a user experience-based cloud service redeployment mechanism kang yu
TRANSCRIPT
A User Experience-based Cloud Service
Redeployment MechanismKANG Yu
Introduction
Overview of Cloud-based Services
Redeploying Service Instances
Experiment
Obtaining User Experience
Conclusion and Future Work
Introduction
• In the emerging cloud computing systems, auto scaling and elastic load balance are keys to host the cloud services.– Auto scaling enables a dynamic allocation of computing
resources to a particular application. In other words, the number of service instances can be dynamically adapted to the request load.
– Elastic load balance distributes and balances the incoming application traffic (i.e., the user requests) among the service instances.
Introduction
• Typical approach of auto scaling and load balance (Amazon EC2)
Introduction
• Unfortunately, current auto scaling and elastic load balance techniques are generally not optimized for achieving best service performance.– Typical auto scaling approaches cannot start or terminate a
service instance at the data center selected according to the distributions of the end users.
– Elastic load balance generally redirects user requests to the service instances merely based on loads of the instances. It does not take the user specifics (e.g., user location) into considerations.
Introduction
• Our contribution:– We model the features of user experience in cloud
service.– We propose a new user experience-based service
hosting mechanism which employs a service redeployment method.
Introduction
• Our method has two advantages:1) It improves current auto scaling techniques by
launching the best set of service instances according to the distributions of end users.
2) It extends elastic load balance. Instead of directing user request to the lightest load service instance, it directs user request to a nearby one.
Introduction
Overview of Cloud-based Services
Redeploying Service Instances
Experiment
Obtaining User Experience
Conclusion and Future Work
Framework of Cloud-Based Services
• A cloud contains several data centers. Physical machines are virtualized as instances in the data center. Service providers would deploy service running on these instances. An end user normally connects to the cloud to get data and run applications
/services. User requests are directed to the service instances.
Framework of Cloud-Based Services
• The connection information especially Round Trip Time (RTT) between a user and an instance can be kept by the cloud provider.
• User experience contains three elements:1. Internet delay between a user and a cloud data
center (This is the most significant part)2. Delay inside the data center3. Time to process the service request
Challenges of Hosting the Cloud Services
• Difficult of foreseeing user experience before actually running the service.
• Internet delay between users and every cloud data center can either be measured or be predicted. ---Different from existing computing infrastructures.
Introduction
Overview of Cloud-based Services
Redeploying Service Instances
Experiment
Obtaining User Experience
Conclusion and Future Work
Measure the Internet Delay
• A request is responded by an instance inside the cloud thus the cloud provider is able to record the RTT from the user to the instance.
Predict the Internet Delay
• A user may not be able to visit many instances deployed in every data center.
• Find similar users and predict the connection.
Obtaining User Experience
Introduction
Overview of Cloud-based Services
Redeploying Service Instances
Experiment
Obtaining User Experience
Conclusion and Future Work
Minimize Average Cost
Minimize Average Cost
Minimize Average Cost
• k-median problem• Algorithms:
1. Brute Force2. Greedy Algorithm3. Local Search Algorithm (3 + ε approximation)4. Random Algorithm
Maximize Close User Amount
• Part of the users may be extremely far away from most of the data centers. They tend to force some service instances deployed in the data center close to them.
• We should also control number of users connected to a single server instance.
• We believe it is acceptable if some responses take a short time less than a threshold T.
Maximize Close User Amount
Maximize Close User Amount
Maximize Close User Amount
• If we view the red nodes as sets– {1,2,3,5}; {1,2,3}; {1,3,4}; {4,5}
• Max k-cover problem• Algorithms:
1.Greedy Algorithm (1-1/e approximation)2.Local Search Algorithm
Introduction
Overview of Cloud-based Services
Redeploying Service Instances
Experiment
Obtaining User Experience
Conclusion and Future Work
Dataset Description
• Deploy our WSEvaluator to 303 distributed computers of PlanetLab invoke to 4302 the Internet services
• A 303 * 4302 matrix containing response-time values
Necessity of Redeployment
Weakness of Auto Scaling
Comparing Algorithms for k-Median
Comparing Algorithms for k-Median
• Theoretical time complexity– Brute Force: – Greedy:– Local Search:
)( NMO k
)( NMkO
)( NMkO tt
Redeployment Algorithms for Max k-Cover
• 20 instances are selected to provide service for 4000 users.
• Expect 200 per server.
Redeployment Algorithms for Max k-Cover
• compare the average cost: max k-cover v.s. k-median
Introduction
Overview of Cloud-based Services
Redeploying Service Instances
Experiment
Obtaining User Experience
Conclusion and Future Work
Conclusion and Future Work
• Our work consists two parts– We propose a framework to address the new features
of cloud.– We formulate the redeployment of service instances
as k-median and max k-cover problems.
• Future Work– Formulate the network capability of service instance
carefully with the amount of users.– Figure out potential users and optimize initial service
instances deployment.
