheduling algorithms. Section VI describes our experimental results and SectionVII concludes. II. R ELATED W ORKS Various management strategies have been developed to effectively reduce server power consumption. Takedai and Takemura [12] propose an extended First-Fit Decreasing algorithm that uses an unique rank as server priority. Their method consolidates virtual machines on high ranking physical machines, instead of minimizing the number of active physical machines. Hu et al. [7] propose a policy called Magnet, which keeps track of all active nodes in a cluster and organizes them into concentric, non-overlapping rings. This scheme can reduce the power consumption greatly by regarding all cluster nodes as a whole. Mazzucco et al. [9], [10] proposed a dynamic allocation policy aiming at satisfying the conflicting goals of maximizing the quality of

users’ experience while minimizing the cost for the provider. Their model considers limited user patience time and the fact that servers energy consumption depends on servers’ utilization. Liu et al. [8] present the GreenCloud architecture, which enables comprehensive online-monitoring, live virtual machine migration, and VM placement optimization. Chase et al. [1] presents the design and implementation of an architecture for resource management in a hosting center operating system, with an emphasis on energy as a driving resource management issue for large server clusters. The major difference between our work and most existing works is the underlining application workload. Most of the energy saving works in the literature focus on the workload of HTTP requests. The virtual machines running web servers are created to process those requests. The goal is to reduce power consumption by using fewer virtual machines or increasing the load of virtual machines, at an acceptable price of increased response time or less QoS guarantee. In contrast, our work focuses on deploying virtual machines to physical machines. These virtual machines can be associated with most kinds of application workload, or general applications. Therefore, the workload considered in our work is virtual machines, not individual HTTP requests. The goal is to deploy these virtual machines to cores so that various performance indices can be satisfied. III. S YSTEM M ODELS The power consumption model and the migration model in our VM provision framework is defined in this section. The power consumption model measures power usage of a physical machine, and the migration model describes the migration mechanism and measures its incurred cost. A. Power Consumption Model The power consumption of a physical machine is assumed to be a function of its CPU load. Previous studies [5], [2] shows that the power consumption of a physical machine is the sum of the idle power (or base power) and a linear function of its CPU load. In order to verify that power consumption is a linear function of CPU load, built-in power meters in the physical machines is used to measure the power usage under different CPU loads. CPU load in this paper is defined as ”the sum of the loading of each core”. For example, if there are four cores with loading 100%, the CPU load of this physical machine will be 400%. The experimental result is shown in Figure 1, where the x-axis is the number of cores with 100% load, and y-axis is the power consumption reading from the power meter. The physical machine used has two CPUs, each with four cores. The result verifies the linearity of the power consumption as a function of the number of cores with 100% load. The shorter line in Figure 1 indicates the power consumption from deploying heavily loaded virtual machines on the

Figure 1.

Power consumption under different CPU loads

cores of the same CPU, and the longer line indicates the power consumption when deploying virtual machines to two different CPUs as much as possible. Figure 1 concludes that both lines are almost linear when the CPU load stays at 100%. The ratio between peak power and idle power can also be found from Figure 1. Peak power Pp is the power consumption when all eight cores of the physical machine have 100% load, which is 800% in Figure 1. Idle power Pi denotes the power consumption when none of the eight cores has any workload. From Figure 1 observation can be made that idle power is about 50% of the peak power. Based on the experimental results and the discussion above, the assumption of a linear power model with 50% idle power is reasonable. Therefore, the power consumption of a physical machine in our simulation is estimated using the linear power model as in Equation 1. P =[

c (1 − α) + α]Pp C

(1)

