Resource clustering algorithm for cloud data centers

Abstract

Recently, the virtual machine deployment algorithm uses physical machine less or consumes higher energy in data centers, resulting in declined service quality of cloud data centers or rising operational costs, which leads to a decrease in cloud service provider’s earnings finally. According to this situation, a resource clustering algorithm for cloud data centers is proposed. This algorithm systematically analyzes the cloud data center model and physical machine’s use ratio, establishes the dynamic resource clustering rules through k-means clustering algorithm, and deploys the virtual machines based on clustering results, so as to promote the use ratio of physical machine and bring down energy consumption in cloud data centers. The experimental results indicate that, regarding the compute-intensive virtual machines in cloud data centers, compared to contrast algorithm, the physical machine’s use ratio of this algorithm is improved by 12% on average, and its energy consumption in cloud data center is lowered by 15% on average. Regarding the general-purpose virtual machines in cloud data center, compared to contrast algorithm, the physical machine’s use ratio is improved by 14% on average, and its energy consumption in cloud data centers is lowered by 12% on average. Above results demonstrate that this method shows a good effect in the resource management of cloud data centers, which may provide reference to some extent.

Keywords

Cloud data centers Clustering algorithm Machine learning Virtual machine deployment

1. Introduction

As more and more enterprises deploy their businesses in cloud data centers, the demand for cloud data centers is greatly enhanced. To further seize markets and users, all cloud service providers further enlarge their deployment in cloud data centers, especially those large-scale cloud service provides which increase investment to promote their market share. For example, Amazon has deployed 23 cloud data centers around the world to date, involving America, Asia, Europe and Africa, to provide prompt and effective cloud service to global users. However, increased cloud data centers will further aggregate the competitions between cloud service providers.

A mature cloud data center can optimize itself to reduce operating costs, and the optimizing means mainly include improving physical machine’s use ratio and bringing down energy consumption in cloud data centers [1]. To improve physical machine’s use ratio, the most direct way is to deploy more virtual machines in this physical machine and keep it at a full load condition as possible. When a cloud user applies for a new virtual machine from cloud data center, a cloud service provider will deploy the virtual machine to a fixed physical machine according to its parameters [2]. Since the virtual machines applied by different users may have different parameters such as CPU, memory and so on, so the deployment algorithm for cloud data centers may produce different mapping results. To bring down energy consumption in cloud data centers, there are two approaches: static energy conservation and dynamic energy conservation [3]. Static energy conservation approach aims at the low power design of computer hardware and starts from processor framework and circuit board structure. It attempts to utilize low-power circuits and devices to reduce the operating power consumption of physical machine and then bring down the power consumption of cloud data centers [4, 5]. Nevertheless, physical machine runs when multiple devices jointly act, and different devices not only depend on each other but also mutually restrict, so this approach has limited effect on reducing energy consumption [6]. Dynamic energy conservation approach is more flexible, which primarily includes dynamic voltage and frequency scaling(DVFS) technique and optimal resource scheduling algorithm. Though DVFS technique can solve the energy consumption problem of a single physical machine, it can’t generate the optimal result due to local optimization when handling with an entire cloud data center. Optimizing the resource scheduling algorithm is an important way of dynamic energy conservation, because it can promptly respond to the true-time situation of cloud data centers and generate an optimal scheduling result [7, 8, 9, 10]. In general, scheduling algorithm has one or more optimization objectives, including completion time, cloud data centers’ energy consumption and load balancing [11, 2, 13, 14].

