Profile-based Static Virtual Machine Placement for Energy-Efficient Data center

Transcription

1 2016 IEEE 18th International Conference on High Performance Computing and Communications; IEEE 14th International Conference on Smart City; IEEE 2nd International Conference on Data Science and Systems Profile-based Static Virtual Machine Placement for Energy-Efficient Data center Fares Alharbi, Yu Chu Tain, Maolin Tang and Tusher Kumer Sarker School of Electrical Engineering and Computer Science Queensland University of Technology Brisbane, Australia {g.tain, Abstract The energy consumption of a data center and hence the carbon footprint from it largely depends on the energy consumption by its active Physical Machines (PMs). Researchers have taken many attempts to minimize the data center energy consumption through the Virtual Machines (VMs) allocation into a minimal number of PMs of homogeneous types. However, the current VM placement strategies do not consider useful information that can be extracted from data logs of data center. This paper presents profile-based VM placement approach to improve energy efficiency of data centers. The approach formulates the energy consumption problem as a profile-based optimization problem. Then, the problem decomposed into multiple smaller ones in a number of intervals. For each intervals, a number VMs and PMs are sorted in terms of resource requirements and energy efficiency respectively. Then, the First Fit-Decreasing(FFD) is adopted to place the sorted VMs to the sorted PMs. Experiments conducted to demonstrate the presented approach with comparisons with the original FFD algorithm. The experimental results have shown that the presented approach can reduce more energy consumption than the original FFD algorithm and is scalable for larger test problems. Keywords data center, virtual machine, physical machine, energy consumption, placement I. INTRODUCTION Now-a-days Cloud Computing is a hot topic for researchers and academic because it provides a number of services to a wide range of users from a single end user to an enterprise owner. It can provide for users on-demand services such as configurable computing resources. Cloud Computing can serve its clients with higher capability and reasonable time in different locations. Virtualization [1] is the technology that has enabled the Cloud data centers to provide on-demand services in the form of Virtual Machines (VMs). A VM is an abstract machine created on the top of a Physical Machine (PM) by sharing its physical resources. A Cloud data center comprises a huge numbers of PMs which are to provide ondemand physical resources to the users located in different geographical locations in the globe through their hosted VMs. However, those PMs consume a huge amount of energy which increase electricity cost and carbon footprint as well. The consumption of energy of a PM depends on its energy profiles and CPU utilization by its hosted VMs. A study of Le et al. [2] shows that a data center energy consumption cost is more than $15M in a year at its average CPU utilization. While Cloud data center is providing enormous services to its users, an huge amount of world electricity is being consumed by it. Jonathan G. Koomey [3] explained that there was a worldwide increment in the electricity used by data centers was around 56% from 2005 to In 2006, data centers in the US used over 1.5% of the overall energy produced that year, and it is expected that the percentage would increase 18% on a yearly basis [4]. Therefore, a lot of pressure is faced by infrastructure providers to decrease the data center energy consumption, and intensive efforts have been given to optimize data center resource usage and hence to decrease the data center energy consumption [2], [5] [7]. The energy consumption of a data center is characterized by its effective resource management. Thus, resource management on cloud environments has attracted a great deal of focus over the course of the past decade [8]. By resource management in this paper we mean the VM placement. VM placement is the key management activity in today s Cloud data center. Inefficient VM placement is one of the crucial problems in cloud data center. For example, placing two VMs on arbitrary two different PMs with same energy efficiency will consume more energy in the data center than the case when these two VMs are placed on one of those two PMs by switching of the second one [6]. Moreover, the resource requirements of a VM is dynamic. Therefore, the VM placement decision is a crucial factor, which should account the VM resource usages profiles. According to Atefeh et al. [9], there are distinct characteristics of each VM, and on the basis of the resources used by the VM, the hosting PM consumes an amount of energy in the data center. When the cloud manager receives a VM request, it chooses the physical resources required to start the request. Placement of VM in cloud data center is a complicated task and if it cannot be carried out effectively, it generates high energy consumption and high carbon emission [9]. There are two kinds of VM placement static VM placement and dynamic VM placement. The static or initial placement means assigning VMs to the PMs during the creation of the VMs. The assignment of VMs to PM means the multiplexing of PMs physical resources to the VMs. In this work, we address the static VM placement problem and consider with two types of physical resources CPU and memory. However, the static placement of VM should consider the resource usage /16 $ IEEE DOI /HPCC-SmartCity-DSS