α denotes the percentage of idle power versus the peak power. In Figure 1 α is set to be 50%. The load of a physical machine is Cc , where c is the total number of cores required by the resident virtual machines, and C is the total number of cores of the physical machine. From the linearity of the power consumption function, Equation 1 can be concluded. The total energy consumption of a physical machine during a time period t is denoted as P t. B. Migration Model Live migration [3] is adopted as our migration model. Live migration allows a server administrator to move a running virtual machine to a different physical machine without interrupting the execution of virtual machine. As a result, the execution time of a virtual machine will not

be affected by migration. Live migration is supported by popular virtualization hypervisor, like Xen [14]. Although live migration does not lengthen the execution time of a virtual machine, it does increase energy consumption of source and destination physical machines where the migration takes place. When migration starts, the destination physical machine will create a process to copy the memory contents from source physical machine, increasing the loading, thus increasing the total power consumption. Experiments are conducted to measure the increased power consumption. Power meters are used to measure the extra energy consumption during migration. From our experiment results, live migration increases about 20% loading to the destination physical machine. The loading of source destination remains the same. Based on the results, the energy consumption during migration is defined as in Equation 2, where Em is the energy consumption of migrating a virtual machine from the source to the destination physical machine, c is the number of cores required by the migrating virtual machine, c0 is the sum of the number of cores required by the virtual machines running on the destination physical machine. Therefore, c c0 C and C are the percentage load of the source and the destination machine respectively, α is the idle power ratio, Pp is the peak power of a physical machine, tM is the time the migration takes, and L is the extra load caused by the migration process. Em

=

Es + Ed c = [( )(1 − α) + α]Pp tM C c0 + {[( ) + L](1 − α) + α}Pp tM C The notations are summarized in table I.

(2)

IV. VM DEPLOY STRATEGY Virtual machine consolidation is an NP-hard problem. The problem can be stated as follows. Virtual machines with different hardware requirements, such as the number of cores and the size of memory, must be deployed to physical machines with fixed hardware capacity in a way that minimizes the number of physical machine used. This consolidation problem is harder than the bin packing problem [13], which is known to be NP-hard [6]. It is easy to see that greedy method for bin packing problem, like First-Fit method, can also be modified and applied to our consolidation problem. The First-Fit method is an approximation algorithm to the bin packing problem, with approximation factor of 2 [15]. Modifications can be made to the First-Fit method to fit into the context of our consolidation problem. The FirstFit method attempts to deploy a virtual machine to the first machine in a physical machine list that can accommodate this virtual machine. If no physical machine is found, then

Table I N OTATIONS USED IN THE POWER AND MIGRATION MODELS Symbol Pp Pi c,c0 C c/C,c0 /C Em Es Ed Ec tr tM L α θ

Meaning peak power of a physical machine idle power of a physical machine total number of cores required by virtual machines total number of cores of the physical machine the load of a physical machine the energy consumption of migrating a virtual machine the energy consumption of the source physical machine the energy consumption of the destination physical machine the energy consumption after consolidation, from the migration ends to the end of execution of the virtual machine remaining execution time of a virtual machine time the migration takes extra load caused by the migration process the percentage of idle power versus the peak power retirement threshold

a new physical machine will be booted to host this virtual machine. Despite that the First-Fit method is a simple and effective heuristic to the bin packing problem, it may not be a good solution to our consolidation problem. A more complicated virtual machine consolidation problem is considered in this paper. In addition to resource requirements, each virtual machine also has an arrival time and an execution time. Unlike objects in the bin packing problem, every virtual machine only exists during its execution period, which starts at its arrival time, and lasts for its execution time. After a Virtual machine finishes it leave the system and the cores it used become idle before they are reassigned to other virtual machines. These idle cores may increase the actual number of physical machine required to run all virtual machines. We propose two virtual machine deployment methods, Dynamic Round-Robin(DRR) and a Hybrid approach for virtual machine consolidation. The objective of these methods is to minimize the number of physical machines used to run all virtual machines. This goal is very important because the number of physical machines used strongly affects total power consumption. A. Dynamic Round-Robin Dynamic Round-Robin method is proposed as an extension to the Round-Robin method. Dynamic Round-Robin method uses two rules to help consolidate virtual machines. The first rule is that if a virtual machine has finished and there are still other virtual machines hosted on the same physical machine, this physical machine will accept no more new virtual machine. Such physical machines are referred

