Automated Control in Cloud Computing:
Challenges and Opportunities
Harold C. Lim, Shivnath Babu, Jeffrey S. Chase, and Sujay S. ParekhACM’s First Workshop on
Automated Control for Datacenters and Clouds, 2009, Barcelona, Spain.
Presenter: Ramya Pradhan, Fall 2012, UCF.
Outline of the presentationResearch problemProposed solutionEvaluation of the proposed solutionStrengthsLimitationsPotential extensions
Research Problem
IaaS provider
Guest using IaaS
Guest’s clients
How to adaptively provision resources?
ChallengesDecoupling control
Cloud controller arbitrate resource requests, select guest VM placements
Application controller determine physical resources needed and communicate
to cloud controller
Control granularityCoarse sensor and actuator information.Noisy sensor measurement
CPU utilization as percentage of VM usage work-conserving scheduler gives noisy measurement
Proposed solutionA feedback driven application control implemented
at the guest’s end.Guest application controllers or slice controllers.
IaaS provider provides sensors and actuators to enable control policies.
Slice controllers use APIs to collect coarse-grained information from sensors and actuators.
Solution: A control technique, proportional thresholding, for coarse-grained actuators with a wide range of actuator values.
Proportional thresholdingIf incoming accumulated sensor value > high threshold, - then request resources- set high threshold to accumulated sensor valuehigh
threshold
low threshold
If incoming accumulated sensor value < low threshold, - then release resources- set low threshold to accumulated sensor value
Why proportional thresholding?Parameters to tune: CPU entitlement and
utilizationTuned using: an integral control
control effort is proportional to the integral of the error
well-suited for coarse-grained actuators actuators have a dynamic target range steady state error is zero
Evaluation of proportional thresholding
Horizontally scalable web service Automat (control interface) Open Resource Control Architecture (underlying architecture
and resource leasing mechanism) Hyperic HQ (gathers CPU utilization)
Sensor measurement average CPU utilization on all leased VMs experiments start with one VM
Additional VMs are obtained using proportional thresholding static thresholding integral control
Evaluation of proportional thresholdingSynthetic workload
time 0: 1000 threads, time 10: 1650 threads, time 40: 1000 threads
Proportional thresholding vs. integral control
Evaluation of proportional thresholdingSynthetic workload
time 0: 1000 threads, 15: 1650 threads, 30: 3200 threads, 45: 2450 threads
Proportional thresholding vs. static thresholding
StrengthsUtilizes accumulated actuator error to better
adapt to dynamic resource provisioning.Suitable for coarse-grained sensor information
provided by cloud providers.Shows self-constraint capability. Performs better resource allocation than integral
control and control using static thresholding.
LimitationsA key parameter, integral gain, in the equation for
integral control is empirically determined.May become application specific
Limited to 3 VMs.Discussion only on horizontal clusters.
Possible ExtensionsExtend to include more VMs.Extend to include vertical clusters.Analyze application of proportional thresholding to
at least one target system that needs complex models for integral gain. shows feasibility of the proposed method
Thank you!