2 profiles of VMs because the workload on VMs changes over time due to different applications hosted by those VMs, which can result in changing resource utilization on PMs. Therefore, the optimal VM placement should be found on the basis of profiling of VMs and PMs for a number of time intervals. To address this challenging problem, in this paper we proposed profile based VM placement for minimizing energy in data center. We have presented an approach, which considers the subsequent time intervals of data center, and resource usage profiles of VMs and PMs. We have evaluated our approach by conducting experiments, and the experimental results show that our approach outperforms a well-known heuristic algorithm, bin backing(ffd) [10], by reducing energy consumption for 24 hours. The contribution of this paper can explicitly identified as follows: The concept of profile used for VM placement to allow us prediction of workload and future pre-scheduling in advance. The VM placement problem is formulated as a constrained optimization problem with a huge solution space. Then it is simplified into number of smaller similar problems which becomes solvable within a reasonable period of time. An approach is presented to define a solution to the simplified optimization problem. The rest of this paper is organized as follows: Section II gives the motivation and describes the related works. Section III presents the formulation of VM placement problem. Section IV describes the approach has been used. Section V evaluates our method and discusses the experiments and its results. Finally, Section VI concludes our work and presents future works. II. MOTIVATION AND RELATED WORK This section gives the motivation of this research, and presents the related works. A. A Motivation Example The static (or initial) VM placement problem is an wellattempted research problem in data center, and many algorithms have been proposed and developed to address this problem. Data center s energy consumption cost highly depends on the energy consumption by its active PMs. Therefore, researchers have taken many attempts to minimize the energy consumption of PMs in the data center. However, the energy consumption of a PM depends on its energy attribute, CPU capacity and the VMs deployed on it. In this section, we give a motivation example for the necessity of energy-efficient VM placement that illustrates how the energy consumption of a data center depends on PMs energy attributes, CPU capacities and the VMs deployed on them. A data center comprises heterogeneous PMs, i.e. PMs with different attributes (e.g. CPU capacity, energy efficiency) and hosts the heterogeneous types VMs as well, for example, VMs with different resources requirements. To make this example simple, we consider all the VMs are created at the same time and executed for the same duration, and do not change the resource requirements during their execution. Consider the attributes of the PMs and VMs are given in Tables I and II respectively. As the energy consumption of a PM depends on the CPU usage of the PM, we have only considered the CPU resource in our example. The energy efficiency of a PM is calculated as the energy consumption of each CPU core, i.e. Max. Energy Energy Efficiency = Consumption PM CPU Capacity TABLE I PMS ATTRIBUTES Max. Energy Consumption (kwhr) CP U Capacity. Min. Energy Consumption (kwhr) Energy Efficiency pm pm pm pm pm TABLE II VMS CPU REQUIREMENTS VM CPU Requirements vm 0 8 vm 1 3 vm 2 3 vm 3 1 vm 4 1 vm 5 1 vm 6 1 vm 7 8 vm 8 7 vm 9 1 Now consider two placements Placement-1: {{vm 8 } pm 0, {vm 1,vm 9 } pm 1, {vm 7 } pm 2, {vm 2,vm 3 } pm 3, {vm 0,vm 4 } pm 4 }; Placement-2: {{vm 3 } pm 0, {vm 0,vm 4 } pm 1, {vm 7 } pm 2, {vm 2,vm 8 } pm 3, {vm 1,vm 9 } pm 4 }. Here the notation {vm i } pm j means the set of VMs, {vm i }, are assigend to pm j. We calculate the energy consumption of a PM using the model proposed in [6]. Placement-1 results in total energy consumption by PMs to kwhr and Placement-2 gives total kwhr energy consumption by the PMs. This simple example illustrates that different VM placement result in different energy consumption. However, there can be many other VMs placement scenarios giving different energy consumption. In this simple example, we have only considered CPU resources and constant CPU requirement by each VM. However, in real scenario the CPU requirements of a VM can vary over time, which has been taken into consideration in this research, and the more resource constraints need to be satisfied for a VM placement. Therefore, the VM placement should be done in strategic way. Motivating from this example, we will formulate a profile based VM placement problem. B. Related Work Most of the VM placement algorithms emphasize on minimizing the number of active PMs so that total energy con- 1046