to as being in “retiring” state, meaning that when the rest of the virtual machines finish their execution, this physical machine can be shutdown. The second rule of Dynamic Round-Robin method is that if a physical machine is in the “retiring” state for a sufficiently long period of time, instead of waiting for the residing virtual machines to finish, the physical machine will be forced to migrate the rest of the virtual machines to other physical machines, and shutdown after the migration finishes. This waiting time threshold is denoted as “retirement threshold”. A physical machine that is in the retiring state but cannot finish all virtual machines after the retirement threshold will be forced to migrate its virtual machines and shutdown. Our Dynamic Round-Robin strategy uses these two rules in order to consolidate virtual machines deployed by the original Round-Robin method. The first rule avoids adding extra virtual machines to a retiring physical machine. The second rule speeds up the consolidation process and enables Dynamic Round-Robin to shutdown physical machines, so that it can reduce the number of physical machine used to run all virtual machines, thus achieve power saving. B. A Hybrid Approach In order to conserve more energy, we combine Dynamic Round-Robin and First-Fit into a Hybrid algorithm. The number of incoming virtual machines is assumed to be a function of time, and follows a probability distribution (e.g., a normal distribution). Our hybrid algorithm will use the incoming rate of virtual machines to guide the scheduling of virtual machines. The main idea of the Hybrid algorithm is to determine which algorithm to use based on the incoming rate of virtual machines. The First-Fit method is suitable for high incoming rate of virtual machines. By knowing there will be a large number of incoming virtual machine, it is not necessary to be conservative in using resources (i.e., not to conserve the number of machines) since the demands will keep coming. On the other hand, the Dynamic Round-Robin performs well for low incoming rate of virtual machines, since resources must be actively reclaimed. Otherwise, physical machines will become idle due to lack of incoming virtual machines. Our Hybrid algorithm works as follows. The Hybrid method uses First-Fit during rush hours to fully utilize the computing power of physical machines, and uses the Dynamic Round-Robin to consolidate virtual machines and reduce energy consumption during non-rush hours. More details on the definition of rush hours will be given in Section VI-C2 V. I MPLEMENTATION This section introduce some design issues that arose during our algorithm implementation.

A. Retirement Threshold The first implementation issue is to determine a proper retirement threshold. As mentioned in section IV-A, a physical machine will be forced to shut down if in “retiring” for a sufficiently long period of time. This waiting time threshold is denoted as “retirement threshold”. If the threshold is too small, there will be a large number of migrations, which increases energy consumption. If the threshold is too large, then it will postpone consolidation and increase the number of physical machines required, which also increases energy consumption. The threshold can be determined by considering the migration time and the remaining execution time of virtual machines. The intuition is that if a virtual machine almost finishes its execution, then it is not necessary to migrate it. On the other hand, if the remaining execution time of a virtual machine is still long, consolidating it with other virtual machines is very likely to save energy. The two equations of the overall energy consumption, one with decision to migrate a virtual machine and the other not to migrate a virtual machine, respectively, is defined in the following. A condition can be found to make these two equations have equal values, which means the energy consumption is the same for either migrating or not migrating. A proper retirement threshold can then be computed from this condition. The energy consumption for migrating the virtual machine is defined as Equation 2. On the other hand, the energy consumption for not migrating the virtual machine can be defined as in Equation 3. Es and Ed are respectively the energy consumption of the source and destination physical machines if not migrating the virtual machine. E