The previous scholars have made some research on scheduling system optimization [15, 16, 17], but they haven’t considered the following issues. First, in reality, cloud service providers often provide cloud service to cloud users in a fixed form, which is mainly reflected in virtual machine’s parameters. Normally, the proportion of virtual machine’s CPU core number and memory size is a fixed value: 1:2 in compute-intensive virtual machine, and 1:4 in general-purpose virtual machine. When applying for a virtual machine, the cloud users may increase CPU core number, meanwhile, the memory size will also be increased, but their proportion is fixed. Hence, the cloud users could choose an appropriate virtual machine according to their own needs. Virtual machines with different proportions also raise the difficulties of deploying virtual machines. Second, physical machines with different operational cycles and resource use ratios will show different deployment effect, and different deployment effects may lead to great discrepancy in energy consumption, but the deployment sequence can be adjusted to change the way of deployment. Third, the energy consumption in cloud data centers is another important factor to restrict the development of cloud computing. Too high energy consumption will increase the operational costs of cloud data centers, thereby increasing the cloud users’ expense and reducing the use ratio of cloud data center. Fourth, the properties of virtual machine and physical machine consist of factors such as CPU use ratio, memory size and so on. Most of previous algorithms only consider CPU use ratio or take CPU use ratio as the main factor, thus causing a large waste in other resources of physical machine.

On account of above reasons, this paper proposes a resource clustering algorithm for cloud data centers(RCDC). This algorithm systematically analyzes the cloud data center model and physical machine’s use ratio, establishes dynamic resource clustering rules through k-mean clustering algorithm, and deploy the virtual machines according to the clustering results, as so to reduce energy consumption in cloud data centers.

2. Cloud data centers model and rules

2.1 Cloud data centers model

This section first introduces the framework of cloud data centers. The cloud data centers studied in this paper refer to the true data of Amazon’s virtual machines and physical machines. The main considerations include virtual machines’ CPU core number, memory size and lease time. Therefore, assume this cloud data center contains the physical machine set PM $=\{\text{pm}_{1},\text{pm}_{2},\ldots,\text{pm}_{n}\}$ , in which, n represents the number of physical machines owned by this cloud data center. The parameters of each physical machine are shown in Table 1.

Table 1
Parameters of physical machines

]-8pt] Parameters	Definition	Unit
pm ${}_{i}$ .tc	Number of CPU cores contained	/
pm ${}_{i}$ .tm	Memory size contained	GB
pm ${}_{i}$ .lc	Number of residual CPU cores	/
pm ${}_{i}$ .lm	Residual memory size	GB
pm ${}_{i}$ .at	Activated moment	/
pm ${}_{i}$ .rt	Operating time	h
pm ${}_{i}$ .et	End moment	/
s pm ${}_{i}$ .st	State (closed $=$ 0, activated $=$ 1)	/
pm ${}_{i}$ .cn	Subordinating cluster number	/
pm ${}_{i}$ .lt	Light-load operating time	min

Similarly, virtual machines applied by cloud users can also be represented as a virtual machine set. VM $=\{\text{vm}_{1},\text{vm}_{2},\ldots,\text{vm}_{m}\}$ , in which, m represents the number of virtual machines in this set. Cloud data centers use scheduling algorithm to deploy virtual machines in the set to physical machines, see each virtual machine in Table 2.

Table 2

Parameters of virtual machines

Parameters	Definition	Unit
vm ${}_{j}$ .tc	Number of CPU cored needed	/
vm ${}_{j}$ .tm	Memory size needed	GB
vm ${}_{j}$ .at	Establishing moment	/
vm ${}_{j}$ .rt	Operating time	h
vm ${}_{j}$ .et	Destroy moment	/

2.2 Dynamic k-means clustering algorithm

The clustering issue is an important research direction in unsupervised learning. The virtual machine’s mapping problem in this paper can be classified into a clustering issue. So, the clustering algorithm in unsupervised learning can be taken as reference to cluster the activated physical machine and form a cluster.

Once a cloud user submits the application for virtual machine lease, the scheduling algorithm in cloud data centers will depend on the current operating condition of physical machines to select suitable physical machines for deployment. Taking six physical machines in cloud data centers as an example, they are all isomorphic physical machines based on Intel Xeon P-8175M. Hence, their CPU core numbers are both 16, and their corresponding memory sizes are both 64 GB. See parameters in Table 3. Use k-means algorithm and set k value as 2. See the clustering results in Fig. 1.