3 sumption by PMs is minimized [11] [13], thus the approaches attempt to minimize the data center energy consumption. These algorithms, however, are applicable for homogeneous types of PMs. On the contrary, in real data centers the PMs are of heterogeneous types. Yang et al. [14] proposed an available-aware and energyefficient VM placement algorithm to avoid the application failure. The approach considered the redundancy of VMs hosting the same application and placed the VMs into different PMs so that if a PM failure occurs, the execution of the application is completed by other VMs on different PMs. However, this approach increases the energy consumption because more number of PMs need to be activated. Moreover, the approach proposed by Yang et al. did not consider the diversity of energy profiles of PMs and the dynamic workload change of VMs. The static VM placement has been studied as a constraint satisfaction problem [5], [15] [18]. In these approaches, the placement of VMs was done satisfying the resource constraints of PMs while aiming to minimize the data center energy consumption. Beloglazov et al. [5], however, only considered the CPU resources in their VM placement problem, which make their approach less realistic for multiple resource constraints satisfaction problem. Though Zhao et al. [15] considered multiple resource constraints of a PM for placing a VM on it, their approach was tested for small number of VMs, experiment was conducted for 32 VMs only, therefore, the scalability of the approach was not tested. All these approaches, however, did not consider of having different energy profiles of PMs, the workloads fluctuations of VMs, which are present in real data center scenario. In addition to this, the scalability experiments were not present in these works, i.e. the data center sizes were not taken into account in these work. The dynamic workload fluctuation was considered by Zhang et al. [19]. However, the approach did not consider the profile based VM placement. Zhang et al. proposed an algorithm that dynamically places a VM when its workload changes, therefore, the approach best addressed the dynamic VM placement. On the other hand, this paper focuses on static VM placement accounting the VM profiles, which avoid the cost of dynamic VM placement. A very close work to our research was proposed in [20] by Vasudevan et al. They utilized profiling techniques against application and VMs in enabling allocation managers to determine the requirements and availability of resources so that allocation can be made easily. The profiles helped the assigning algorithms towards concluding the most suitable VM for hosting the application. This would enable the initial step in constructing profiles which would relate to specific data related to functions associated with energy use, CPU functions, memory usage, application completion times and the frequency. The profiles mainly relate to data associated with CPU utilization and the time in completing the applications. On creating the profiles, the system updates CPU utilization data and completion times of application regularly towards ensuring the efficiency of the system. However, the approach proposed by Vasudevan et al. considered the applications assignment to the VM. On the contrary, in the context of our research, we consider the VMs assignment onto the PMs giving more efficient approach for minimizing data center energy consumption. The above state-of-the-art works aimed to address the static VM placement problem minimizing the data center energy consumption. However, there remains a gap between these stat-of-the-art works and the real data center scenario the profiles of the VMs, i.e. workload changes of VMs over time, was not considered, the heterogeneity of energy profiles of PMs was ignored and the VM placement did not consider for longer period of intervals. Motivating from these research gaps, in this paper we propose and develop a profile based VM placement in the data centers with heterogeneous PMs, in terms of resource capacities and energy attributes, for a longer period (24 hours). The VMs profiles related to the requirements of CPU and memory, arrival time and end time of each VM; and the CPU and memory usages of a VM