= Es + Ed c = [ (1 − α) + α]Pp tr C c0 + [ (1 − α) + α]Pp tr (3) C The energy consumption of migrating and not migrating the virtual machine strikes a balance while Equation 2 equals to Equation 3. The two equations have the same value under the condition that the remaining execution time of a VM tr equals to θ, which is defined in Equation 4. L (4) α While a physical machine is going to retire, instead of deciding whether to migrate each virtual machine by their remaining execution time tr , which is not known, the threshold θ is used as the retirement threshold. If a virtual machine with remaining execution time t0r less equal than θ, meaning not migrating it is more energy saving, the virtual machine will eventually finish before the physical machine θ = (1 − L)tM +

starts to migrate virtual machines. On the other hand, if t0r is greater than θ, meaning migrating is a better choice, the virtual machine will be migrated after the retirement threshold exceed. In Equation 4, L, α, and tM are all constants. Therefore, the threshold θ can be calculated in advance. Notice that the remaining execution time tr is not needed while running our Dynamic Round-Robin. The threshold θ is calculated and fixed. A retiring physical machine will be shut down after a waiting time of θ. Every unfinished virtual machine will be migrated no matter how long the remaining execution time is. B. Retirement Criteria The other implementation issue is retirement criteria. By migrating every virtual machine on physical machines which reach the retirement threshold, it may not minimize overall power consumption. For example, if a source physical machine has a high load since it hosts many virtual machine, and will be forced to retire and shut down because one of the virtual machine has finished. All the rest of the virtual machines need to be migrated to other physical machines, which increases the power consumption and reduces the resource availability of other physical machines. A method is proposed to solve this problem in the following way. Instead of using time as the only threshold, the load of the physical machine is also considered as a parameter of retirement criteria. After a virtual machine finishes, the hosting physical machine will not accepts any new virtual machines, and will check its loading. If the loading is still over a certain threshold, then this physical machine will not retire until the load decreases under the threshold. The intuition is that by reducing the number of virtual machine migration, more energy can be saved. This method also reduces the network traffics caused by migration. VI. E XPERIMENTAL R ESULTS A. Settings and test cases For small-scale experiments, a homogeneous 8-node cluster is used where each physical machine has two-way quad-core CPUs and 24GB memory. A linear power-load model with idle power Pi set to be 50% of the peak power is adopted, as mentioned in Section III-A. A live migration model is also established. The migration cost L is 0.025(1/8 × 0.2), which means each migration process increases 20% load to one of the eight cores in the machine. A time unit is assumed to be one minute, and migrating a virtual machine takes two time units (tM = 2). Test cases are generated as follows. Each test case contains a set of virtual machines, with each virtual machine being one of the three pre-defined types from EC2 (small, large, extra large), arrival time, and execution time. A virtual machine has a core and a memory requirement according to one of the three EC2 standard instances. The execution time

of a virtual machine ranges from 2 to 12 hours randomly. The arrival time of virtual machines is generated from a normal distribution to reflect the fact that the amount of incoming virtual machines is related to the daily routine of users, who tends to submit more jobs during work hours. B. Small Scale Experimental Results The purpose of this experiment is to verify our power model and migration model. A physical machine has two Intel(R) Xeon(R) E5620 CPUs running at 2.40GHz, and each CPU has four cores and 24GB memory. Xen Hypervisor is installed on each physical machine for virtualization. Two sets of power data are collected. One is estimated by our power model, and the other is the actual power usage measured by power meters in the machine. A virtual machine provision and consolidation system is implemented. The experiment is conducted in the following way. The system creates virtual machines according to their arrival time, and a virtual machine scheduling algorithm determines which physical machine should host the incoming virtual machine. After a virtual machine is created on the physical machine, it starts running and keeps the core busy with 100% load. Power consumption is recorded every minute, and the average power consumption is calculated after the experiments. Table II P OWER CONSUMPTION RESULTS USING POWER - SAVE F IRST-F IT AND DYNAMIC ROUND -ROBIN . Power-Saving First-Fit Dynamic RR Improvement in power saving