Table 3
Examples for physical machines clustering

PM	lc	lm	at	rt	et
pm ${}_{1}$	8	32 GB	1000	3500	4500
pm ${}_{2}$	6	24 GB	1000	3600	4600
pm ${}_{3}$	2	8 GB	1000	3700	4700
pm ${}_{4}$	12	48 GB	1000	10600	11600
pm ${}_{5}$	10	40 GB	1000	10700	11700
pm ${}_{6}$	8	32 GB	1000	10800	11800

Figure 1.

Clustering results of physical machines.

The K-means clustering algorithm is an iterative solution clustering analysis algorithm. Its steps are to pre-divide the data into K groups, then randomly select K objects as the initial cluster centers, and then calculate each object and each seed. The distance between cluster centers, each object is assigned to the nearest cluster center. The cluster centers and the objects assigned to them represent a cluster.

2.3 Resource clustering rules

After physical machines clustering is finished, when there is new application for virtual machines, the scheduling algorithm in cloud data centers could deploy them reasonably according to the clustering results. At this time, not only the operating time of physical machines but also their use ratio should be considered. To increase the operating time of physical machines would lead to the increase in energy consumption of these physical machines, thereby raising the energy consumption in cloud data centers. A low use ratio of physical machines would make more physical machines be activated, resulting in a waste of system performance. In view of above, we propose the dynamic resource clustering rules.

2.3.1 Rule 1-Dynamic resource clustering rule

When deploying the applications for a new virtual machine, the resource management system in cloud data centers will schedule according to the activated physical machines’ clustering results. First, figure out the destroy moment of such virtual machine and its distance to the clustering center of physical machine, and then cluster this virtual machine into the nearest cluster. Second, find out a physical machine in this cluster which residual resources could meet the need of this virtual machine. In case that there is a physical machine in this cluster that can meet the needs, the virtual machine should be deployed in the physical machine that may maximize the use ratio. If there is no such physical machine that can meet the needs, a new physical machine should be activated to complete the deployment of such virtual machine. Finally, update the physical machine clustering center.

The parameters of physical machine in Table 3 are still used for demonstration. Their clustering results are shown in Fig. 1. Assume the parameters of virtual machine are shown in Table 4, which includes the application for two virtual machines, and moments are established based on these to cluster and deploy the virtual machine. First, according to destroy moment of vm ${}_{1}$ , use dynamic clustering rules to cluster it, and cluster such virtual machine into the second cluster. Meanwhile, to maximize the use ratio of physical machine, choose pm ${}_{6}$ to deploy. Since no new physical machine is activated and all physical machines haven’t reached the end moment, so it is unnecessary to update the clustering results. Second, cluster vm ${}_{2}$ , and cluster this virtual machine into the second cluster too. Meanwhile, to maximize the use ratio of physical machines, pm ${}_{6}$ is still used to deploy. Similarly, since no new physical machine is activated and all physical machines haven’t reached the end moment, so it is unnecessary to update the clustering results. So far, two virtual machines are deployed.

Table 4
Parameters of virtual machines to be deployed

VM	tc	tm	at	rt	et
vm ${}_{1}$	4	16 GB	3500	2	10700
vm ${}_{2}$	2	8 GB	4000	2	11200

In the above example, dynamic resource clustering rules are used to complete the deployment of two virtual machines, which doesn’t increase the operating time of physical machine, but improves the use ratio of corresponding physical machines, thereby reducing the energy consumption in cloud data centers. The next section will introduce the core algorithm RCDC of this paper.

3. RCDC algorithm

Above section firstly analyzes the cloud data center model, then further studies dynamic k-means clustering algorithm, finally proposes to use dynamic resource clustering rules in the clustering and deployment of virtual machine. By using this rule and machine learning theory, a Resource Clustering Algorithm for Cloud Data Centers(RCDC) is put forward. This algorithm can reduce the energy consumption in cloud data centers by improving the use ratio of physical machine.

