Volume o1 issues in engineering and advanced technology on Mar 2016

OPTIMAL RESOURCE PROVISIONING FOR RAPIDLY INCREASING WORKLOAD IN VIRTUALIZED CLOUD COMPUTING ENVIRONMENTS .Ananthraj K1,Varadharajan S2 PG scholar, Asst Professor Surya Group of Institutions, Vikkravandi. Abstract--As no of users accessing the cloud increasing exponentially it is necessary to provide the end user with quality service in minimal operation cost without degradation in performance. Cloud computing enables the feature of elasticity by dynamically adding or removing virtual machines instantly. However effective resource management in virtualized environment is still a challenging task when the workload increases rapidly or if the workload has never seen before. So the resources allocated to the VMs needs to be re-configured dynamically for optimal resource provisioning. In this paper, we introduce an optimal resource provisioning scheme that automatically reconfigures itself to adapt to rapidly changing workload and runs iteratively to make sure whether resource provisioned to virtual machine is optimal by reducing the resource excessively allocated or allocating resource that is below the requirement. Our proposed system is able to find an optimal configuration which is capable of reducing the cost due to over provisioning, and maintaining higher resource utilization, thus reducing infrastructure and management costs. Keywords-cloud computing: virtualization: virtual machine

1.

INTRODUCTION:

Cloud Computing [1-3] is becoming an increasingly popular enterprise model in which computing resources are made available ondemand to the user as needed. The unique value proposition of Cloud Computing creates new opportunities to align IT and business goals. Cloud computing is essentially a powerful computing paradigm in which tasks are assigned to a combination of connections, software and services accessed over a network. The vast processing power of Cloud Computing is made possible though distributed, large-scale computing clusters, often in concert with server virtualization software, like VMware ESX Server [4] and Xen [5], and parallel processing. This network of servers and connections is collectively known as the Cloud. The computing resources may be maintained within the client enterprise, or made available by a service provider. Computing at the scale of the Cloud allows users to access supercomputer-level computing power. Users can access the enormous and elastic resources whenever they need them. For this reason, Cloud Computing is also described as on-demand computing. Elastic resource provisioning is one of the most attractive features provided by Infrastructure as a Service (IaaS) clouds [2]. Unfortunately, deciding when to get more resources, and how many to get, is hard in the face of dynamically changing application workloads and service level objectives (SLOs) that need to be met. Existingcommercial

IaaS clouds such as Amazon EC2 [2] depend on the user to specify the conditions for adding or removing servers. However, workload changes and interference from other co-located applications make this difficult. Virtual machine (VM) technology enables multiple VMs, each running a traditional operating system (OS) and hosting one or more services, to share the resources on the same physical machine. It relies on a VM monitor, residing in between underlyinghardware and guest OSes, for resource allocation between the VMs. The monitor facilitates ondemand resource reconfiguration in a timely manner. There are reasons for online VM resource reconfiguration. When created from a template or arrival at a new host through live migration [7], VMs may need to be reconfigured so as to be incorporated to the new machine for better performance. Due to the time-varying resource demand of typical network applications, it is usually necessary to re-allocate resources to each hosted VM for overall performance. In service consolidation with heterogeneous applications, it is not easy to figure out the best settings for VMs with distinct resource demands. Modern VM monitors provide a rich set of configurable resource parameters for fine-grained run-time control, which makes the problem more complicated. Server virtualization has a key request for performance isolation. In practice, applications sharing the physical server still have chance to interfere with

