雲端大型多人線上遊戲遊戲下基於負載預測之資源分配方法

國

立

交

通

大

學

資訊學院資訊科技（IT）

產業研發碩士專班

碩

士

論

文

雲端大型多人線上遊戲遊戲下基於負載預測之

資源分配方法

SVM-based load prediction for resource allocation in MMOG clouds

研 究 生：吳昭霖

指導教授：王國禎 博士

雲端大型多人線上遊戲遊戲下基於負載預測之資源

分配方法

SVM-based load prediction for resource allocation in

MMOG clouds

研 究 生：吳昭霖 Student：Zhao-Li Wu

指導教授：王國禎 Advisor：Kuochen Wang

國 立 交 通 大 學

資訊學院資訊科技（IT）產業研發碩士專班

碩 士 論 文

Submitted to College of Computer Science National Chiao Tung University in Partial Fulfillment of the Requirements

for the Degree of Master

in

Industrial Technology R & D Master Program on Computer Science and Engineering

雲端大型多人線上遊戲下基於負載預測之資源分配方法

學生：吳昭霖 指導教授：王國禎 博士

國立交通大學資訊學院產業研發碩士專班

摘 要

近年來大型多人線上遊戲(MMOG)已成為主流遊戲，其使用者超過數百萬

人。對大型多人線上遊戲來說，值得研究的課題非常的多，其中又以資源

分配為重要的研究課題之一。如何有效的進行資源分配使遊戲可以正常運

作，又可使其滿足使用者之QoS，如讓使用者滿意之反應時間，是為產業與

學界共同研究的課題。本論文利用支持向量機來進行資源預測並結合雲端

的資源分配，以便有效地進行資源分配以提升使用者之QoS。支持向量機能

運算大量的資料並具備快速計算的能力，使其被廣泛地運用在各種預測上。

由於其預測準確率高於類神經網路，因此更能有效用於MMOG負載的預測。

我們的方法是基於支持向量機於多伺服器架構下進行MMOG負載的預測，即

利用每個地圖區域的虛擬機器之歷史負載量來預測該區域未來負載量的層

級，從而進行有效的資源分配，以滿足使用者反應時間的要求。實驗結果

顯示，我們基於支持向量機的預測方法在預測準確率上比基於類神經網路

的預測方法高12.24%，且其在虛擬機器的使用數量上也減少8%。

關鍵詞：雲端計算、負載預測、大型多人線上遊戲、類神經網路、資源分

配、支持向量機。

SVM-based load prediction for resource

allocation in MMOG clouds

Student: Zhao-Li Wu Advisor: Dr. Kuochen Wang

Industrial Technology R & D Master Program of Computer Science College

National Chiao Tung University

Abstract

In recent years, massively multiplayer online games (MMOGs) have hundreds to

thousands of active concurrent players so as to have become a popular research topic in

academia and industry. MMOGs consume huge resources due to a massive number of players.

To utilize resources efficiently, we integrate MMOGs with cloud computing environments.

An MMOG game world is composed of game regions. MMOG load management allocates

resources (i.e. virtual machines, VMs) to each game region. In this paper, we propose an

SVM (support vector machine)-based load prediction scheme to first forecast cloud resources

needed based on the loading class of each game region in an MMOG cloud. Then we allocate

resources needed to achieve reasonable response time with less resources used. Simulation

results show that the proposed SVM-based load prediction is 12.24% better than neural

network-based load prediction in term of prediction accuracy. In addition, the proposed

SVM-based load prediction reduces 8% of the number of VMs used compared with the neural

network-based load prediction.

Keywords: cloud computing, load prediction, MMOG, neural network, resource allocation,

Acknowledgements

Many people have helped me with this thesis. I am in debt of gratitude to my thesis

advisor, Dr. Kuochen Wang, for his intensive advice and guidance. I would also like to show

my appreciation for all the classmates in the Mobile Computing and Broadband Networking

Laboratory for their invaluable assistance and inspirations. The support by the National

Science Council under Grants NSC 99-2221-E-009-081-MY3 is also gratefully acknowledged.

