Kyu-Sik Chung

Autonomous learning for efficient resource utilization of dynamic VM migration

Dynamic migration of virtual machines on a cluster of physical machines is designed to maximize resource utilization by balancing loads across the cluster. When the utilization of a physical machine is beyond a fixed threshold, the machine is deemed overloaded. A virtual machine is then selected within the overloaded physical machine for migration to a lightly loaded physical machine. Key to such threshold-based VM migration is to determine when to move which VM to what physical machine, since wrong or inadequate decisions can cause unnecessary migrations that would adversely affect the overall performance. We present in this paper a learning framework that autonomously finds and adjusts thresholds at runtime for different computing requirements. Central to our approach is the previous history of migrations and their effects before and after each migration in terms of standard deviation of utilization. We set up an experimental environment that consists of extensive real world benchmarking problems and a cluster of 16 physical machines each of which has on average eight virtual machines. We demonstrate through experimental results that our approach autonomously finds thresholds close to the optimal ones for different computing scenarios and that such varying thresholds yield an optimal number of VM migrations for maximizing resource utilization.

A systematic approach to classifier selection on combining multiple classifiers for handwritten digit recognition

Much research has been done on combining multiple classifiers for handwritten character recognition to improve the performance of the classifier. Given a fixed set of classifiers using the same or different kinds of feature set, they focus on a methodology to combine all of the classifiers. In this paper, given a variable set of classifiers, we focus on a methodology to determine which subset of classifiers achieves the optimal combination results. In order to evaluate the dependency between classifiers, we propose a similarity measure between them which can be calculated from the errors generated by each classifier. This similarity measure allows us to compare the performance of one combination case relative to those of the other cases without performing any experiments. Using five individual neural net classifiers with different feature sets [gradient, structural, UDLRH (up-down left-right hole), mesh and LSF (large stroke feature)], we perform handwritten digit recognition experiments. With three combination methods [majority voting, Borda count and LCA (linear confidence accumulation)], we perform combination experiments for all possible cases of three classifiers selected from among the above five. Then, we compare their rankings in terms of the recognition rate with that in terms of the similarity measure. This comparison shows the effectiveness of the proposed method.

Performance comparison of several feature selection methods based on node pruning in handwritten character recognition

The paper presents a performance comparison of several feature selection methods based on neural network node pruning. Assuming the features are extracted and presented as the inputs of a 3 layered perceptron classifier, we apply the five feature selection methods before/during/after neural network training in order to prune only input nodes of the neural network. Four of them are node pruning methods such as node saliency method, node sensitivity method, and two interactive pruning methods using different contribution measures. The last one is a statistical method based on principle component analysis (PCA). The first two of them prune input nodes during training whereas the last three do before/after network training. For gradient and upper down, left right hole concavity features, we perform several experiments of handwritten English alphabet and digit recognition with/without pruning using the five feature selection algorithms, respectively. The experimental results show that node saliency method outperforms the others.

Video caption image enhancement for an efficient character recognition

For improved recognition of videotexts, we have focused on image enhancement techniques. Since the videotexts are low-resolution and mixed with complex backgrounds, image enhancement is a key to successful recognition of the videotexts. Especially in Hangul characters, several consonants cannot be distinguished without sophisticated image enhancement techniques. In the paper, after multiple videotext frames containing the same captions are detected and the caption area in each frame is extracted, five different image enhancement techniques are serially applied to the image: multi-frame integration, resolution enhancement, contrast enhancement, advanced binarization, and morphological smoothing operations. We have tested the proposed techniques with the video caption images containing both Hangul and English characters from various video sources such as cinema, news, sports, etc. The character recognition results are greatly improved by using enhanced images in the experiment.

A Dynamic Server Power Mode Control for Saving Energy in a Server Cluster Environment

All the servers in a traditional server cluster environment are kept On. If the request load reaches to the maximum, we exploit its maximum possible performance, otherwise, we exploit only some portion of maximum possible performance so that the efficiency of server power consumption becomes low. We can improve the efficiency of power consumption by controlling power mode of servers according to load situation, that is, by making On only minimum number of servers needed to handle current load while making Off the remaining servers. In the existing power mode control method, they used a static policy to decide server power mode at a fixed time interval so that it cannot adapt well to the dynamically changing load situation. In order to improve the existing method, we propose a dynamic server power control algorithm. In the proposed method, we keep the history of server power consumption and, based on it, predict whether power consumption increases in the near future. Based on this prediction, we dynamically change the time interval to decide server power mode. We performed experiments with a cluster of 30 PCs. Experimental results show that our proposed method keeps the same performance while reducing 29% of power consumption compared to the existing method. In addition, our proposed method allows to increase the average CPU utilization by 66%.