Volume o1 issues in engineering and advanced technology on Mar 2016 each other. In [15, 9], the authors showed that bad behaviours of one application in a VM could adversely affect the others’ in Xen [3] VMs. The problem is not specific to Xen, it can also be observed on other virtualization platforms due to the presence of centralized Virtual machine scheduling. The involvement of the virtualization layer in the execution of VMs causes uncertainties in VMs’ performance, which makes the VM configuration problem even harder. We model the behaviour and performance of different types of applications and IT resources to adaptively transform end user’s service requests We use analytical performance (queueing network system model) [5] and workload information to supply intelligent input about system requirements to an application provisioner, which improves the efficiency of the system. Such improvements ensure better achievement of QoS targets, while reducing costs due to improved utilization of IT resources. Workload information allows application provisioner to better understand workload demands in terms of resource needs, hence improving resilience to uncertainties and reducing estimation errors that may lead to unacceptable application performance and resource usage (over/under provisioning). Analytical performance model allows the system to predict the effects of a provisioning schedule on target QoS. This model helps application provisioner to predict what mixes of infrastructures is most suited for a given application and when and how system capacity should be scaled up or down. Such models are simple and still efficient in delivering expected QoS because they are based on information available to the application provisioner. Therefore, aspects related to network connections between physical machines hosting the VMs, as well as other infrastructure-level aspects controlled by infrastructure providers are abstracted because this information is not disclosed to the application provisioner. Reinforcement learning (RL) is a process of learning by interactions with dynamic environment, which generates the optimal control policy for a given set of states. It requires no domain knowledge of the controlled system and is able to generate policies optimizing a long-term goal. RL approaches to the design of computer systems

involve several important issues. It is non-trivial to apply RL methods in real systems due to the exponentially increased state space when systems scale up. In online system management, interaction-based RL policy generation suffers from slow adaptation to new policies. Previous studies showed the feasibility of RL in optimizing server allocation [21, 23], power management [22] and designing self-optimizing memory scheduler [11]. There are no reports so far, to the best of our knowledge, on the use of RL in VM-level resource management. Designing a RL-enabled controller to automate VM configuration process poses unique challenges. In server consolidation, physical machines usually host a large number of heterogeneous VMs concurrently. The resource demands from individual VMs vary over time. The applicability of the RL approaches to large scale VM online auto-configuration problem deserves investigation. In this paper, we propose an iterative RL-based Decision Maker for virtual machine autoconfiguration . The central design of our approach is the use of model-based RL algorithms for scalability and adaptability. We define the reward signal based on the summarized performance of each VM. By maximizing the long run reward, IRL automatically directs the virtual machine configuration to a good (if not optimal) one. In controlled environment, our approach was able to find the best (optimal) configuration for single and multiple VMs running homogeneous workloads. In larger systems, I-RL showed good adaptation of policies in online auto configuration with heterogeneous VMs. Although, there is no optimality guarantee for the derived configurations in large systems, I-RL was able to direct arbitrary initial configuration to a better one without performance penalties in any of the VMs. We also introduce a Self-Adaptive and Self Configured Optimal Resource Provisioning System(SASCORPS) to allocate optimal hardware resource to VM instance by reconfiguring the VM iteratively .The input from the I-RL will passed to SASCORP to understand current load of virtual machines to destroy VM running idle or passing its application instance to others with less workload and also reconfigure VM instance by providing additional resources if seems less beyond the requirement. Our proposed scheme will helps to address the problems lies with current optimal resource configuring methods. Thereby handling the resources efficiently will reduce operational and maintenance cost required, and also helps Data centres to avoid concerns regarding the effective

Volume o1 issues in engineering and advanced technology on Mar 2016 management in order to achieve growth in performance, Availability, Scalability. The rest of the paper is organised as follows. Section 2 describes system architecture and modules associated with proposed solution. Related work is discussed in section 3.Finally; section 4 concludes our paper by describing the advantages by our approach. 2.

OPTIMAL RESOURCE PROVISIONING SCHEME:

In this section we present an overview of our system architecture, and the modules that drive our design. And the challenges in determining optimal configurations of virtual machines in shared environment. The hypervisor provides the guest OS, also called a guest domain in xen based VMM. The hypervisor performs functions such as CPU scheduling, memory mapping and I/O handling for guest domains. 2.1 THE SYSTEM ARCHITECTURE: Optimal resource provisioning in dynamically changing virtualized environments poses various challenges and complexity in provisioning resources. We carefully designed our system to provision optimal resources in rapidly increasing workload or the workload has never seen before.

