![• Many internet applications can benefit from
an auto scaling property where their resource
usage can be scaled up and down
automatically by the cloud service provider.
• “Pay as you go” model.
• Fault tolerance.
• It consists of a load balancing switch,
A set of application servers, and a set of
backend storage servers.
INTRODUCTION[1]](https://image.slidesharecdn.com/autoaticscalingofinternetapplicationsforcloud-150303030921-conversion-gate01/85/Autoatic-scaling-of-inter-net-applications-for-cloud-4-320.jpg)
Autoatic scaling of inter net applications for cloud
- 1. AUTOATIC SCALING OF INTER NET APPLICATIONS FOR CLOUD COMPUTING SERVICES P.CHANDU 118T1A0503 ponnagantichandu@gmail.com
- 3. INTRODUCTION • Benefits of cloud computing service is the resource elasticity • A business customer can scale up and down its resource usage as needed without upfront capital investment or long term commitment. The amazon EC2 service, for example, allows users to buy as many virtual machine (VM) instances as they want and operate them much like physical hardware.
- 4. • Many internet applications can benefit from an auto scaling property where their resource usage can be scaled up and down automatically by the cloud service provider. • “Pay as you go” model. • Fault tolerance. • It consists of a load balancing switch, A set of application servers, and a set of backend storage servers. INTRODUCTION[1]
- 5. • The applications store their state information in the backend storage servers. It is important that the applications themselves are stateless so that they can be replicated safely. • The storage servers may also become overloaded, but the focus of this work is on the application tier. Example: the Google app engine service. Fig .two tiered architecture for internet applications
- 6. SYSTEM ARCHITECTURE • We encapsulate each application instance inside a virtual machine the use virtual machine to provide isolation among untrusted users. • Each server run the XEN hypervisor. • The system described as some decisions: Application placement Load distribution. Effectiveness techniques in ACFIAFCS: • Ballooning technique • Mirroring technology
- 7. DEFINITIONPROBLEM • Multi dimensional resource demands • Vector bin packing. Automatic scaling problem we want to slove is defined as follows: Our goals are to maximize the demand satisfaction ratio, minimize the placement change frequency and minimize energy consumption. Our optimization objectives can be expressed as follows
- 8. • To simplify the problem described above, we make the assumption that the servers are homogeneous with uniform capacity. • Then the auto scaling problem is similar to the Class Constrained Bin Packing (CCBP) problem. • The only difference is that the CCBP problem does not have the "Minimize the placement change frequency "goal. • Therefore, in order to solve our problem, we modified the CCBP model to support the "Minimize the placement change frequency "goal
- 9. CCBPFORAUTOMATICSCALING OFRESOURCES • Bin packing problem: a series of items of different sizes need to be packed into a minimum number of bins. • The goal is to pack the items in to a minimum number of bins. • We can model our resource allocation as the Class Constrained Bin Packing (CCBP) problem where each server is a bin and each class represents an application. Items from a specific class represent the resource demands of the corresponding application. The class constraint reflects the practical limit on the number of applications a server can run simultaneously. For J2EE applications • Load unit • Most online CCBP algorithms do not support item departure. • minimize application placement changes by adjusting the existing load distribution.
- 10. • Fig. 3 shows an example where we have two servers and three applications. Each server can run at most two applications simultaneously . We set the load unit to 20% Each server can thus satisfy 5 units of load (i.e., ). Let the resource demands of the applications be 60%, 20%, and 120% of the server capacity, respectively. • This translates into 3, 1, and 6 items to be packed for their corresponding classes. • The CCBP problem is NP-hard.
- 11. Detail of our algorithm: Our algorithm belongs to the family of color set algorithms Application load increase: The load increase of an application is modeled as the arrival of items with the corresponding color . Application load decrease: The algorithm has the freedom to choose which item of that color to remove. Application joins and leaves: When the last item of a specific color leaves, that color can be removed from its color set.
- 12. If there are, we allocate the new colors to the unfilled sets first using the following add_new_colorsprocedure.in that use a three procedures that are a Procedure add_new_colors Procedure consolidate_unfilled_sets Procedure fill(s1 ,s2 )
- 13. ANALYSIS OF THE APPROXIMATION RATIO The quality of a polynomial time algorithm A is measured by its approximation ratio R(A) to the optimal algorithm OPT: Where is the list of the input sequence and and are the number of bins used under the algorithm and the optimal algorithm, respectively the problem is NP-hard, no polynomial time optimal algorithm exists. Hence, the OPT algorithm here serves as an unrealistic upper bound.
- 15. EXPERIMENTS • Load shifting: We first evaluate the effectiveness of load shifting in our algorithm to avoid application placement change. • Auto scaling :we reduce the request rate of the application gradually to emulate that the flash crowd is over. The algorithm scales down the server resources accordingly to conserve energy. In experiment demonstrate the effectiveness of our system in real experiments
- 16. SIMULATIONS • This section evaluates the performance of our application scheduling algorithm in large scale simulation. • We assume that all servers are homogeneous. We define the“demand ratio”as the ratio between the aggregate demand of all applications and the aggregate capacity of all servers. • The simulator invokes our scheduling algorithm according to the prepared input and computes performance metrics such as the number of apms in the system, the decision time, the application demand satisfaction ratio, and the number of placement changes.
- 17. Application demand ratio Scalability : We evaluate the scalability of the algorithm by increasing both the number of servers and the number of applications from 1000 to 10,000. Fig. Left shows how the decision time increases with the system size. The middle figure shows that the demand satisfaction ratio is independent of the system size .The right figure shows that the number of placement changes increases linearly with the system size.
- 18. Application number: Next we vary the ratio between the applications and the servers by increasing the number of applications from 200 to 2000. The number of servers is fixed at 1000. The results are shown in Fig The left figure shows that the decision time is longer (but remains under 0.5 second) when there are more applications and when the class constraint is larger. The middle figure shows that the satisfaction ratio is little affected by the number of applications. The right figure shows that when the demand is high, more applications lead to more placement changes .
- 19. CONCLUSION We presented the design and implementation of a system that can scale up and down the number of application instances automatically based on demand. We developed a color set algorithm to decide the application placement and the load distribution. Our system achieves high satisfaction ratio of application demand even when the load is very high. It saves energy by reducing the number of running instances when the load is low. There are several directions for future work. Some cloud service providers may provide multiple levels of services totheir customers. When the resources become tight, they may want to give their premium customers a higher demand satisfaction ratio than other customers. In the future, we plan to extend our system to support differentiated services but also consider fairness when allocating the resources across the applications.