Contents

Abstract (in Chinese)……….……….i

Abstract ... iii

Contents ... v

List of Figures ... vii

List of Tables ... viii

Chapter 1 Introduction ... 1 1.1 Motivation ... 1 1.2 Resource allocation ... 2 1.3 Load prediction ... 2 1.4 Research objective ... 3 1.5 Thesis organization ... 3 Chapter 2 Preliminaries ... 5 2.1 SVM ... 5 2.2 The Applications of SVM ... 6

2.3 Pros and cons comparison of the proposed SVM-based and neural network-based load prediction ... 6

Chapter 3 Related Work ... 8

3.1 Neural network-based load prediction ... 8

3.2 Qualitative comparison of proposed SVM-based with neural network-based load prediction ... 8

Chapter 4 Proposed SVM-based Load Prediction for Resource Allocation ... 10

4.1 SVM-based resource allocation architecture for MMOGs ... 10

4.3 SVM-based load prediction ... 14

4.4 Resource assignment ... 17

Chapter 5 Performance Evaluation ... 18

5.1 Simulation setup and evaluation metrics ... 18

5.1.1 Prediction accuracy ... 19

5.2 Comparison between SVM-based and NN-based load prediction ... 19

Chapter 6 ... 20

Conclusion ... 20

6.1 Concluding remarks ... 20

6.2 Future work ... 21

List of Figures

Figure 1: Resource allocation for each region [13]. ... 3

Figure 2: SVM hyperplane and the support vector [9]. ... 6

Figure 3: Proposed SVM-based resource allocation architecture for MMOGs. ... 10

Figure 4: Preprocessing procedure and the MMOG player datasets format. ... 13

Figure 5: Preprocessing procedure for deriving the loading class of each game region j at time interval i. ... 14

Figure 6: SVM using the loads in previous time intervals to predict the future load of a region. ... 15

Figure 7: Flowchart of the proposed SVM-based load prediction and resource assignment. 16 Figure 8: Prediction accuracy comparison of the three load prediction schemes. ... 20

List of Tables

Table I. Pros and cons comparison of the proposed SVM-based and neural network-based

load prediction ... 7 Table II. Qualitative comparison of the proposed SVM-based with neural network-based

load prediction. ... 9 Table III. MMOG loading level and its SVM category. ... 17 Table IV. Simulation parameters [1,16]. ... 18

Chapter 1

Introduction

1.1 Motivation

In recent years, the development of massively multiplayer online games (MMOGs) has been

booming. In [6], it is reported that over 55% of Internet users are also online gamers, while

the number of MMOG players grows to 21 million in 2011. Today, MMOGs causes

high interest in academia and industry. MMOGs need to consume huge resources due to a

massive number of players. MMOG game servers are responsible for communicating

information among players. The challenges of MMOGs include scalability, reasonable

resource allocation, QoS, fairness among players and meeting response time requirement of

players. An important goal is to allocate or release resources (such as virtual machines - VMs)

to meet the response time requirement of players and to have efficient resource utilization for

power saving. In this paper, we pay attention to the resource allocation problem in MMOGs

and use the multi-server architecture in the cloud. Cloud computing integrates distributed

computing and virtualization techniques. It has powerful computing ability and high

scalability. Cloud computing applications are for distributed computing, data storage, and

sharing of various software. Cloud computing environments are suitable for MMOGs. By the

statistics of WoW (World of Warcraft), there are millions of players online [3]. So we

leverage the powerful computing ability and high scalability of cloud computing

environments to deal with a large number of MMOG players. In an MMOG cloud, it flexibly

uses VMs instead of physical machines to reduce response time and enhance QoS of players.

That is, it can allocate the resources of an MMOG cloud more effectively. In [3], it mentioned

system QoS modeling/prediction, and (iii) player loyalty modeling/prediction. Based on these

reason, we use players’ log files for predicting the VMs load in the proposed scheme.