A. Workload Modelling and Load Balancer: Estimates future demands foe the application instances. This information will passed to the load balancer component; which solves an analytical model based on the observed system performance and predicted load to decide the number of VM instances that should be allocated to an application. B. Iterative RL Based Decision Maker: RL offers two advantages [8]. First, it does not require a model of either the system in consideration or the environment dynamics. Second, RL is able to capture the effect of the delayed consequences in a decision-making task. An iterative based RL decision enable the agent performs trial-and-error interactions with environment, each of which returns an instantaneous reward. The reward information is propagated backward temporarily in repeated iterations, eventually leading to an approximation of the value function. C. Self-Adaptive and Self Configured Optimal Resource Provisioning Module(SASCORP) We explicit optimize resource efficiency by introducing a metric to measure a VM’s capacity settings. The metric synthesizes application performance and resource utilization. When employed as feedbacks, it effectively punishes decisions that violate applications’ SLA and gives users incentives. The learning agent operates on a VM’s running status which is defined on the utilization of multiple resources. We employ a Cerebellar Model Articulation Controllerbased Q table for continuous state representation. The resulted RL approach is robust to workload changes because state on low-level statistics accommodates workload dynamics to a certain extent. 2.1.1 Workload Modelling and Load Balancer (load predictor and performance modeller):

Fig. 1. Proposed mechanism for optimal resource provisioning.

Workload Modelling is the component that is responsible for generating estimation (prediction) of request arrival rate.This information is used to compute the exact number of application instances required for meeting QoS targets and resource utilization goals. Prediction can be based on different information; for example, it can be based on historical data about resources usage, or based on statistical models derived from known application workloads.The workload analyzer alerts

Volume o1 issues in engineering and advanced technology on Mar 2016 the load predictor and performance modeller when service request rate is likely to change. This alert contains the expected arrival rate and must be issued before the expected time for the rate to change, so the load Balancer has time to calculate changes in the system and the application provisioned has time to deploy or release the required VMs. Load Balancer (Load Performance Modeller):

Predictor

operations inside and outside the loop are computed in constant time. Number of iterations in the loop depends on finding the number of required virtualized application instances: maximum number of application instances possible is dependent on both policy applied by the application provider and its previous negotiation with IaaS provider, and minimum number of virtualized application instances is updated during execution.

and

This component models the system as a network of queues whose model parameters are obtained via system monitoring and load prediction models. Monitoring data can be obtained via regular monitoring tools or by Cloud monitoring services such as Amazon Cloud- Watch4. The queuing network model considered by the load predictor and performance modeller is depicted in Figure 1. End-users in the model are represented by the generated requests, whereas application provisioned and application instances are the processing stations for these requests. Application provisioned is modelled to have a M/M/1 request queuing station. On the other hand each virtualized application instance has a M=M=1=k queue. Therefore, inter arrival and service time distributions are exponentially distributed during each specific analysis interval. The k parameter, queue size, is defined according to the negotiated service time (Ts) and execution time of a single request (Tr),. If number of requests in a VM exceeds k, the request is rejected by the admission control system and thus not forwarded to the applicationprovisioner. This guarantees that requests are either rejected or served in a time acceptable by clients 𝑻𝒓

K=𝑻𝒔

When workload analyzer updates the estimation of arrival rate, the load predictor and performance modeller checks whether current pool of virtualized application instances aresufficient to meet QoS. To make this verification, it first obtains current service times for each application instance, which is used along with the estimated arrival rate to predict the overall request response time, rejection rate, resource utilization, and maximum number of VMs. If response time or rejection is estimated to be below QoS, or if the utilization is predicted to be below a minimal utilization threshold, the number of VM instances serving applications is updated according to Algorithm 1.Computing time of Algorithm 1 is dominated by the execution of the repeat loop between lines 4 and 22: all the