Estimated Power 662.874 W

Actual Power 638.664 W

644.242 W 2.8%

618.865 W 3.1%

Table II compares the power consumption from Dynamic Round-Robin (DRR) we propose and the Power-Saving First-Fit method. The Power-saving First-Fit is the same as the First-Fit heuristic except that when a machine becomes idle, it is shut down to conserve energy. Energy consumption is estimated according to our power model before the experiment, and then compare the results with actual power meter reading after the experiment. As can be seen from Table II, there is very small difference between the estimated power and the measured power. The reason for this discrepancy could be caused by the assumption that a virtual machine always has 100% load, while the actual load of a virtual machine is more complicated. It takes some time for a virtual machine to initialize, and the workload could vary during the execution of a virtual machine. Another possible reason for this discrepancy could be that the power overhead is slightly overestimated in our power model, therefore the actual power is smaller than the estimated power. Nevertheless, the estimation error is within

5%, and the power saving percentage is consistent, either in estimated or measured power. As a result, the estimation from our power model is believed to be accurate enough to evaluate different scheduling algorithms. C. Large Scale Simulation Results Simulations have been made to see how much energy can be saved by our algorithms for a large number of incoming virtual machines. Our simulator simulates a data center with 600 8-core Xeon servers connected by Gigabit Ethernet switches. In each simulation run (i.e. test case), there are 2000 incoming virtual machines. The simulator records the numbers of physical machine used, the number of migrations, and the total energy usage at every time unit of the simulation. 1) Different Retirement Thresholds: The first simulation applies different retirement thresholds to observe how these thresholds affect performance. Recall that retirement threshold is the length of time from a virtual machine finishes its job to the time that the virtual machine on the physical machine is migrated. The retirement threshold is set to be 0, 5, 10 and 30 minutes in the simulation for comparison purpose. CPU load is also used to determine whether to retire a physical machine, as described in Section V. That is, despite the time in retiring state, a physical machine migrates the hosting virtual machines if the loading of physical machine decreases under a certain threshold, 50% in our case, during retiring.

The first four are the results of Dynamic Round-Robin using different retirement thresholds, the fifth bar(marked as “’D’) uses CPU load as the retirement criteria, and the last two bars are the results of power-saving Round-Robin and FirstFit. From Figure 2 a conclusion can be made that when the retirement threshold decreases, power saving increases. One thing worth to mention is that using physical machine load as the retirement threshold outperforms using time as the threshold. As can be observed from the Figure 2, the power consumption of using load as threshold(bars marked by “D”) always uses less energy than those using time as the retirement threshold. 2) Performance comparison: Performance results from Dynamic Round-Robin and the Hybrid method are compared with other algorithms, including First-Fit, Round-Robin, and Best-Fit, with First-Fit as the baseline for normalization. Note that all these algorithms are implemented in a powersaving manner. That is, they all shut down idle physical machines to conserve energy. Average power consumption is used as the major performance metric, and other metrics like average number of physical machines used, and the number of migrations are also considered. The Hybrid method requires a definition of rush hour. Since the arriving time of virtual machines are assumed to follow a normal distribution, rush hours is defined to be “the time interval centered at the mean of the arriving time, and plus/minus a standard derivation”.

Figure 2. Average power consumption under different retirement threshold using DRR, comparing with RR, and FF. Each test case contains 2000 virtual machines.

Figure 3. Comparison of average power consumption between different algorithms. Each test case has 2000 virtual machines.

Figure 2 illustrates the average power consumption of three power-saving algorithms (that is, they all shut down idle machines). The average power consumption (y-axis) is normalized by using the power consumption of the powersaving First-Fit algorithm as the baseline. Every set of bars represent the results of a test case from different algorithms.

Figure3 illustrates the normalized power consumption of five algorithms from 10 test cases. The Dynamic RoundRobin and the Hybrid method use physical machine load as the retirement threshold, since it conserves most energy from the discussion in Section VI-C1. In addition, the Hybrid method outperforms all other algorithms, and save about 3%

more power than the power-saving First-Fit method does in every case.