would be considered as changeable during time intervals based on load requirements. III. PROBLEM FORMULATION In this section, we use profiling to find optimal place for VM on PM to optimize energy. Also in our work, we consider next T time intervals to optimize VM allocation that can minimize energy consumption for the whole time intervals. A. Profiling We will use profiling of VMs and PMs that helps allocation manager to identify availability to satisfy requirements of requested by VMs. Profiling of VM has been done as any newly created VM to data center will go under profiling with its arrival time, life cycle, dynamic requirements of CPU and memory of its life cycle. PM profile contains its energy efficiency, CPU and memory capacities, CPU and memory usages. In this work, the profile has been created randomly. B. Mathematical Modelling Our model will read VMs and PMs attributes from the profile has been discussed in section III-A. Given a data center consisting of a set of PMs with their resource capacities and energy profiles; a set of VMs, for which a placement to be found, with their resource requirements profiles; find a static VM placement for each of T time intervals such that the total energy consumption is minimized while satisfy the placement constraints. The mathematical formulation of the problem can be given as follows: Input: The following are the inputs to the problem: (i) A set of PMs, P = pm i, where pm i is a PM in the data center with its CPU capacity, CPU pmi, memory capacity, RAM pmi and energy efficiency, η pmi. (ii) A set of VMs, V = vm i, where vm i is a VM to be placed in the data center with its CPU requirements profiles, CPU vmi, memory requirements profiles, RAM vmi. 1047

4 Output: A static placement of V for each T intervals, P k = { vm i,pm j }, where each 2-tuple, vm i,pm j, illustrates the placement of vm i to pm j and represented by binary value X ij =1. Constraints: For each VM placement, vm i,pm j, the following two resource constraints must be satisfied: (i) CPU pmj CPU vmi, X ij =1. (ii) RAM pmj RAM vmi, X ij =1. Constraint (i) indicates that total CPU capacity of pm j must be greater or equal to the total CPU requirements by the VMs deployed on pm j. Constraint (ii) denotes that total memory capacity of pm j must be greater or equal to the total memory requirements by the VMs deployed on pm j. Objective: The obtained placement, P k, minimizes the total energy consumption in the data center. This formulation based on integer programming [21] which has been used to allocate teachers to courses in optimal place on timetable. The assignment of vm i onto pm j is given a binary decision variable X ij,i as: { 1 if vm i is allocated to pm j X ij = (1) 0 otherwise C. Constraint Optimization Energy cost: The cost of energy for allocating vm i to pm j is denoted by C ij. Our model calculates the energy cost by using model [6]. This model calculates the energy based on CPU utilization. Our model reads CPU requirement of VMs and CPU capacity of PMs from profile to calculate CPU utilization for vm i by using (2). CPU utilization: The CPU utilization of pm j,μ ij, is defined in the following: μ ij = CPU vm i, X ij =1 (2) CPU pmj The energy cost of pm j for vm i, C ij, with X ij = 1 is calculated as follows using the model in [6] C ij = ( ) e pmj max e pmj min μij + e pmj min μ j (3) where e pmj max and e pmj min are the energy consumption of pm j when all of its CPU is completely utilized and there is no VM on pm j respectively; and μ j is the CPU utilization of pm j before allocating vm i to pm j. D. Simplifying Problem Formulation The presented approach in this paper minimizes the energy consumption of a data center for a period of time which is divided into a number of intervals as t =1, 2, 3,...,T.We use costs, C ij, as matrix [ ] C ij to allocate VM to PM. For each interval, t, the energy consumption of data center is minimized by placing the newly arrived VMs and running VMs (running VMs, however, do not change their PMs as static VM placement has been considered in this research) as ( P V ( ) ( ) ) min C ij t Xij t. The total energy consumption of j=1 i=1 the data center is thus minimized by adding the minimized energy consumption of each interval and is given by (4). mine ( t ) = Subject to: T ( ( P V ( ) ( ) )) min C ij t Xij t (4) t=1 j=1 i=1 CPU pmj CPU vmi, X ij =1; RAM pmj RAM vmi, X ij =1; X ij =1; vm i ; Equation 5 ensure the resources requirements has been satisfied for each intervals. It has been explained in section III-B. IV. PROPOSED APPROACH The presented approach uses the mathematical model to find near optimal solution for minimizing energy consumption in data centers for any period of time divided into a number of intervals. The approach uses the profiles of the VMs for which the placement have to be found and profiles of the PMs. The profile supplies to our model dynamic