Algorithm 1 Adaptive VM provisioning. Data: QoS metrics: Ts and Rej(Gs) Data: Tm: monitored average request execution time Data: k: application instance queue size Data: λ: expected arrival rate of requests Data: MaxVMs: maximum number of VMs allowed Result: m: number of application instances able to meet QoS 1 m current number of application instances; 2 min 1; 3 max MaxVMs; 4 repeat 5 old m←m; 6 λsi ←λ/m; 7 Pr(Sk)← expected rejection in a M/M/1/kqueue given λsi and Tm; 8 Tq ← expected response time in a M/M/1/k queuegiven Pr(Sk),λsi , and Tm; 9 ifPr(Sk) and Tq do not meet QoS then 10 m ← m + m/2; 11 min ← m + 1; 12 if m > max then 13 m ← max; 14 end 15 else if utilization is below threshold then 16 max ← m; 17 m ← min + (max - min)/2; 18 ifm ≤ minthen 19 m ← oldm; 20 end 21 end 22 untiloldm = m ; 23 return m;

2.1.2 Iterative RL based Decision Maker: Reinforcement learning based decision maker runs iteratively until the enough VM instance provisioned to satisfy the given workload metrics to provide optimal resource configuration.The RL-enabled agent performs trialand-errorinteractions with the environment, each of which returns an instantaneous reward. The reward information is propagated backward temporally in repeated interactions, eventually leading to an approximation of the value function. The optimal policy is essentially choosing the action that maximizes the value function in each state. The interactions consist of exploitations and

Volume o1 issues in engineering and advanced technology on Mar 2016 explorations. Exploitation is to follow the optimal policy, while exploration is the selection of a random action. Exploitations allow the agent to select the best decisions each time, whileexplorations capture the change of the environment so as to enable the refinement of existing policy. The VM configuration task fits within the agent environment framework. Consider the agent as an VM controller residing in dom0. The states are VM resource allocations and possible changes to the allocations form the actions. The environment comprises the dynamics underlying the virtualized platform. Each time the controller adjust the VM configurations, it receives performance feedback from individual VMs. A RL problem is usually modelled as a Markov Decision Process (MDP). Formally, for a set of environment states S and a set of actions A, the MDP is defined by the transitionprobability Pa(s, s) = Pr(st+1 = s_|st = s, at = a) that action a in state s at time t will lead to state st+1 at time t + 1 and an immediate reward function R = E[rt+1|st =s, at = a, st+1 = s_]. At each step t, the agent perceives its current state st ∈ S and the available action set A(st).By taking action at ∈ A(st), the agent transits to the next state st+1 and receives an immediate reward rt+1 from the environment. The value function of taking action a in state s can be defined as: Q(s, a) = E {

k ∞ 𝑘=0 𝛾 rt+k+1|st=

s, at = a},

Where 0 ≤ γ <1 is a discount factor helping the Q(s,a)’s convergence The reward function. The long-term cumulative reward is the optimization target of RL. In the VM configuration task, the desired configurations are the ones which optimize system-wide performance. The immediate rewards are the summarized VM(s) performance feedbacks on the resulted new configuration. The performance of individual VM is measured by a score which is the ratio of currentthroughput (thrpt) to a reference throughput plus possible penalties when response time (resp) based SLAs (Service Level Agreement) are violated:

Score =

𝑡𝑕𝑟𝑝𝑡 𝑟𝑒𝑓 _𝑡𝑕𝑟𝑝𝑡

-penalty,

0, 𝑖𝑓𝑟𝑒𝑠𝑝 < 𝑆𝐿𝐴; Penalty = 𝑓 𝑥 = 𝑟𝑒𝑠𝑝 , 𝑖𝑓𝑟𝑒𝑠𝑝 ≥ 𝑆𝐿𝐴.

under SLA constraints in current hardware settings. We obtained the reference for one application by dedicating the physical hostand giving more than enough resources to the corresponding VM. A low score indicates either lack of resource or interference between VMs, both of which should be avoided in making allocation decisions. Then, the immediate reward is the summarized scores over all VMs. As suggested by virtualization benchmarks [6, 28] for summarized performance, we define the reward as: 𝑛