Figure 5. The normalized average percentage of powered-on physical machines with ROUNDROBIN as the baseline

Figure 4. Comparison of average number of powered-on physical machines per minute between different algorithms. Each test case has 2000 virtual machines.

Figure 4 shows the average number of powered-on physical machines per minute. The number of powered-on physical machines is strongly related to the power consumption because even if a machine is idle, a powered-on machine still consumes 50% power. As can be observed from Figure 4, the Hybrid method reduces the average number of poweredon physical machines by 5.5%, and Dynamic Round-Robin reduces the number by about 4.5%. D. Comparison with Eucalyptus Eucalyptus [4], [11] is a widely used open-source Cloud middleware. There are three scheduling options in Eucalyptus: GREEDY, ROUNDROBIN and POWERSAVE. GREEDY and ROUNDROBIN do not shut down machines nor put machines to sleep. POWERSAVE is equivalent to the powersaving First-Fit method described in the previous section. We measure the power consumptions of the three scheduling algorithms in Eucalyptus, and compare the results with our proposed algorithms. Since GREEDY and ROUNDROBIN do not shut down any physical machine, the number of powered-on physical machines in GREEDY and ROUNDROBIN is equivalent to the number of physical machines required to process all incoming virtual machines. Figure 5 and Figure 6 show average number of powered-on physical machines and average power consumption from the five different scheduling methods, where ROUNDROBIN, the default method of Eucalyptus, is the baseline. The five algorithms include two of our proposed algorithms (Dynamic Round-Robin and Hybrid method), GREEDY, ROUNDROBIN, and POWERSAVE. From Figure 5 and Figure 6, observation can be made that the number of powered-on physical machines and the power

Figure 6. The normalized average percentage of power consumption with ROUNDROBIN as the baseline

consumption reduce significantly using our algorithms, when compared with those from Eucalyptus. The Dynamic RoundRobin and the Hybrid method reduce 56.4% and 55.9% power consumption respectively when compared with the power consumption required by ROUNDROBIN policy. Furthermore, our algorithms save 3% more power on average than POWERSAVE scheduling policy in Eucalyptus. E. Summary We demonstrate the effectiveness of our algorithms in both small scale experiments and large scale simulations. Experiments have been run on an 8-node homogeneous environment, and the experimental results show that proposed Dynamic Round-Robin consume less energy than the First-Fit method. The experimental results also validate our system model. Large scale simulations have also been

performed. The simulation results show that using physical machine load as the retirement threshold save more energy than using time as the retirement threshold. The average power consumption from different algorithms are compared and analyzed. Dynamic Round-Robin reduces the average number of power-on physical machine by 4.5%, and reduce average power consumption by 2%. The Hybrid method has the best performance among all algorithms – it saves the number of power-on physical machines by 5%, and average power consumption by 3%. Comparison between our power consumption results with the three scheduling options in Eucalyptus are also presented, and the result shows that the saving is about 45% in power when compared with their default setting. VII. C ONCLUSION Power saving is an important issue for data centers. Large data center can consume million Watts per year. Cloud data centers such as Amazon EC2 provide users resizable compute capacity in the cloud. Instead of directly manipulate physical machines, virtual machine instances are created and controlled by users. How to deploy these virtual machines to physical machines to achieve power saving in homogeneous environment become an important issue. We propose two methods, Dynamic Round-Robin and Hybrid, to deploy the virtual machines for power saving purpose. Dynamic Round-Robin is an extension from original Round-Robin, with two rules that help the virtual machines to consolidate. Hybrid is the combination of Dynamic Round-Robin and First-Fit. Simulations and experiments have been conducted to verify our algorithms. The average power consumption is used as performance metrics and the result of power-saving First-Fit is used as baseline. Results show that Dynamic Round-Robin and Hybrid method we proposed can reduce the number of power-on physical machine and average power consumption compare to other deploy algorithms with power saving. Also, the comparison between our algorithms and the three scheduling algorithms in Eucalyptus shows significant improvement in saving power. R EFERENCES [1] Jeffrey S. Chase, Darrell C. Anderson, Prachi N. Thakar, Amin M. Vahdat, and Ronald P. Doyle. Managing energy and server resources in hosting centers. In Proceedings of the eighteenth ACM symposium on Operating systems principles, SOSP ’01, pages 103–116, New York, NY, USA, 2001. ACM. [2] Gong Chen, Wenbo He, Jie Liu, Suman Nath, Leonidas Rigas, Lin Xiao, and Feng Zhao. Energy-aware server provisioning and load dispatching for connection-intensive internet services. In Proceedings of the 5th USENIX Symposium on Networked Systems Design and Implementation, NSDI’08, pages 337–350, Berkeley, CA, USA, 2008. USENIX Association.