Based on above-mentioned conclusions, RCDC algorithm mainly includes seven steps as below:

Step 1:
Cluster the activated physical machine set;
Step 2:
Choose virtual machines successively according to the their establishing moment in virtual machine application queue, and cluster them using dynamic resource clustering rules.
Step 3:
Find out whether there is a physical machine in the cluster of this virtual machine that can meet the needs of virtual machine. If not, go to Step 4. If there is only one such physical machine, go to Step 5. If there are more than one such physical machine, go to Step 6.
Step 4:
Activate new physical machines and deploy the virtual machine. Simultaneously, update the clustering results. Then, go to Step 7.
Step 5:
Deploy virtual machines in this physical machine. Then, go to Step 7.
Step 6:
Choose the physical machine that can maximize use ratio to complete the deployment of such virtual machine. Then, go to Step 7.
Step 7:
Judge if there is undeployed virtual machine. If yes, back to Step 2. Otherwise, end the entire clustering algorithm.

From above steps, we can find that RCDC algorithm can deploy the virtual machine into the virtual machine that can maximize the use ratio of physical machines without extending the operating time of physical machine, thereby improving the use ratio of physical machines and further reducing the energy consumption in cloud data centers. In the next section, we will further validate its efficiency by experiment.

Table 5
Experimental parameters

Parameters Value

Number of virtual machines {10000, 20000, 30000, 40000, 50000}

Number of physical machines {10000, 20000, 30000, 40000, 50000}

Proportion of virtual machines’ CPU and memory size {2(computing-intensive), 4(general-purpose)}

Number of virtual machines’ CPU {2, 4, 8, 16}

Figure 2.
The change of physical machines’ use ratio with the number of virtual machine (computing-intensive).

Figure 3.
The change of physical machines’ use ratio with the number of virtual machine (general-purpose).

4. Experimental results

Parameters	Value
Number of virtual machines	{10000, 20000, 30000, 40000, 50000}
Number of physical machines	{10000, 20000, 30000, 40000, 50000}
Proportion of virtual machines’ CPU and memory size	{2(computing-intensive), 4(general-purpose)}
Number of virtual machines’ CPU	{2, 4, 8, 16}

In this section, to verify the effect of RCDC algorithm, we use CloudSim to implement this algorithm. Meanwhile, we simulate and implement FF algorithm, BF algorithm and EPS algorithm in CloudSim [18]. By reference to parameters of true virtual machines and physical machines provided by Amazon, the Amazon-customized AWS Graviton2 CPU-based physical machine is used to deploy virtual machines, which power consumption is 0.256 W. See other parameters in Table 5. This section will validate the average use ratio of physical machine and the energy consumption of cloud data centers. The energy consumption of cloud data centers is shown in Eq. (1). The definitions of $pm_{i}$ and $rt_{i}$ have been described in Table 1. The $p_{i}$ is the power consumption of the physical machine.

$\displaystyle\text{Total\_energy}=\sum_{i=1}^{n}{\left({pm_{i}rt_{i}*p_{i}/360% 0}\right)}$ (1)

4.1 Change of algorithmic effect with the number of virtual machine

In this part, we validate the change of physical machines’ use ratio and cloud data centers’ energy consumption with the number of virtual machines. The number of virtual machines’ CPU is set as 2, and virtual machines are computing-intensive and general-purpose types. See other parameters in Table 5.

Figure 4.

The change in energy consumption of cloud data centers with the number of virtual machines (computing-intensive).

Figure 5.

The change in energy consumption of cloud data centers with the number of virtual machines (general-purpose)..

It can be seen from Figs 2 and 3 that, with the change of the number of virtual machines, the average use ratio of physical machines basically remains stable in four algorithms. Compared to FF algorithm, BF algorithm and EPS algorithm, the average use ratio in RCDC algorithm is improved by 13%, 13% and 8%. This is mainly because RCDC algorithm cluster the virtual machines first when deploying them, so that a virtual machine could only be deployed into the physical machine with similar destroy moment. Meanwhile, the physical machine that can maximize use ratio is prior selected to deploy. By this way, no physical machine will extend operating time or lower its use ratio due to the deployment of virtual machine, thus the ensemble average use ratio of physical machine is enhanced ultimately.

