a method for transparent admission control and request scheduling in e-commerce web sites s....

12
Transparent Transparent Admission Control Admission Control and Request and Request Scheduling in Scheduling in E-Commerce Web Sites E-Commerce Web Sites S. Elnikety, E. Nahum, J. S. Elnikety, E. Nahum, J. Tracey and W. Zwaenpoel Tracey and W. Zwaenpoel Presented By Tyrel Russell Presented By Tyrel Russell

Upload: alban-lee

Post on 05-Jan-2016

213 views

Category:

Documents


1 download

TRANSCRIPT

Page 1: A Method for Transparent Admission Control and Request Scheduling in E-Commerce Web Sites S. Elnikety, E. Nahum, J. Tracey and W. Zwaenpoel Presented By

A Method for Transparent A Method for Transparent Admission Control and Admission Control and Request Scheduling in Request Scheduling in E-Commerce Web SitesE-Commerce Web Sites

S. Elnikety, E. Nahum, J. Tracey and S. Elnikety, E. Nahum, J. Tracey and W. ZwaenpoelW. Zwaenpoel

Presented By Tyrel RussellPresented By Tyrel Russell

Page 2: A Method for Transparent Admission Control and Request Scheduling in E-Commerce Web Sites S. Elnikety, E. Nahum, J. Tracey and W. Zwaenpoel Presented By

The IdeaThe Idea

Reduce contention at the database serverReduce contention at the database server

No modification of existing systemsNo modification of existing systems

Increase response time and maintain peak Increase response time and maintain peak throughputthroughput

Page 3: A Method for Transparent Admission Control and Request Scheduling in E-Commerce Web Sites S. Elnikety, E. Nahum, J. Tracey and W. Zwaenpoel Presented By

OutlineOutline

IntroductionIntroduction

The Gatekeeper ProxyThe Gatekeeper Proxy

Experimental EnvironmentExperimental Environment

Conclusion and DiscussionConclusion and Discussion

Page 4: A Method for Transparent Admission Control and Request Scheduling in E-Commerce Web Sites S. Elnikety, E. Nahum, J. Tracey and W. Zwaenpoel Presented By

The ProblemsThe Problems

Two problems for e-commerce web sitesTwo problems for e-commerce web sites OverloadOverload ResponsivenessResponsiveness

Solution is admission controlSolution is admission control Measures performanceMeasures performance Enables overload protectionEnables overload protection Improves response timeImproves response time

Page 5: A Method for Transparent Admission Control and Request Scheduling in E-Commerce Web Sites S. Elnikety, E. Nahum, J. Tracey and W. Zwaenpoel Presented By

The ArchitectureThe Architecture

Page 6: A Method for Transparent Admission Control and Request Scheduling in E-Commerce Web Sites S. Elnikety, E. Nahum, J. Tracey and W. Zwaenpoel Presented By

The Gatekeeper ProxyThe Gatekeeper Proxy

Intercepts dynamic web content requestsIntercepts dynamic web content requests

Requests generated by Java servlets on the application serverRequests generated by Java servlets on the application server

Estimates expected service times for each typeEstimates expected service times for each type

Determines the capacity of the systemDetermines the capacity of the system

Schedules the requests to improve performanceSchedules the requests to improve performance

Page 7: A Method for Transparent Admission Control and Request Scheduling in E-Commerce Web Sites S. Elnikety, E. Nahum, J. Tracey and W. Zwaenpoel Presented By

Estimated LoadEstimated Load

Generates estimated load from expected Generates estimated load from expected service times for each servletservice times for each servlet

This is a measure based on recent This is a measure based on recent executions of requestsexecutions of requests

Much greater variation between servlets Much greater variation between servlets than within servletsthan within servlets

Service times vary based the loadService times vary based the load

Page 8: A Method for Transparent Admission Control and Request Scheduling in E-Commerce Web Sites S. Elnikety, E. Nahum, J. Tracey and W. Zwaenpoel Presented By

System CapacitySystem Capacity

Determined offlineDetermined offlineCapacity is defined as the maximum load level Capacity is defined as the maximum load level that produces the highest capacitythat produces the highest capacityFound by searching the space using candidate Found by searching the space using candidate values and experimentally verifying themvalues and experimentally verifying them

0

20

40

60

80

100

120

0 20 40 60 80 100 120 140 160

Page 9: A Method for Transparent Admission Control and Request Scheduling in E-Commerce Web Sites S. Elnikety, E. Nahum, J. Tracey and W. Zwaenpoel Presented By

Admission ControlAdmission Control

Maintains running estimate of system loadMaintains running estimate of system load

When request is made, the work estimate When request is made, the work estimate is checked against the capacityis checked against the capacity

If there is room, it is admitted otherwise If there is room, it is admitted otherwise the request is queuedthe request is queued

When jobs are finished, the queued When jobs are finished, the queued requests are admitted when the room requests are admitted when the room allowsallows

Page 10: A Method for Transparent Admission Control and Request Scheduling in E-Commerce Web Sites S. Elnikety, E. Nahum, J. Tracey and W. Zwaenpoel Presented By

Request SchedulingRequest Scheduling

To reduce average response time, they To reduce average response time, they use the shortest-job first policyuse the shortest-job first policy

Queues request based on their expected Queues request based on their expected processing timesprocessing times

To prevent starvation, they used an agin To prevent starvation, they used an agin mechanism which enforces an upper mechanism which enforces an upper bound on the length of time the request bound on the length of time the request can stay in the queuecan stay in the queue

Page 11: A Method for Transparent Admission Control and Request Scheduling in E-Commerce Web Sites S. Elnikety, E. Nahum, J. Tracey and W. Zwaenpoel Presented By

Experimental EnvironmentExperimental Environment

Locking is done by using concurrency Locking is done by using concurrency control at the application servercontrol at the application server

They don’t control the serving of images They don’t control the serving of images as they are stored at the web serveras they are stored at the web server

Evaluate throughput and response time as Evaluate throughput and response time as a function of loada function of load

Page 12: A Method for Transparent Admission Control and Request Scheduling in E-Commerce Web Sites S. Elnikety, E. Nahum, J. Tracey and W. Zwaenpoel Presented By

Discussion and ConclusionDiscussion and Conclusion

Is the concavity assumption of capacity reasonable?Is the concavity assumption of capacity reasonable?Is the sliding window average of execution time actually Is the sliding window average of execution time actually a good measure of expected execution time?a good measure of expected execution time?Is the assumption that servlets will have relatively Is the assumption that servlets will have relatively uniform loads justified?uniform loads justified?Is locking at the server side a good idea? For example, Is locking at the server side a good idea? For example, what happens if multiple application servers are used for what happens if multiple application servers are used for the same database?the same database?They use a finite-state Markov model is used to model They use a finite-state Markov model is used to model user client behaviour. How were the transition user client behaviour. How were the transition probabilities generated and is this realistic?probabilities generated and is this realistic?How does holding up requests at the proxy effect the How does holding up requests at the proxy effect the scheduling and optimization mechanisms of the scheduling and optimization mechanisms of the database server?database server?