Reward=𝑓 𝑥 =

𝑛 𝑖=1 𝑤𝑖

∗ 𝑠𝑐𝑜𝑟𝑒𝑖 𝑖𝑓∀𝑠𝑐𝑜𝑟𝑒 > 0; −1, 𝑜𝑡𝑕𝑒𝑟𝑤𝑖𝑠𝑒;

where wi is the weight for the ith VM. We strictly refuse the configurations that lead to user SLA violation by assigning a reward of -1 if any VM’s score is negative. In the case of soft SLA thresholds, the reward function can be revised correspondingly to tolerate transient SLA violations. The state space. In the VM configuration task, the State space naturally matches the VM configurations. In the driver domain, VM configurations are fully observable. States defined on VM configurations are deterministic in that Pa(s, s_) = 1, which simplifies the RL problem. Wedefine the RL state as the global resource allocations: (mem1, time1, vcpu1, · · · ,memn, timen, vcpun). where memi, timei and vcpui are the ith VM’s memorysize, scheduler credit and virtual CPU number, respectively. 2.1.2.1 Solutions to the RL-workload metrics: The solution to a RL task is an optimal policy that maximizes the cumulative rewards at each state. It is equivalent to finding an estimation of Q(s, a) which approximates its actual value. The experience-based solution is based on the theory that the average of the sample Q(s, a) values collected approximates the actual value of Q(s, a) given sufficiently large number of samples. A sample is in the form of (st, at, rt+1). The basic RL algorithms in experience based solution are called temporal-difference (TD) methods, which update Q(s, a) at each time a sample is collected: Q(st, at) = Q(st, at)+α*[rt+1+γ*Q(st+1, at+1)−Q(st,at)],

𝑆𝐿𝐴

The reference throughput (ref thrpt) values are the maximum achievable application performance

where α is the learning rate and γ is the discount factor. The Q values are usually stored in a look-up table and updated by writing the new values to the

Volume o1 issues in engineering and advanced technology on Mar 2016 corresponding entries in the table. In the VM configuration task, the RL-based agent issues reconfiguration actions following a _-greedy policy. With a small probability _, the agent picks a random action. 2.1.3Self-Adaptive and Self Configured Optimal Resource Provisioning System (SASCORPS): VM capacity management relies on precise operations that set resources to desired values assuming the observation of the instant reconfiguration effect. However, in fine-grained cloud management, such as in [21], [23],Within the management interval the effect of a reconfiguration cannot be correctly perceived. The work in [23] showed up to 10 minutes delayed time before a memory reconfiguration stabilizes. Similar phenomenon was also observed in CPU. VM capacity management task defined as an autonomous learning process in an interactive environment. The framework is general inthe sense that various learning algorithms can be incorporated. Although the efficacy or the efficiency of the capacity management may be compromised, the complexity of the management task does not grow exponentially with the number of VMs or the number of resources. After a VM submits its SLA profile to App-agent and registers with Host-agent and Decision-maker. (SASCORPS)considers the VM capacity to be multidimensional, including CPU, memory and I/O bandwidth. This is one of the earliest works that consider these three types of resources together. A VM’s capacity can be changed by altering the VCPU number, memory size and I/O bandwidth. The management operation to one VM is defined as the combination of three meta operations on each resource: increase, decrease and nop. 2.1.3.1 VM Running Status VM running status has a direct impact on management decisions. A running status should provide insights into the resource usage of the VM, from which constrained or over-provisioned resource can be inferred. We define the VM running status as a vector of four tuples. (ucpu; uio; umem; uswap); Where ucpu, uio, umem, uswap denote the utilization of CPU, I/O, memory and disk swap, respectively. As discussed above, memory utilization cannot be trivially determined. 2.1.3.2 Feedback Signal The feedback signal ought to explicitly punish resource allocations that lead to degraded