resource requirements of VMs with their arrival times and life cycles; and resource capacities and energy efficiencies of PMs. Our approach, described by the pseudocode in Algorithm 1. The algorithm iterates between steps 2 and 51 to calculate the energy consumption for each interval and the total energy consumption is found by adding the energy consumption of all intervals. The energy consumption of each interval, interval, is calculated from Step 3 to Step 49. Each interval is divided into a number of fixed time slots. A time slot is a duration at which the resource requirements of a VM remain unchanged. As the CPU requirements for a slot may be different from the requirements of other slots, the energy consumption for each slot is calculated and then energy consumption of all slots are added to get the total energy consumption of a PM for an interval. The proposed approach uses heuristic algorithm to find the placement at each interval. In the heuristic approach, the VMs are sorted in descending order of their CPU requirements and PMs are sorted according to their energy inefficiencies. The heuristic approach places the most CPU requirement VM to the most energy efficient PM first and the same process iterates for the remaining VMs in that interval. The presented approach is divided into two main parts between steps 6 and 20, which updates the data center resource status at the beginning of the interval; and between steps 21 and 46, an appropriate PM is found for each of the VMs of an interval. The lifetime of a VM can span more than one interval. Therefore, it is required to update the resource usages status of each PM at the beginning of interval before finding the placement for new VMs. Suppose the interval duration is 2 minutes and slot duration is 30 seconds and then (5) 1048

5 total number of slots in an interval is =4. Let the life cycle of a VM, vm i, is 5 minutes with CPU requirements {2, 3, 2, 2, 4, 1, 3, 2, 1, 3}. Then if vm i is created at interval 1 on PM, pm j, it will finish its execution at interval 3 as 5 2 = 2.5. The CPU usages of pm j in interval 1 will be {2, 3, 2, 2} for its 4 slots. Let we need to find the placement of VMs at interval 2. Before doing this we need to update resource usages status of pm j and in that case the CPU usages of pm j will be {4, 1, 3, 2} as first 4 slots of vm i has already finished. This process is implemented as the steps between 8 and 11. However, some VMs can finish their execution within an interval, which are needed to be removed from the data center by relinquishing their consumed resources, Step 13and 14. If there is no VM in a PM after an interval, then the PM should be switched off to minimize the energy consumption according to Step 18. The second core part of our approach finds the suitable PM for a VM according to steps from 21 to 46. To do this, two PMs are found for each VM one of which is active and another one is inactive. This is because, it is always not the case that placing a VM onto an active PM instead of a switched off PM will give minimum energy consumption. For example, placing a VM with 2 core CPU requirements on an inactive PM with 4 core CPU capacity, 100 kwh maximum energy consumption and 20 kwh minimum energy consumption will result in 50 kwh increased of energy consumption in data center as calculated in Step 32. On the other hand, if the VM is placed into an active PM with CPU capacity 4 core and maximum and minimum energy consumption of 120 kwh and 10 kwh respectively, then the increase of data center energy consumption will be 55 kwh, calculated by Step 31. Thus for each VM, the probable energy consumption of placing the VM on an active PM and an inactive PM are calculated. The PM that gives minimum energy consumption for that VM is chosen for the final placement. Once the placement of each VM is found for an interval, the energy consumption of that interval is calculated in Step 48 using model proposed in [6]. V. EVALUATION This section describes the baseline algorithm that has been used for comparison, presents the design of experiments and discusses the experimental results. A. Baseline Algorithm In real data centers, the PMs are of heterogeneous types. The very popular algorithm for VMs placement that has been considered heterogeneity of PMs CPU capacities is First-Fit- Decreasing (FFD) [10]. In the FFD, the VMs and PMs are ordered in decreasing of their CPU requirements and CPU capacities. After that the VM with most CPU requirement (first VM to be placed) is attempted to place on highest CPU capacity PM and this PM is added to the target PMs list if the placement of that VM is possible. The target PMs list is also sorted in descending of their CPU capacities. The subsequent VMs are placed in similar way, however, the target PMs for these VMs are searched in the target PMs list. If no PM is Algorithm 1 Heuristic algorithm for profile