A Method for Optimal Bandwidth Utilization in IEEE 802.11 WLAN Networks

This paper proposes a load sharing scheme to maximize network bandwidth utilization in IEEE 802.11 WLAN networks using the SSED (Service Set Identifier) hiding. For this purpose, the proposed scheme keeps checking the available bandwidths of a group of wireless routers; selects the most bandwidth-optimal one; makes it visible to clients. Such that, whenever a client connects to a wireless router, only the selected one is visible to it while others are hiding. We implemented the proposed scheme with modifying the firmware of ASUS WL- 500 G wireless router and performed experiments. Experimental results show 35.6% performance increase in the bandwidth utilization compared to the conventional scheme.

An Improved Estimation Model of Server Power Consumption for Saving Energy in a Server Cluster Environment

In the server cluster environment, one of the ways saving energy is to control server`s power according to traffic conditions. This is to determine the ON/OFF state of servers according to energy usage of data center and each server. To do this, we need a way to estimate each server`s energy. In this paper, we use a software-based power consumption estimation model because it is more efficient than the hardware model using power meter in terms of energy and cost. The traditional software-based power consumption estimation model has a drawback in that it doesn`t know well the computing status of servers because it uses only the idle status field of CPU. Therefore it doesn`t estimate consumption power effectively. In this paper, we present a CPU field based power consumption estimation model to estimate more accurate than the two traditional models (CPU/Disk/Memory utilization based power consumption estimation model and CPU idle utilization based power consumption estimation model) by using the various status fields of CPU to get the CPU status of servers and the overall status of system. We performed experiments using 2 PCs and compared the power consumption estimated by the power consumption model (software) with that measured by the power meter (hardware). The experimental results show that the traditional model has about 8-15% average error rate but our proposed model has about 2% average error rate.

Load Balanced Parallel Prime Number Generator with Sieve of Eratosthenes on Cluster Computers

Algorithms of finding prime numbers up to some integer N by Sieve of Eratosthenes are simple and fast. However, even with the time complexity no greater than O(N In In N), it may take weeks or even years of CPU time if N is large like over 15 decimal digits. No known shortcut was reported yet. We develop efficient parallel algorithms to balance the workload of each computer, and to extend memory limit with disk storage to accommodate Giga-bytes of data. Our experiments show that a complete set of up to 14-digit prime numbers can be found in a week of computing time (CPU time) using our 8 1.6 GHz Pentium-4 PCs with Linux and MPI library. Also, by sieve of Eratosthenes, we think it is very unlikely that we can compute all primes up to 20 digits even using the fastest computers in the world.

An Energy Efficient Cluster Management Method based on Autonomous Learning in a Server Cluster Environment

Energy aware server clusters aim to reduce power consumption at maximum while keeping QoS(Quality of Service) compared to energy non-aware server clusters. They adjust the power mode of each server in a fixed or variable time interval to let only the minimum number of servers needed to handle current user requests ON. Previous studies on energy aware server cluster put efforts to reduce power consumption further or to keep QoS, but they do not consider energy efficiency well. In this paper, we propose an energy efficient cluster management based on autonomous learning for energy aware server clusters. Using parameters optimized through autonomous learning, our method adjusts server power mode to achieve maximum performance with respect to power consumption. Our method repeats the following procedure for adjusting the power modes of servers. Firstly, according to the current load and traffic pattern, it classifies current workload pattern type in a predetermined way. Secondly, it searches learning table to check whether learning has been performed for the classified workload pattern type in the past. If yes, it uses the already-stored parameters. Otherwise, it performs learning for the classified workload pattern type to find the best parameters in terms of energy efficiency and stores the optimized parameters. Thirdly, it adjusts server power mode with the parameters. We implemented the proposed method and performed experiments with a cluster of 16 servers using three different kinds of load patterns. Experimental results show that the proposed method is better than the existing methods in terms of energy efficiency: the numbers of good response per unit power consumed in the proposed method are 99.8%, 107.5% and 141.8% of those in the existing static method, 102.0%, 107.0% and 106.8% of those in the existing prediction method for banking load pattern, real load pattern, and virtual load pattern, respectively.

Dynamic Shutdown of Server Power Mode Control for Saving Energy in a Server Cluster Environment

In order to ensure high performance, all the servers in an existing server cluster are always On regardless of number of real-time requests. They ensure QoS, but waste server power if some of them are idle. To save energy consumed by servers, the server power mode control was developed by shutdowning a server when a server is not needed. There are two types of server power mode control depending on when a server is actually turned off if the server is selected to be off: static or dynamic. In a static mode, the server power is actually turned off after a fixed time delay from the time of the server selection. In a dynamic mode, server power is actually turned off if all the services served in the server are done. This corresponds to a turn off after a variable time delay. The static mdoe has disadvantages. It takes much time to find an optimal shutdown time manually through repeated experiments. In this paper, we propose a dynamic shutdown method to overcome the disadvantages of static shutdown. The proposed method allows to guarantee user QoS with good power-saving because it automatically approaches an optimal shutdown time. We performed experiments using 30 PCs cluster. Experimental results show that the proposed dynamic shutdown method is almost same as the best static shutdown in terms of power saving, but better than the best static shutdown in terms of QoS.