application performance, and meanwhile encouraging a free-up of unused capacity. It also acts as an arbiter when resource are contented. We define a real-valued reward as the feedback. Whenever there is a conflict in the aggregated resource demand, e.g. the available memory becomes less than the total requested memory, (SASCORPS) set the reward to -1 (penalty) for the VMs that require an increase in the resource and a reward of 0 (neural) to other VMs. In this way, some of the conflicted VMs may back-off leading to contention relaxation. Note that, although conflicted VMs may give up previous requests, Decision-maker will suggest a second best plan, which may be the best solution to the resource contention. Reward =

𝑦𝑖𝑒𝑙𝑑 𝑐𝑜𝑠𝑡

,

Where yield = Y (x1, x2,..., xm)= 1, 𝑦 𝑥𝑖 = 𝑒

𝑚

,

𝑖𝑓𝑥𝑖𝑠𝑎𝑡𝑖𝑠𝑓𝑖𝑒𝑠𝑖𝑡𝑠𝑆𝐿𝐴;

|𝑥𝑖 −𝑥𝑖 |) ( 𝑥𝑖 −𝑝∗ 𝑚

𝑚 𝑖=1 𝑦 (𝑥𝑖 )

− 1,

𝑜𝑡𝑕𝑒𝑟𝑤𝑖𝑠𝑒,

𝑦 (𝑥𝑖 )𝑖

And cost =1 + 𝑖=1 . Note that the metric yield 𝑚 is a summarized gain over m performance metrics x1; x2; _ _ _ ; xm. The utility function y(xi) decays when metric xi violates its performance objective x 0 i in SLA. Cost is calculated as the summarized utility based on n utilization status u1; u2; _ _ _ ; un. Both the utility functions decay under the control of the decay factors of p and k, respectively. We consider throughput and response time as the performance metrics and ucpu, uio, umem, uswap as the utilization metrics. The reward punishes any capacity setting that violates the SLA and gives incentives to high resource efficiency. 2.1.3.3 Self –adaptive Learning Engine At the heart of (SASCORPS)is a self-adaptive learning agent using CMAC based queuing table responsible for each VM’s capacity management. Reinforcement learning is concerned with how an agent ought to take actions in a dynamic environment so as tomaximize a long term reward [27]. It fits naturally within (SASCORPS) feedback driven, interactive framework. RL offers opportunities for highly autonomous and adaptive capacity management in cloud dynamics. It assumes no priori knowledge about the VM’s running environment. It is able to capture the delayed effect of reconfigurations to a large extent. A RL problem is usually modelled as a Markov Decision Process (MDP). Formally, for a set of

Volume o1 issues in engineering and advanced technology on Mar 2016 environment states S and a set of actions A, the MDP is defined by the transition probability Pa(s; s0) = Pr(st+1 = s0jst = s; at = a) and an immediate reward function R = E[rt+1jst =s; at = a; st+1 = s0]. Q(s, a) = E {

k ∞ 𝑘=0 𝛾 rt+k+1|st=

s, at = a},

where 0 _ < 1 is a discount factor helping Q(s; a)’s convergence. The optimal policy is as simple as always select the action a that maximizes the value function Q(s; a) at state s. Algorithm 2Update the CMAC-based Q value function 1: Input st, at, st+1, rt; 2: Initialize δ = 0; 3: I[at][0] = get_index(st); 𝑗 ≤𝑛𝑢𝑚 _𝑡𝑖𝑙𝑖𝑛𝑔𝑠 4: Q(st; at) = 𝑗 =1 Q [I[at][j]]; 5: at+1 = get_next_action(st+1); 6: I[at+1][0] = get_index(st+1); 𝑗 ≤𝑛𝑢𝑚 _𝑡𝑖𝑙𝑖𝑛𝑔𝑠 7: Q(st+1; at+1) = 𝑗 =1 Q[I[at + 1][j]]; 8: δ = rt-Q(st, at + γ *Q(st+1, at+1)); 9: for i = 0; i < num tilings; i + +do 10:/*If SLA violated, enable fast adaptation*/ 11: if rt < 0 then 12:θ[I[at][i]]+ = (1:0/num _ tilings) *δ; 13: else 14:θ[I[at][i]]+ = (α/num _tilings) * δ; 15: end if 16: end for