Figure 6.

The change of physical machines’ use ratio with the capacity of virtual machines (computing-intensive).

Figure 7.

The change of physical machines’ use ratio with the capacity of virtual machines (general-purpose).

From Figs 4 and 5, it can be seen that the energy consumption of cloud data centers in four algorithms increases with the rising number of virtual machines. Compared to FF algorithm, BF algorithm and EPS algorithm, RCDC algorithm reduces by 11%, 11% and 7%, respectively. The unactivated physical machines are in inactive state, so it is unnecessary to calculate their power consumption. Therefore, the power consumption mainly depends on the operating time of activated physical machines. As RCDC algorithm adopts dynamic resource clustering rule in the deployment of virtual machines, such rule clusters these virtual machines into the cluster closest to their destroy moment, thereby lessening the possibility that the operating time of physical machine is extended to the largest extent. Finally, the operating time of physical machine and the energy consumption in cloud data centers are lower than those of the contrast algorithms.

Figure 8.

The change of cloud data centers’ energy consumption with the capacity of virtual machines (computing-intensive).

Figure 9.

The change of cloud data centers’ energy consumption with the capacity of virtual machines (general-purpose).

4.2 The change of algorithmic effect with the capacity of virtual machine

In this part, we will verify the change of the physical machines’ use ratio and cloud data centers’ energy consumption with the capacity of virtual machines. The number of virtual machines is set as 50000, including computing-intensive and general-purpose types. See other parameters in Table 5.

From Figs 6 and 7, it can be seen that the average use ratio of physical machines in four algorithms basically remain stable with the capacity of virtual machines. Compared to FF algorithm, BF algorithm and EPS algorithm, the use ratio in RCDC algorithm is improved by 23%, 23% and 18%. This is mainly because RCDC algorithm aims at maximizing the use ratio of physical machines when deploying virtual machines in the cluster, so as to enhance the ensemble average use ratio of physical machines as possible.

From Figs 8 and 9, it can be seen that the energy consumption of cloud data centers in the four algorithms increase with the rising capacity of virtual machines. Compared to FF algorithm, BF algorithm and EPS algorithm, the use ratio in RCDC algorithm is improved by 18%, 18% and 17%. RCDC algorithm can utilize dynamic resource clustering rules to cluster virtual machines into the cluster closest to their destroy moment, so as to reduce the operating time of physical machines and the energy consumption in cloud data centers.

5. Conclusions

Through analyzing the cloud data center model and physical machines’ use ratio systematically, using k-means clustering algorithm to establish dynamic resource clustering rules, and deploying virtual machines according to the clustering results, a Resource Clustering Algorithm for Cloud Data Center(RCDC) is proposed. When deploying virtual machines with this algorithm, on the premise of not extending the operating time of physical machines, the use ratio of physical machines is maximized, so as to improve use ratio and reduce energy consumption in cloud data centers. Finally, we use CloudSim to implement this algorithm. Compared to existing algorithms, RCDC algorithm can effectively improve physical machines’ use ratio and reduce energy consumption in cloud data centers.

Footnotes

Fund project

Youth Science Foundation Project of National Natural Science Foundation of China:Research on performance optimization of blockchain data communication considering trust and weight(61802301); Shaanxi Provincial Natural Science Basic Research Plan – General item (2019JQ-056); Xi’an Science and Technology Planning Project (2020KJRC0101).

References

Hassan

H.A.

Salem

S.A.

and Saad

E.M.

, A smart energy and reliability aware scheduling algorithm for workflow execution in DVFS-enabled cloud environment, Future Generation Computer Systems 112 (2020), 431–448.

Tseng

F.H.

Jheng

Y.M.

Chou

L.D.

Chao

H.C.

and Leung

V.C.M.

, Link-aware virtual machine placement for cloud services based on service-oriented architecture, IEEE Transactions on Cloud Computing 8 (2020), 989–1002.

