authors: mianyu wang, nagarajan kandasamy, allon guez, and moshe kam proceedings of the 3 rd...
TRANSCRIPT
ADAPTIVE PERFORMANCE CONTROL OF COMPUTING SYSTEMS
VIA DISTRIBUTED COOPERATIVE CONTROL:
APPLICATION TO POWER MANAGEMENT IN COMPUTING
CLUSTERSAuthors: Mianyu Wang, Nagarajan Kandasamy, Allon Guez, and Moshe Kam Proceedings of the 3rd International Conference on Autonomic Computing,
ICAC 2006, Dublin, Ireland
Presenter: Ramya Pradhan,
Fall 2012, UCF.
Outline
Research problem Proposed solution Evaluation of proposed solution Strengths Limitations Proposed extensions
Research ProblemServer cluster Clients
Power Consumption
How to balance power consumption with time-varying workload and QoS?
Proposed solution
Fully decentralized and cooperative control framework using optimal control theorybalance cluster operating frequency and
average response timescalable due to problem decompositionfault-tolerant due to cooperative controlno intra-cluster communication
Proposed solution using optimal control
Optimal control“uses predictive approach that generates sequence
of control inputs over a specified lookahead horizon while estimating changes in operating conditions.”
System ModelSystem state: queue sizeConstrained control input: operating frequencyOutput: average response time
Distributed control framework
Server cluster Global request buffer ClientsDynamic
Controllers
Evaluation
System settingse-commerce
○ Virtual store consisting of 10000 objects○ response time uniformly chosen between
(4,11) msrequest distribution
○ popularity○ temporal locality
cluster of four servers
EvaluationAdaptive power consumption
EvaluationAdaptive power consumption during processors’ failure
Strengths
Development of a communication-less framework for distributed optimization
Implementation of the framework of power consumption and guarantee QoS
Usage of distributed frameworkautonomous controllersno single point of failurecapable of self-* properties
Limitations Main concept: decomposing power
management into optimal control problems for each server, based on the assumption that resource provisioning and allocation can also be decomposed into such problems; this may not always be possible.
Adding new servers adds to the overhead in predicting its behavior by all other servers. Results for adding servers is not presented.
Possible extensions
Study the system under dynamic adding and removing of servers
Experiment with perturbations when servers are optimally performingremove servers that almost always
guanrantee QoS and see how other servers respond
add more servers to observe how estimating the other servers’ behavior affects guarantee of QoS
Thank You!