One advantage of CMAC is its efficiency in handling limited data. Similar VM states will generate CMAC indexes with a large overlap. Thus, updates to one state can generalize to the others, leading to accelerated RL learning process. One update of the CMAC-based Q table only needs 6:5 milliseconds in our testbed, in comparison with the 50-second update time in a multi-layer neural network [23]. Once a VM finishes an iteration, it submits the four-tuple (st; at; st+1; rt) to Decisionmaker. Then the corresponding RL agent updates the VM’s Q table using Algorithm 2. In thealgorithm, we further enhanced the CMACbased Q table with fast adaptation when SLA violated. We set the learning rate _ to 1 whenever receives a negative penalty. This ensures that “bad” news travels faster than good news allowing the learning agent quickly response to the action.

3.

RELATED WORK:

Zhu and Agrawal [15] propose a dynamic mechanism for VM provisioning based on control

theory considering user budget. However, such an approach considers reconfiguration of available virtual instances (increase or decrease their capacity) and not increasing/decreasing number of instances for a customer, conversely to our approach that applies the latter approach for VM provisioning. Bi et al. [16] propose a model for provisioning multitier applications in Cloud data centers based on queueing networks. However, such a model does not perform recalculation of number of required VMs based on expected load and monitored performance as does our approach. Cloud computing allows cost-efficient server consolidation to increase system utilization and reduce cost. Resource management of virtualized servers is an important and challenging task, especially when dealing with fluctuating workloads and performance interference. Traditional control theory and machine learning have been applied with success to the resource allocation in physical servers; see [1], [18], [25], [15], [16], [24], [8] for examples. Recent work demonstrated the feasibility of these methods to automatic virtualized resource allocation. Early work [22], [26] focused on the tuning of the CPU resource only. Padala, et al. employed a proportional controller to allocate CPU shares to VM-based multi-tier applications [22]. This approach assumes nonwork- conserving CPU mode and no interference between co-hosted VMs, which can lead to resource underprovisioning. Recent work [14] enhanced traditional controltheory with Kalman filters for stability and adaptability. But the work remains under the assumption of CPU allocation. The authors in [26] applied domain knowledge guided regression analysis for CPU allocation in database servers. The method is hardly applicable to other applications in which domain knowledge is not available. Approaches that adjust resource allocation until performanceis satisfied. Chieu et al. describe a thresholdbased approach to scale web applications dynamically [30]. Industrial efforts have to consider both performance and versatility. Thus, most resource management modules deployed in commercial IaaS clouds exploit loadbased and threshold-based approaches to adjust resource allocations. These approaches include AutoScaling [20] and Rightscale [21]. This kind of approach only adds fixed numbers of VM instances in a matter of tens of minutes and, as such, may not provision resources in time when workload increases.

Volume o1 issues in engineering and advanced technology on Mar 2016 4.

CONCLUSIONS:

In this work, we present an optimal resource configuration scheme that allows self-adaptive virtual machine resource provisioning. Nevertheless, there are several limitations of this work. First, the management operations are discrete and are in a relatively coarse granularity. Second, the RL-based capacity management still suffers from initial performanceconsiderably. Future work can extend these provisioning models byCombining control theory with reinforcement learning. It offers opportunities for the control theory to provide fine grained operations and stable initial performance.

