Simulation Results

Figure 11, Figure 12, Figure 13 and Figure 14 are the simulation results of our experiments. The results of NN-Load are derived by using a neural network tool provided by MATLAB 7.11.0 (R2010b) to predict server load. The type of the neural network we chose is dynamic time series, which takes d past values and d past prediction results as input and predicts the next value. In the NN-Load experiment, since the length of a time slot is 5 minutes, we set d = 24, which means the inputs are the CPU usages and the prediction result of past two hours and output is the prediction value of next 5 minutes. The NN-Player+DRP experiment uses the same setting of neural network; the difference between NN-Load experiment is that the inputs are 24 past values and 24 past prediction value of player numbers in the next time slot. The predicted number of players is a parameter for load model to compute an estimated load.

22

Figure 11: Maximum number of allocated VMs comparison in all time slots for three different approaches.

Figure 12: Over-allocation rate comparison for three different approaches.

Figure 11 illustrates the number of allocated VMs of three experiments, the points in this figure we plot is the maximum value of each two hours. It can be observed that our approach allocates the same number of VMs as DRP approach at most of time. Figure 12 is the over-allocation rate comparison for three different approaches. This figure shows that NN-Load approach is more resource efficient, but

23

the following two figures will show the defects of NN-Load approach.

Figure 13: Probability of under-allocation for each experiment.

Figure 14: Average rate of under/over allocation.

Figure 13 and Figure 14 show that our proposed scheme decreases the

24

probability of under-allocation. The results of NN-Load experiment shows that the neural network achieve good accuracy in predicting load. However, it has higher probability to under-allocate resources than other two experimental setups. The result of NN-Player+DRP experiment shows that probability of under-allocation will decrease, since load model will overestimate the resource requirements. However, neural network predictor is based on trial-and-error design strategy, so the prediction results cannot fit the actual value perfectly. Thus, there’re still 2.16% under-allocation event occurred. Finally, our proposed DRP-HA scheme can avoid most of under-allocation event, which only 18 events over 4255 records. The average over-allocation rate is 66.58%, which can be handled by single VM.

25

Chapter 5 Conclusion

5.1 Concluding Remarks

In this paper, we have presented a hotspot anticipation (HA) scheme to enhance dynamic resource provisioning (DRP), called DRP-HA for MMOGs in cloud computing environments. If avatars are aggregated into groups and become hotspots in a virtual world, interactions between avatars in hotspots will cause extra load to the zone server. Our proposed DRP-HA employs a finite state machine model to monitor the movement of avatars in a virtual world. By combining the state of each avatar in a game zone with a neural network (NN) predictor to forecast the number of players in the next time slot (called NN-player+DRP-HA), we may figure out the potential workload produced by hotspots, and then allocate appropriate computing resources to support the game zone.

Experimental results have supported that the proposed NN-Player+DRP-HA scheme can avoid most of under-allocation events with an acceptable over-allocation rate. Compared with a representative dynamic resource provisioning method, called NN-Player+DRP, the proposed NN-Player+DRP-HA reduces the probability of under-allocation events from 2.16% to 0.42% (80% improvement) in terms of CPU capacity, under the premise that the CPU over-allocation rate is within the capacity of one VM.

5.2 Future Work

In this paper, we focus on reducing the under-allocation events in terms of CPU capacity by considering the interaction of avatars. In the future, by taking different

26

behavior of avatars that result in different loads into account, we may further reduce under/over-allocation rates of the dynamic resource provisioning for MMOGs.

27

References

[1] Y. Lee; K. Chen; Y. Cheng; C. Lei, “World of Warcraft avatar history dataset,” in Proc. the Second Annual ACM Conference on Multimedia Systems MMSys’11, San Jose, California, USA, Feb. 2011, pp. 123-128.

[2] S. Wang; S. Dey, “Modeling and Characterizing User Experience in a Cloud Server Based Mobile Gaming Approach,” in Proc. IEEE Global Communications Conference GLOBECOM’09, Hilton Hawaiian Village, Honolulu, Hawaii, USA, Dec. 2009, pp. 1-7.

[3] “Boss (video gaming),” [Online.] Available:

http://en.wikipedia.org/wiki/Boss_(video_gaming)

[4] P. Bod´ık, C. Sutton, A. Fox, D. Patterson and M. Jordan, “Response-Time Modeling for Resource Allocation and Energy-Informed SLAs,” in Workshop on Statistical Learning Techniques for Solving Systems Problems (MLSys), Whistler, Canada, 2007.

[5] S. Legtchenko, S. Monnet and G. Thomas, "Blue Banana: resilience to avatar mobility in distributed MMOGs," in Proc. IEEE/IFIP International Conference on Dependable Systems and Networks (DSN), vol., no., June 28 2010-July 1 2010, pp.171-180.

[6] J.L. Miller and J. Crowcroft, "Avatar movement in World of Warcraft battlegrounds," in Workshop on IEEE Network and Systems Support for Games (NetGames), vol., no., 23-24 Nov. 2009, pp.1,6.

[7] V. Nae, A. Iosup and R. Prodan, "Dynamic resource provisioning in massively multiplayer online games," IEEE Transactions on Parallel and Distributed Systems, pp. 380-395, Mar. 2011.

28

[8] D.T. Ahmed and S. Shirmohammadi, "A Dynamic Area of Interest Management and Collaboration Model for P2P MMOGs," IEEE/ACM International Symposium on Distributed Simulation and Real-Time Applications, vol., no., pp.27-34, 27-29 Oct. 2008.

[9] H. Liang, I. Tay, M. F. Neo, W. T. Ooi and M. Motani, “Avatar mobility in networked virtual environments: Measurements, analysis, and implications.”

CoRR, abs/0807.2328,2008.

[10] M. Mao and M. Humphrey, “A Performance Study on the VM Startup Time in the Cloud,” in Proc. IEEE 5th International Conference on Cloud Computing, CLOUD’12, June 2012, pp. 423-430.

[11] Y. Lee and K. Chen, "Is server consolidation beneficial to MMORPG? a case study of World of Warcraft," in Proc. IEEE 3rd International Conference on Cloud Computing (CLOUD) , July 2010, pp. 435-442.