International Journal of Computer Applications (0975 – 8887) Volume 73– No.11, July 2013 18 Overload Avoidance Model using Optimal Placement of Virtual Machines in Cloud Data Cetres Narander Kumar Department of Computer Science, B. B. Ambedkar University, Lucknow, U.P., India. Shalini Agarwal Department of Computer Science, B. B. Ambedkar University, Lucknow, U.P., India. Vipin Saxena Department of Computer Science, B. B. Ambedkar University, Lucknow, U.P., India. ABSTRACT Cloud data centres improve CPU utilization of their servers (physical machines or PMs) through Virtualization (virtual machines or VMs). Over virtualised and under virtualized PMs suffer performance degradation and power dissipation respectively. This work presents a stochastic modular scheme for allocating VM requests to a PM by avoiding overloading of PM and keeping the global load characteristics under specified QoS goal. The proposed approach categorizes PMs into three groups (Under Load, Normal Load, Over Load) in a way that minimizes number of PMs in Under Load and Over Load groups and maximizes number of PMs in Normal Load group. We compute VM request rejection probability, response time, service time and number of PMs that are overload or under loaded for evaluating the performance of our model. The results show that these parameters do not degrade with increasing arrival rate. Thus the proposed model is simple yet efficient approach for VM placement problem. Keywords Markov Chain, Virtual Machine, Physical Machine, Virtual Chunks, VM Consolidation. 1. INTRODUCTION Cloud Computing is based on Autonomic Computing Model with access to a vast and shared pool of resources (e.g. servers, storage, applications, services) that are subject to rapid provisioning and deployment at the user‟s site with minimal effort on the part of resource service provider. It is becoming a pervasive model [1] in which achieving high reliability and scalability of applications and enabling the cloud resource provider to maximize CPU utilization subject to the constraints imposed by the need to optimize QoS are certain conflicting objectives [1][12][16]. Therefore, developing a model that captures resiliency and yet be tractable, is an area of active research [3]. The integral component of the next generation cloud data centers is the virtualization [9][10] of compute and storage resources which enables a number of virtual machines (VMs) to be deployed and scaled on a single physical machine (PM) or host, corresponding to the input workload. A VM refers to the software implementation of a computer that runs its own OS and applications as if it were a physical machine. Most cloud data centers implement automatic VM placement algorithms [2][4][5][8][9][10][12][15][19][20][22][23], in which the most suitable host is selected by categorizing VM resource requirements and its anticipated expansion while optimizing the placement goals. There are numerous placement goals which include maximizing the usage of available resources, power saving by switching idle hosts to sleep mode, meeting SLA and optimizing server consolidation [11][14][16][25]. In order to meet one or more of the placement goals and depending upon the dynamic workload submitted to the hosts, the placement controller component allows VMs to move from one physical machine to another, under state consistency preservation constraints. This form of movement is termed as live migration, in the literature[11][14][16][25](a sub process of VM consolidation), which is desirable, yet, being a resource intensive operation, consumes several CPU cycles and appreciable network bandwidth, affects the performance of applications as well as resource usage of the migrating and collocated VMs. Therefore, a model that avoids overloading and under loading of PMs directly influences VM migrations. PMs become under loaded or overload by dynamic departure and deployment of VMs to PMs [13]. This work proposes a VM placement algorithm that places VMs to PMs in such a way so as to keep the number of overloaded and under loaded PMs to a minimum so that the interval between successive VM consolidations is maximized. The VM placement algorithm is described as a continuous time Markov Chain model, which has the knowledge of the current and future load characteristics of the PMs in the cloud data centre in terms of computed probabilities. The model is essentially a collection of interacting sub-models that exchange their outputs to minimize VM request rejection probability and the response time. The proposed model efficiently represents the dynamics of today‟s cloud centers as compared to the earlier monolithic models that were more restrictive in nature. The rest of the paper is organized in the following manner. In section 2, we discuss the related work, followed by a CPU utilization model in section 3. We present System Model of VM placement technique in section 4 and the associated sub models with mathematical formulations in sections 5, 6, 7, 8 respectively Section 9 presents the VM placement Algorithm. Simulation and numerical result analysis is given in section 10. Finally we give conclusion and future directions in section 11. 2. RELATED WORK Stochastic models for data centre performance have been proposed in literature [1][5][6]. In [1], the authors proposed a statistical model for PM overload detection for stationary workload that maximizes mean inter migration time. The authors give the categorization of dynamic VM consolidation as periodic, heuristic-based and thresh-hold based. A degree constraint is introduced in [2], and using this constraint a model of virtual machine allocation problem is developed. However, after theoritical modeling several heuristics has been proposed to solve the on-line version of the problem. An efficient and responsive economic resource allocation in high-