based VM placement 1: Energy total := 0.0 2: for interveal =1to total number of intervals do 3: Energy interval := 0.0 4: get the sorted the VMs list in the interval in descending order of their average CPU requirements, V i 5: update the resource usages of each PM in the data center for interval following steps 6 to 19 6: for each PM pm k in the data center do 7: for each VM, vm i,inapm k do 8: get the arrival Interval ID of vm i, vm arr i 9: get the execution duration of vm i, vm exe vm 10: if i Interval Duration > (interval vmarr i ) then 11: update the resource ( usages of pm k for the resource requirements of vm vm i from point ) i Interval Duration (interval vm arr i ) to the end of its life cycle 12: else 13: release the consumed resource by vm i to pm k 14: remove vm i from the data center 15: end if 16: end for 17: if there is no VM on pm k then 18: make the status of pm k in inactive mode 19: end if 20: end for 21: for each vm j V i do 22: find the set two most energy efficient PMs, P l, whenever possible, one of inactive PM and the another one of active PM and each pm k P l in must satisfy each type of resource requirements for vm j //Resource constraints are satisfied by checkresourceconstraints() method 23: if P l then 24: energy := : for each pm k P l do 26: EnergyCons := : for each slot of CPU requirements of vm j do slot cpu 28: calculate CPU utilization of pm k, U slot vm = vm j j pm cpu k 29: if pm k is already active then ( ) E max pm E min k pm k 30: EnergyCons+ := 31: else 32: EnergyCons+ := i U slot vm j ( ) E max pm E min k pm U slot k vm + j E min pm k 33: set the status of pm k to active 34: end if 35: end for 36: if energy == 1.0 OR energy > EnergyCons then 37: energy := EnergyCons 38: choose pm l for vm j 39: end if 40: end for 41: update the resource usages of pm l by the resources requirements of vm j 42: allocate vm k to pm l 43: else 44: Print no placement is found for vm j 45: end if 46: end for 47: for each PM, pm i, in the data center do 48: ( calculate the energy) consumption for interval as Energy interval+ := E max pm E min i pm U cpu i pm + E min i pm U cpu i pm i 49: end for 50: Energy total+ :=Energy interval 51: end for found in target PMs list, then an inactive PM is selected where the VM can be placed and that PM is included in the target PMs list. The process iteratively finds the VMs placement and minimizes the energy consumption in the data center by reducing the number of active PMs while taking into consideration into the heterogeneity of CPU capacities. However, the FFD does not consider the different energy profiles (or energy efficiencies) of the PMs. The energy consumption of a PM depends on its energy efficiency, for example, a PM with 1049

6 larger capacity may be less energy efficient and vice versa. Therefore, placing a high CPU demand VM to the more CPU capacity and less energy efficient PM results in more energy consumption. Our approach considers the energy profiles of the PMs as the primary attribute for VMs placement unlike the FFD. B. Experiment Design As the proposed VM placement approach is designed to deal with the placement of hundreds of VMs, it is essential to evaluate it on a large-scale infrastructure. However, it is difficult to conduct large-scale experiments on a real infrastructure, especially when it is necessary to reproduce the experiment with the same conditions to compare different algorithms. Therefore, a simulation experiment has been chosen as a way to evaluate the presented approach. The performance of the approach mostly influenced by features of the test problems. We attribute a test problem by changing the size of problem test. The size of a test problem is determined by the number of VMs and PMs. Therefore, the evaluation has been done based on size of test problems by varying number of VMs and PMs. In first experiment, data center comprises 100 PMs, and the variation of VMs start from 600 to In the second experiment, we have fixed number of VMs to 250 VMs and we have varied number of PMs from 50 to 250 to test suitability of our approach. Table III and table IV show our test problems in the experiments. Data center has been simulated with consideration of heterogeneous PMs and VMs, for example, VMs and PMs have different configurations. VMs and PMs configuration has been chosen randomly. The CPU requirements of VMs were generated randomly and the values were taken between 1 and 8 MIPS, the memory requirements were generated randomly from the range [10, 20] GB, the life cycle of a VM chosen randomly from 1 to 100 minutes. The CPU capacities and memory capacities of PMs were randomly picked up from the set of values {10, 11,...,20} MIPS and from the range [20, 40] GB respectively. We have simulated both our approach and FFD algorithm in Java language and has been implemented on a desktop computer which has configuration of Intel Core