The load of MMOGs depends on the number of active concurrent players [12]. Therefore,

fast and accurate load prediction algorithms are required to dynamically allocate resources for

MMOGs so as to provision VMs to game sessions with the minimum resources dynamically.

1.2 Resource allocation

A game world is composed of regions. Each region is supported by multiple VMs of

servers.Assume that each VM has the same CPU, memory, and network bandwidth capability.

The methods of resource allocation can be classified into two categories, real-time resource

allocation and pre-allocation [10]. The real-time resource allocation scheme dynamically

assigns resources when the user has the demand. The pre-allocation scheme can achieve better

system resource stability, but a good load prediction algorithm is needed. In this paper, we

combine real-time resource allocation with a load prediction scheme to predict the user load

and allocate resources efficiently. By prediction, resources can be allocated before players’

login to the MMOG servers.

1.3 Load prediction

Load prediction is based on historical data to forecast the load level of each region in

advance for VM assignment. Load prediction can help to achieve efficient real-time resource

allocation. It can help reduce resource allocation time to meet the response time requirement

1.4 Research objective

SVM (support vector machine) is a suitable tool for predicting MMOGs load. In this

paper, we propose an efficient real-time resource allocation scheme to support the workload

generated by MMOGs players. The task involves predicting the load of each region in an

MMOG game world using historic players’ log files. We use SVM to predict MMOGs game

world load. We integrate current MMOG multi-server architecture with a cloud computing

environment to achieve better performance.

1.5 Thesis organization

The rest of this paper is organized as follows. In Chapter 2, we introduce the basic concept

of SVM. In Chapter 3, we review an existing prediction-based real-time resource provisioning

scheme for MMOGs. We present the proposed SVM-based load prediction scheme in detail in

Chapter 4. In Chapter 5, we show simulation results that include comparison with neural

VM

VM

VM

VM

VM VM

VM

VM VM

network-based load prediction [13] in terms of prediction accuracy. In Chapter 6, we conclude

Chapter 2

Preliminaries

2.1 SVM

The support vector machine (SVM) was developed by Vapnik (1995) [1]. The main idea

of SVM comes from the binary classification, namely to find a hyperplane as a segmentation

of the two classes to minimize the classification error. The maximum margin hyperplane

gives the maximum separation between decision classes. The training data are closest to the

maximum margin hyperplane; it is called support vectors [1]. SVM analyzes data and

recognize patterns for classification and regression analysis (i.e. time sequence analysis).

SVM concentrates on applying regression to short-term electricity load forecasting. Its

forecasting accuracy outperforms other forecasting models. SVM has good performance on

reducing the outliers and improves the prediction accuracy efficiently. The SVM parameters

are defined as follows: , where Xi is a training vector,

Yi represents a class. Figure 2 shows the SVM hyperplane and the support vector w [9].

2 }, 1 , 1 { , ..., , 1 }, , {X_i Y_i i n 　Y_i  　X_i R

2.2 The Applications of SVM

SVM can be applied to some applications, such as flood forecasting, power prediction

[18], image recognition (large dataset), three-dimensional classification, text classification,

and time series prediction[1] [18].

2.3 Pros and cons comparison of the proposed SVM-based

and neural network-based load prediction

Table I shows the pros and cons between SVM-based and the NN-based load prediction.

For the SVM-based load prediction, it has low performance with a small amount of data. The

NN-based load prediction has slow training speed, complicated neural nodes, and high

wT _{x + b = [+1, -1]} Class +1 Class -1 w

 

Table I. Pros and cons comparison of the proposed SVM-based and neural

network-based load prediction

Approach Pros Cons

SVM-based (proposed)

• Fast training speed[18] • Easy to implement [1] [18] • Optimal solution [17]

• Low performance with a small amount of data

Neural network (NN)-based [5]

[16]

• Accepting different input data formats [15]

• Ambiguous inference ability [16]

• High resource consumption [15]

• Slow training speed [15] • Complicated neural node

Chapter 3

Related Work

3.1 Neural network-based load prediction

In [5], it proposed a neural network (NN) prediction-based method for dynamic resource

provisioning and scaling of MMOGs in distributed grid environments. The load prediction

method predicts the future game world entity distribution from history trace data by using the

neural network-based method. It developed generic analytical game load models used to

forecast the future hot-spots that congest the game servers and make the overall environment

fragmented and unplayable [5]. Finally, a resource allocation service performs dynamic load

distribution, balancing, and migration of entities that keep the game servers reasonably loaded

such that the real-time QoS requirements are maintained[5].