[3] Christopher Clark, Keir Fraser, Steven Hand, Jacob Gorm Hansen, Eric Jul, Christian Limpach, Ian Pratt, and Andrew Warfield. Live migration of virtual machines. In Proceedings of the 2nd conference on Symposium on Networked Systems Design & Implementation - Volume 2, NSDI’05, pages 273– 286, Berkeley, CA, USA, 2005. USENIX Association. [4] Eucalyptus. http://open.eucalyptus.com/. [5] Xiaobo Fan, Wolf-Dietrich Weber, and Luiz Andre Barroso. Power provisioning for a warehouse-sized computer. In Proceedings of the 34th annual international symposium on Computer architecture, ISCA ’07, pages 13–23, New York, NY, USA, 2007. ACM. [6] Michael R. Garey and David S. Johnson. Computers and Intractability; A Guide to the Theory of NP-Completeness. W. H. Freeman & Co., New York, NY, USA, 1990. [7] Liting Hu, Hai Jin, Xiaofei Liao, Xianjie Xiong, and Haikun Liu. Magnet: A novel scheduling policy for power reduction in cluster with virtual machines. In CLUSTER, pages 13–22. IEEE, 2008. [8] Liang Liu, Hao Wang, Xue Liu, Xing Jin, Wen Bo He, Qing Bo Wang, and Ying Chen. Greencloud: a new architecture for green data center. In Proceedings of the 6th international conference industry session on Autonomic computing and communications industry session, ICAC-INDST ’09, pages 29–38, New York, NY, USA, 2009. ACM. [9] Michele Mazzucco, Dmytro Dyachuk, and Ralph Deters. Maximizing cloud providers’ revenues via energy aware allocation policies. Cloud Computing, IEEE International Conference on, 0:131–138, 2010. [10] Michele Mazzucco, Dmytro Dyachuk, and Marios Dikaiakos. Profit-aware server allocation for green internet services. Modeling, Analysis, and Simulation of Computer Systems, International Symposium on, 0:277–284, 2010. [11] Daniel Nurmi, Rich Wolski, Chris Grzegorczyk, Graziano Obertelli, Sunil Soman, Lamia Youseff, and Dmitrii Zagorodnov. The eucalyptus open-source cloud-computing system. In Proceedings of the 2009 9th IEEE/ACM International Symposium on Cluster Computing and the Grid, CCGRID ’09, pages 124–131, Washington, DC, USA, 2009. IEEE Computer Society. [12] Shingo Takeda and Toshinori Takemura. A rank-based vm consolidation method for power saving in datacenters. Information and Media Technologies, 5(3):994–1002, 2010. [13] Vijay V. Vazirani. Approximation Algorithms, chapter Bin Packing, pages 74–78. Springer Verlag, 2001. [14] Xen. http://www.xen.org. [15] Binzhou Xia and Zhiyi Tan. Tighter bounds of the first fit algorithm for the bin-packing problem. Discrete Appl. Math., 158:1668–1675, August 2010.