i CPU of of 3.60 GHz and GB RAM. We have implemented FFD [10] algorithm to provide as baseline algorithm for our evaluations, and compared with our approach based on performance criteria as energy cost. TABLE III EXPERIMENT 1 Test Problem VM PM TABLE IV EXPERIMENT 2 Test Problem VM PM Fig. 1. Experimental results of varying VMs number C. Experimental Results Varying the number of VMs: In the first experiment, we have varied the number of VMs between 600 and 2200; and done the placement to a fixed set of 100 PMs. In Fig. 1, the experimental results for test problems in Table III, show that the presented approach outperforms the FFD algorithm [10] for all test problems. The presented approach also shows its scalability, i.e. the energy consumption increases almost linearly with the increased number of VMs. Therefore, the experimental results conclude that the proposed approach can be applied to find a placement for a large number of VMs due to its scalability with giving energy efficient placement compared to the well-known heuristic algorithm, FFD. In terms of computation time, the computation time of the proposed approach increases almost linearly with the increase of the number of VMs as shown in Fig. 2. However, computation time of finding the placement for 1800 VMs is less than that for 1400 VMs. The reason behind that, the VMs in 1800 VMs test problem have less resources requirements than the VMs in 1400 VMs test problem. As a result, the VMs in 1800 VMs test problem need to check less number of PMs for satisfying the resource constraints and finding the VMs placement compared to the 1400 VMs test problem, and hence giving less computation time than 1400 VMs test problem. From Fig. 2 it is clear that, if all the points of computation times of the proposed algorithm are connected excluding the computation time for 1800 VMs test problem, then it gives a 1050

7 Fig. 2. Computation time of varying VMs number Fig. 4. Computation time of varying PMs number Fig. 3. Experimental results of varying PMs number straight line which indicates that the computation times of the proposed approach for placing different number of VMs, lays around a straight line and a straight line can be drawn using linear regression. This concludes the good scalability of the proposed algorithm. Varying the number of PMs: In the second set of experiments, we have varied the number of PMs between 50 and 250 keeping the VMs to 250 with same attributes throughout the experiments. In Fig. 3, the experimental results for the test problems in Table IV, show that the proposed approach outperforms the FFD algorithm for all test problems and as expected the energy consumption decreases with the number of PMs increased. This is because for more number of PMs the proposed approach finds the set of most energy efficient PMs for the placement, and therefore, energy consumption is minimized for larger data center size with heterogeneous PMs. For the FFD, the energy consumption fluctuates arbitrarily for the test problems. The reason behind this is that, the PMs profiles and VMs profiles were created randomly, and as a result the PMs with larger CPU capacity can be of less energy efficient and vice versa. The FFD chooses the PMs with most CPU capacity first for hosting the VMs, and therefore, for some test problems the selected PMs become more energy efficient and for some test problems the chosen PMs are less energy efficient. Apart from the random energy consumption of FFD, the graph shows that the proposed approach performs better over the FFD for all test problems indicating the suitability of the proposed algorithm for any size data center. Fig. 4 shows computation time graphs for finding placement of a set of VMs in different size data centers. The computation time graph of the proposed approach shows linear tendency with an increased of computation time for 150 PMs test problem compared to its neighbor test problems. This is because due to random creation of PMs profiles, the 150 PMs test problem comprised with PMs with less resources capacities, and as a result each VM in this test problem need to check more number of PMs for its placement and which increased the computation time for the test problem. However, the computation time graph becomes a straight line when a linear regression is applied on the computation time points. Furthermore, the difference in computation times of 50 PMs and 250 PMs test problems is very small, only 60 ms, which indicates a very good scalability of the proposed approach. VI. CONCLUSION AND FUTURE WORK The energy consumption in Cloud data centers is growing up yearly which is a challenge facing by industry and researchers. This paper has presented approach to minimize energy on data center by using a novelty of profile-based VM placement. The approach has adopted FFD to place VMs into PMs. The presented approach has been evaluated through comparing 1051