REFERENCES [1]

[2]

[3]

[4]

[5]

[6]

[7]

[8]

[9]

[10]

[11]

[12]

[13] [14]

Jinzhao Liu, Yaoxue Zhang, Yuezhi Zhou, Member, IEEE, Di Zhang, and Hao Liu”Aggressive Resource Provisioning for Ensuring QoS in Virtualized Environments” IEEE TRANSACTIONS ON CLOUD COMPUTING, VOL. 1, NO. 1, JANUARY 2014 Rodrigo N. Calheiros, Rajiv Ranjan, and Rajkumar Buyya. ”Virtual machine provisioning based on analytical performance and QoS in cloud computing environments”, ICPP 2011, pp. 295-304. IEEE, 2011. Zhang, Qi, et al. ”Dynamic Heterogeneity-Aware Resource Provisioning in the Cloud.” IEEE transactions on cloud computing. vol. 2, No. 1. pp. 14-28. IEEE, 2014. Michael Armbrust, et al. ”A view of cloud computing”, Communications of the ACM, vol. 53, no. 4, pp. 50-58, ACM, 2010. Hao Liu, et al. ”Receiving buffer adaptation for high-speed data transfer”, IEEE transactions on computers. Vol. 62, pp. 2278-2291, IEEE, 2013. Hiep Nguyen, et al. ”AGILE: Elastic Distributed Resource Scaling for Infrastructure-as-a-Service”, In Proc. of Autonomic Computing,International Conference on, 2013. Nikolas R. Herbst, Samuel Kounev, Ralf Reussner. ”Elasticity in Cloud Computing: What It Is, and What It Is Not”, ICAC, 2013. Peter Bodik, et al. ”Characterizing, modeling, and generating workload spikes for stateful services”, In Proceedings of the 1st ACMsymposium on Cloud computing, pp. 241-252. ACM, 2010. Emiliano Casalicchio, et al. ”Autonomic Resource Provisioning in Cloud Systems with Availability Goals”, Proceedings of the InternationalConference on Cloud and Autonomic Computing, IEEE, 2013. S. Ranjan and E. Knightly, ”High-Performance Resource Allocation and Request Redirection Algorithms of Web Clusters”, TPDS, vol. 19, no. 9, pp. 1186-1200, Sep. 2008. Jichuan Chang, et al. ”Workload diversity and dynamics in big data analytics: implications to system designers”, In Proceedings ofthe 2nd Workshop on Architectures and Systems for Big Data, pp. 21-26. ACM, 2012.James Hamilton, ”The Cost of Latency”, http://perspectives. mvdirona.com/2009/10/31/TheCostOfLatency.aspx. Apache Foundation, ”The Apache Cassandra Project”, http:// cassandra.apache.org/ ”Chinese holiday bookings bring down hosting site”,http://www.zdnet.com/chinese-holiday-bookingsbring-downhosting- site-7000010063.

[15] Nedeljko Vasi, Dejan Novakovi, Svetozar Miuin, Dejan Kosti, and Ricardo Bianchini. ”Dejavu: accelerating resource allocation in virtualized environments”, ACM SIGARCH Computer ArchitectureNews, vol. 40, no. 1, pp. 423-436. ACM, 2012. [16] Kai Shen, Ming Zhong, Sandhya Dwarkadas, Chuanpeng Li, Christopher Stewart, and Xiao Zhang. ”Hardware counter drivenon-the-fly request signatures”, ACM SIGOPS Operating Systems Review, vol. 42, no. 2, pp. 189-200. ACM, 2008. [17] RUBiS, http://rubis.ow2.org/. [18] Leslie Pack Kaelbling, et al. ”Reinforcement Learning: A Survey”, Journal of Artificial Intelligence Research, vol. 4, pp. 237-285, 1996.Shun-Feng Su, Tao, T., Ta-Hsiung Hung. ”Credit assigned CMACand its application to online learning robust controllers”, Systems,Man, and Cybernetics, Part B: Cybernetics, IEEE Transactions on,vol.33, pp. 202-213, 2003 [19] Brian F. Cooper, et al. ”Benchmarking cloud serving systems withYCSB”. In Proceedings of the 1st ACM symposium on Cloud computing,pp. 143-154. ACM, 2010. [20] Auto Scaling, http://aws.amazon.com/autoscaling/.