3.2 Qualitative comparison of proposed SVM-based with

neural network-based load prediction

The NN-based scheme [5] bases neural network to foresee future hot-spots to allocate

resources for the grid MMOG. However, it did not propose any mechanism to allocate

resources and had low prediction accuracy to predict the MMOG load. The above problems

motivate us to design a more efficient SVM-based load prediction for MMOGs. We propose

proposed SVM-based load prediction scheme will be described in Chapter 4. In Table II, the

prediction technique indicates the operation type of the prediction [5] [16]. The fairness

indicates that the same number of VMs is allocated to each game region at first. The

prediction accuracy indicates the fraction of prediction correctness for each load prediction

scheme.

Table II. Qualitative comparison of the proposed SVM-based with neural

network-based load prediction.

Method Prediction technique Fairness among players Prediction Accuracy SVM-based load prediction (proposed) SVM Yes 90% - 95% Neural network (NN)-based[5] load prediction Neural network Yes 82% - 87%

Chapter 4

Proposed SVM-based Load Prediction

for Resource Allocation

4.1 SVM-based resource allocation architecture for

MMOGs

The proposed SVM-based resource allocation architecture for MMOGs is shown in Figure 3.

We obtain MMOG player history datasets from a character server. In a monitor server, there

are three main components preprocessing module, SVM-based prediction module, and

resource assignment module. The first component, preprocessing module, calculates the

SVM-based prediction

Resource Assignment MMOG player history dataset

Character server Preprocessing Monitor server Resource Allocation Datacenter

players in MMOGs. After load prediction, we allocate the resources to each region in the

datacenters of the MMOG cloud.

4.2 Preprocessing

MMOG loads are dynamic because of the player interaction [8]. Thus, accurate and fast load

prediction is necessary to dynamically allocate resources for MMOGs. In our MMOG cloud

system, we obtain MMOG player history datasets from the character server. Figure 4 shows

the preprocessing procedure and the MMOG player datasets format. The detailed

preprocessing procedure is described as follows:

Preprocessing procedure

1) Read MMOG player history datasets from the character server.

2) Count the number of players (NP) in each region at each time interval.

3) Calculate the loading class of each region.

The formula of estimating the loading level of region j at time interval i, Lj :

ij ij NV NP ij L  ₍₁₎

where NPij = total number of players in region j at time interval I and NVij = number of VMs

used in region j at time interval i. We define MAX_LOAD vm = the maximal number of

players in a VM. The upper threshold of Lij is Tu and the lower threshold of Lij is Tl, which are

defined as follows: vm ij vm ij u MAX LOAD NV LOAD MAX NV T   _  _ (2) ij vm ij l NV LOAD MAX NV T ( 1) _ 1 (3)

The loading state of a region can be: overloading, underloading, or normal.

Underloading: L_ijT_l

(5)

Normal: T_l L_ijT_u

(6)

In addition the loading class of a region is either: class 1 or class 2. Class 1 means the loading

state is overloading or underloading and has the value +1 in SVM. Class 2 means the loading

state is normal and has the value -1 in SVM. The actions taken by each region based on its

loading class is as follows:

If the loading class of a region is class 1, then the number of VMs in this region should

be adjusted,

else if the loading class of a region is class 2, then the number of VMs in this region does

not need to be adjusted.

Figure 5 shows the preprocessing procedure for deriving the loading class of each region j at

START

Count the number of players in each region

at each time interval Read MMOG players log file

A

(a): Preprocessing procedure.