8 with original FFD algorithm, and the experimental has shown that the presented approach finds the VM placement giving near optimal solution for energy consumption compared to the FFD for two different set of experiments by varying the number of PMs in a data center and by placing a fixed set of VMs in different size data centers. The experimental results also have shown the good scalability of the presented approach. In our future work, a more robust algorithm will be proposed that will consider the real data center traces for comparison and also the dynamic VM placement will be considered. ACKNOWLEDGMENT This research is supported by Shaqra University at Saudi Arabia through the Saudi Arabian Culture Mission in Australia. REFERENCES [1] 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 ACM SIGOPS Operating Systems Review, vol. 37, no. 5. ACM, 2003, pp [2] K. Le, R. Bianchini, J. Zhang, Y. Jaluria, J. Meng, and T. D. Nguyen, Reducing electricity cost through virtual machine placement in high performance computing clouds, in Proceedings of 2011 International Conference for High Performance Computing, Networking, Storage and Analysis. ACM, 2011, p. 22. [3] J. G. Koomey, Worldwide electricity used in data centers, Environmental Research Letters, vol. 3, no. 3, p , [4] J. Hamilton, Cooperative expendable micro-slice servers (cems): low cost, low power servers for internet-scale services, in Conference on Innovative Data Systems Research (CIDR09)(January 2009). Citeseer, [5] A. Beloglazov, J. Abawajy, and R. Buyya, Energy-aware resource allocation heuristics for efficient management of data centers for cloud computing, Future generation computer systems, vol. 28, no. 5, pp , [6] T. K. Sarker and M. Tang, A penalty-based genetic algorithm for the migration cost-aware virtual machine placement problem in cloud data centers, in Neural Information Processing. Springer, 2015, pp [7] Y. Wu, M. Tang, and W. Fraser, A simulated annealing algorithm for energy efficient virtual machine placement, in 2012 IEEE International Conference on Systems, Man, and Cybernetics (SMC). IEEE, 2012, pp [8] B. Jennings and R. Stadler, Survey on virtual machine placement techniques in cloud computing environment, International Journal on Cloud Computing, vol. 23, no. 3, pp , [9] A. Khosravi, S. K. Garg, and R. Buyya, Energy and carbon-efficient placement of virtual machines in distributed cloud data centers, in Euro- Par 2013 Parallel Processing. Springer, 2013, pp [10] Y. Ajiro and A. Tanaka, Improving packing algorithms for server consolidation, in Int. CMG Conference, vol. 253, [11] A. Beloglazov and R. Buyya, Adaptive threshold-based approach for energy-efficient consolidation of virtual machines in cloud data centers, in Proceedings of the 8th International Workshop on Middleware for Grids, Clouds and e-science, vol. 4. ACM, [12] R. Bianchini and R. Rajamony, Power and energy management for server systems, Computer, vol. 37, no. 11, pp , [13] A. Verma, P. Ahuja, and A. Neogi, pmapper: power and migration cost aware application placement in virtualized systems, in Middleware Springer, 2008, pp [14] Z. Yang, L. Liu, C. Qiao, S. Das, R. Ramesh, and A. Y. Du, Availabilityaware energy-efficient virtual machine placement, in 2015 IEEE International Conference on Communications (ICC). IEEE, 2015, pp [15] H. Zhao, Q. Zheng, W. Zhang, Y. Chen, and Y. Huang, Virtual machine placement based on the vm performance models in cloud, in 2015 IEEE 34th International Performance Computing and Communications Conference (IPCCC). IEEE, 2015, pp [16] X. Lin, Z. Liu, and W. Guo, Energy-efficient vm placement algorithms for cloud data center, in International Conference on Cloud Computing and Big Data in Asia. Springer, 2015, pp [17] J. Xu and J. A. Fortes, Multi-objective virtual machine placement in virtualized data center environments, in Green Computing and Communications (GreenCom), 2010 IEEE/ACM Int l Conference on & Int l Conference on Cyber, Physical and Social Computing (CPSCom). IEEE, 2010, pp [18] X. Li, Z. Qian, S. Lu, and J. Wu, Energy efficient virtual machine placement algorithm with balanced and improved resource utilization in a data center, Mathematical and Computer Modelling, vol. 58, no. 5, pp , [19] Z. Zhang, C.-C. Hsu, and M. Chang, Cool cloud: A practical dynamic virtual machine placement framework for energy aware data centers, in Cloud Computing (CLOUD), 2015 IEEE 8th International Conference on. IEEE, 2015, pp [20] M. Vasudevan, Y.-C. Tian, M. Tang, and E. Kozan, Profiling: An application assignment approach for green data centers, in Industrial Electronics Society, IECON th Annual Conference of the IEEE. IEEE, 2014, pp [21] A. Sheikh and S. A. Khan, Integer programming approach for optimal resource allocation in workflow automation design, in 9th International Multitopic Conference, IEEE INMIC IEEE, 2005, pp