Chaisiri

Bu-Sung

and Niyato

, Optimal virtual machine placement across multiple cloud providers, in 2009 IEEE Asia-Pacific Services Computing Conference (APSCC), 2009, pp. 103–110.

Guo

Gong

Fang

Khargonekar

P.P.

and Geng

, Energy and network aware workload management for sustainable data centers with thermal storage, IEEE Transactions on Parallel and Distributed Systems 25 (2014), 2030–2042.

Rjoub

Bentahar

and Wahab

O.A.

, BigTrustScheduling: Trust-aware big data task scheduling approach in cloud computing environments, Future Generation Computer Systems 110 (2020), 1079–1097.

Barik

Farooqui

Lewis

B.T.

and Shpeisman

, A black-box approach to energy-aware scheduling on integrated CPU-GPU systems, in 2016 IEEE/ACM International Symposium on Code Generation and Optimization (CGO), 2016, pp. 70–81.

Ferreto

T.C.

Netto

M.A.S.

Calheiros

R.N.

and Rose

S.A.F.D.

, Server consolidation with migration control for virtualized data centers, Future Gener Comput Syst 27 (2011, 1027–1034.

Zhu

Yang

L.T.

Chen

Wang

Yin

and Liu

, Real-time tasks oriented energy-aware scheduling in virtualized clouds, IEEE Transactions on Cloud Computing 2 (2014), 168–180.

Ding

Fan

Zhao

Kang

Yin

and Zeng

, Q-learning based dynamic task scheduling for energy-efficient cloud computing, Future Generation Computer Systems 108 (2020), 361–371.

10.

Guo

Min

and Zhang

, Cost-minimizing bandwidth guarantee for inter-datacenter traffic, IEEE Transactions on Cloud Computing 7 (2019), 483–494..

11.

Liang

Dong

Wang

and Zhang

, Memory-aware resource management algorithm for low-energy cloud data centers, Future Generation Computer Systems 113 (2020), 329–342.

12.

Liang

Dong

Wang

and Zhang

, A low-power task scheduling algorithm for heterogeneous cloud computing, The Journal of Supercomputing 76 (2020), 7290–7314.

13.

Zhu

and Tang

, Collaborative optimization of service composition for data-intensive applications in a hybrid cloud, IEEE Transactions on Parallel and Distributed Systems 30 (2019), 1022–1035.

14.

Haratian

Safi-Esfahani

Salimian

and Nabiollahi

, An adaptive and fuzzy resource management approach in cloud computing, IEEE Transactions on Cloud Computing 7 (2019), 907–920.

15.

WC.

and Psounis

, Resource-constrained replication strategies for hierarchical and heterogeneous tasks, IEEE Transactions on Parallel and Distributed Systems 31 (2020), 793–804.

16.

Farahnakian

Pahikkala

Liljeberg

Plosila

Hieu

N.T.

and Tenhunen

, Energy-aware VM consolidation in cloud data centers using utilization prediction model, IEEE Transactions on Cloud Computing 7 (2019), 524–536.

17.

Yuan

and Tian

, Practical privacy-preserving mapreduce based k-means clustering over large-scale dataset, IEEE Transactions on Cloud Computing 7 (2019), 568–579.

18.

Wang

Dou

and Yu

, refsourceA Virtual Machine Scheduling Method for Trade-Offs Between Energy and Performance in Cloud Environment, 2016.

Resource clustering algorithm for cloud data centers

Abstract

Keywords

1. Introduction

2. Cloud data centers model and rules

2.1 Cloud data centers model

Table 1 Parameters of physical machines

Table 3 Examples for physical machines clustering

2.3.1 Rule 1-Dynamic resource clustering rule

Table 4 Parameters of virtual machines to be deployed

5. Conclusions

Footnotes

Fund project

References

Table 1
Parameters of physical machines

Table 3
Examples for physical machines clustering

Table 4
Parameters of virtual machines to be deployed