Time interval 1 Region 1 NP1, 1 Region 2 NP1, 2 Region 3 NP1, 3 Region 4 NP1, 4 Region j NP1, j ... . Region 1 NPi, 1 Time interval i Region 2_NP i, 2 Region 3 NPi, 3 Region j NPi, j ... . . Time

interval 2 Region 1_{NP2, 1} Region 2_{NP2, 2}

Region 3

NP2, 3

Region j

NP2, j ...

(b): MMOG player datasets format for recording the number of players in each region at each time interval.

Figure 4: Preprocessing procedure and the MMOG player datasets format.

4.3 SVM-based load prediction

In this section, we describe the proposed SVM-based load prediction for MMOG

resource allocation. As shown in Figure 6, the SVM-based prediction bases on the loads in For each region j (j =1~56), derive the

load (Lij ) of region j at time interval i

Tl ≦ Lij ≦ Tu

Region j at the interval i belongs to class 2

Region j at the time interval i belongs to

class 1 No

End

A

Figure 5: Preprocessing procedure for deriving the loading class of each game

transformed to the SVM format, as shown in section 2.1. Then resources can be assigned to

each region based on predicted loading class. The flowchart of the SVM-based prediction and

resource assignment is shown in Figure 7. Table III summarizes MMOG load levels and their

associated SVM-based loading classes.

tn+1 tn tn-1 t1 Time

SVM-based

Prediction

Figure 6: SVM using the loads in previous time intervals to predict the future load of a

Transform the loading class of each

region into SVM format

Train the history game datasets in

SVM format

Adjust the number of VMs in this region

Calculate load prediction accuracy

END START

For each region j, predict its loading class at tn+1 Read MMOG history game datasets from t1 to tn after preprocessing Region j belongs to class 1 Region j belongs to class 2 Overloading or underloading Normal A A

4.4 Resource assignment

Based on the results of load prediction, we propose the rule of resource assignment to add or

release VMs in each region. To allocate resources in an MMOG cloud efficiently, we assume

that the VMs in the cloud have the same specification. The resource assignment objectives are

to meet the QoS, such as the response time requirement of players and to reduce the waste of

resources. We adjust the number of VMs used in region j at time interval i+1 (NVij), as follows:

NVi+1j = NPij / MAX_LOADvm ₍₇₎

Note that NPij is number of players in region j at time interval i and MAX_LOADvm is the

maximum number of players in the VM.

Table III. MMOG loading level and its SVM category.

MMOG load level Load level

range SVM category

Resource assignment type

Overloading LjTu Class 1 Adjustable

Normal Tl LjTu Class 2 Unchanged

Chapter 5

Performance Evaluation

In this chapter, we first describe simulation setup and evaluation metrics. Then, we compare

the proposed SVM-based load prediction with the NN-based load prediction [5] in terms of

prediction accuracy.

5.1 Simulation setup and evaluation metrics

We ran libsvm[1] for SVM-based load prediction. SVM related simulation parameters are

shown in Table IV. MMOG running environment setting are as follows:

(1) We ran the libsvm [1] software package for MMOG load prediction in Intel 2.67 GHz CPU, 4GB, Windows 7 OS.

Table IV. Simulation parameters [1,19].

Parameter Value

Kernel type Radius Basic Function

Gamma 125

Degree 2

Scale -1 ~ 1

Label 1-12 (time interval) Index 1-56 (region number)

5.1.1 Prediction accuracy

The prediction accuracy is defined as the number of correct game region load predictions

divided by the total number of game region load predictions which is expressed as follows:

₍₇₎

5.2 Comparison between SVM-based and NN-based load

prediction

As shown in Figure 8, the proposed SVM-based load prediction scheme achieves

12.24% and 26.77% better prediction accuracy than the NN-based load prediction and the last

value-based load prediction, respectively. The proposed SVM-based load prediction reduces

8% and 78.3% of the number of VMs needed compared with the neural network-based load

prediction and the last value-based load prediction, respectively.

s prediction load region game of number Total s prediction load region game correct of Number = accuracy Prediction

Chapter 6

Conclusion

6.1 Concluding remarks

In this thesis, we have presented an efficient SVM-based load prediction scheme for

resource allocation in MMOG clouds. We use historic load data of players in each region to

predict the future load. By using the load prediction results, we can allocate (or release) an Figure 8: Prediction accuracy comparison of the three load prediction schemes.

0 10 20 30 40 50 60 70 80 90 100 1002 1003 1004 1005 1006 1007 1008 1009 1010 10 1 1 1012 1013 1014 1015 1016 P re dic tion a c c ura c y (% ) Time Last value-based SVM-based (Proposed) NN-based

SVM-based load prediction reduces 8% of the number of VMs used compared with the neural

network-based load prediction. With a higher prediction accuracy and a smaller number of

VMs used, our SVM-based load prediction is feasible for efficient resource allocation in

MMOG clouds.

6.2 Future work

We will integrate the proposed SVM load prediction scheme to an efficient resource

allocation method for MMOG clouds. In addition, we will combine a fuzzy membership

function to each input datum of SVM and design a fuzzy-based SVM load prediction scheme

Bibliography

[1] Libsvm, software available at http://www.csie.ntu.edu.tw/ cjlin/libsvm, 2001.

[2] C. Hsu, C. Chang, and C. Lin, “A practical guide to support vector classication,”

http://www.csie.ntu.edu.tw/~cjlin/papers/guide/guide.pdf, 2003.

[3] Y. Lee, K. Chen, Y. Cheng, and C. Lei, "World of warcraft avatar history dataset," in the

ACM Multimedia Systems, Feb. 2011.

[4] K. W, “Resource allocation for massively multiplayer online games using fuzzy linear

assignment technique,” in Proc. of 5th IEEE Consumer Communications and Networking

Conference, pp. 1035 – 1039, Jan. 2008.

[5] R. Prodan and V. Nae, “Prediction-based real-time resource provisioning for massively

multiplayer online games,” in the Future Generation Computer Systems, Vol. 22, Issue 3,

pp. 785-793, July. 2009.

[6] Z. Wan and X. Xu. “An analysis of WoW players’ game hours,” in Proc. ACM

SIGCOMM Workshop on Network and System Support for Games, pp. 47-52, 2008.

[7] R. Suselbeck, G. Schiele, and C. Becker. “Peer-to-peer support for low-latency massively

multiplayer online games in the cloud,” in Proc. Network and Systems Support for Games,

pp. 1-2, Apr. 2009.

[8] V. Nae, A. Iosup, and R. Prodan, “Dynamic resource provisioning in massively

multiplayer online games,” in the IEEE Transactions on Parallel and Distributed Systems,

[12] H. Thorsten, B. Thomas, and H. Robert, “A peer-to-peer architecture for massive

multiplayer online games.” in Proc. ACM SIGCOMM Workshop on Network and System

Support for Games, Oct. 2006.

[13] S. Zhi, L. Yang, and Y. Tao. “Short-term load forecasting based on LS-SVM optimized

by bacterial colony chemotaxis algorithm,” in Proc. International Conference on

Information and Multimedia Technology, pp. 306-309, Dec. 2009.

[14] M. Mohandes , “Support vector machines for short-term electrical load forecasting.” in

the International Journal of Energy Research, pp. 335-345, 2002.

[15] http://en.wikipedia.org/wiki/Neural_network.

[16] P. Marco, “Peer-to-peer architecture for real-time strategy MMOGs with intelligent

cheater detection,” in Proc. International ICST Conference on Simulation Tools and

Techniques, pp. 1-8, Mar. 2010.

[17] Y. Li and T. Fang, “Application of fuzzy support vector machines in short-term load

forecasting,” in Proc. International Conference on Rough Sets, Fuzzy Sets, Data Mining,

and Granular computing, pp. 363-367, 2003.

[18] M. Javier and A. Muñoz, “Support vector machines with applications,” in Statist. Sci,

Vol. 21, pp. 322-336, Dec. 2006.

[19] L. Gang, et al. “Short-term load forecasting using support vector machine with SCE-UA

algorithm,” in Proc. International Conference on Natural Computation, pp. 290-294,