1. Autonomic Scaling of Cloud Computing Resources
using BN-based Prediction Models
Dr. Abul Bashar,
Assistant Professor
College of Computer Engineering and Sciences
Prince Mohammad Bin Fahd University
Al-Khobar, KSA
IEEE CLOUDNET 2013: 2nd International Conference on Cloud Networking

2. Introduction & Motivation Related Work Proposed Approach
Outline
Introduction & Motivation
Related Work
Proposed Approach
Implementation Details
Results and Discussion
Future Work and Conclusion
CLOUDNET 2013, 13th November 2013

3. Motivation : Scalability of Cloud Computing
Cloud Computing’s popularity: Quality service provisioning
Benefits: Reduced CAPEX and OPEX, Pay-as-you-go IT service, “Unlimited” computing resources
Challenges: Dynamic Provisioning, Resource usage optimization, Ensuring QoS/SLA
Motivation: Develop autonomic resource scaling solution
Machine Learning: Autonomic, Scalable and Predictive solutions
Our contribution: Bayesian Networks-based Scalability Control
CLOUDNET 2013, 13th November 2013

4. Related Work and Research Objectives
Existing ML-based Datacenter Management Systems
ARIMA models for resource prediction of cloud applications
String Matching Algorithms for cloud resource forecasting
Discrete Time Markov Chains for long-term demand predictions
Time Delay Neural Networks for predicting future workloads
Bayesian Networks (BN) for detecting failures in a cloud datacenter
Observations
Numerous ML-based solutions exist for predictive resource scaling
BN has not been used for prediction of resource demands
Our proposed objectives
To study prominent ML-based Cloud Computing Management methodologies
To model and implement BN based Decision Support System
To assess the performance of BNSC for predictive/diagnostic reasoning and decision-making
CLOUDNET 2013, 13th November 2013

5. Bayesian Network Representation
BN is a probabilistic graphical model, a mapping of physical system variables into a visual and intuitive model
Directed Acylic Graph structure : using nodes and arcs
Encodes conditional independence relation among system random variables
Defined mathematically using joint probability distribution formulation
Inference feature : Repeated use of Baye’s rule to estimate unobserved nodes based on evidence of observed nodes
CLOUDNET 2013, 13th November 2013

6. Conceptual Framework : BN-based DSS
DMS (Datacenter Management System): Monitors and provides monitoring data
DSS (Decision Support System): Uses DMS data and builds predictive models
Structural Learning: PC & NPC algorithms
Parameter Learning: EM Algorithm
Validation procedure: k-fold cross validation
Inference & Prediction: repeated Baye’s rule / classifier function
CLOUDNET 2013, 13th November 2013

7. Experimental Setup Details
Characteristics of Workload Demands
Cloud Datacenter Simulation in OPNET
BN Nodes Definition
CLOUDNET 2013, 13th November 2013

8. Simulation Results : BN Model
BN model provides the structural relationships among the nodes
Cause and effect nodes : parent child relationships
Marginal probabilities of all the nodes (shown in the monitor windows)
Useful in scenarios when there are numerous variables
CLOUDNET 2013, 13th November 2013

9. Conditional Probability Tables
Simulation Results : CPTs
Conditional Probability Tables
EM Algorithm for learning parameters (CPTs)
Strength of relationships between the nodes
𝑷 𝑪𝑷𝑼_𝑼𝒔𝒂𝒈𝒆=𝑯𝒊𝒈𝒉 𝑾𝒐𝒓𝒌𝒍𝒐𝒂𝒅_𝑫𝒆𝒎𝒂𝒏𝒅=𝑳𝒐𝒘 =𝟎.𝟎𝟓𝟕
CLOUDNET 2013, 13th November 2013

10. Simulation Results : Influence Diagram
BNSC (Bayesian Networks-based Scalability Control)
Utility node named Reward is added along with an action node named Scalability_Control
Reward node values depend on the states of Response_Time node for making scaling decisions
Decisions of Scale_Up or Scale_Down are the actions of Scalability_Control node
Utility Table for Reward Node of BNSC
CLOUDNET 2013, 13th November 2013

11. Simulation Results : Predictive Reasoning
Sample decision to Scale_Up when Workload_Demand is High
The system is under the influence of heavy workload demand
BNSC rightly decides to scale up the resources with a reward of +88.9
CLOUDNET 2013, 13th November 2013

12. Simulation Results : Diagnostic Reasoning
Sample decision to find the probable cause of Low Response_Time
BNSC is now scaling down the resources with a reward of
Diagnoses the reason for low response time (Workload_Demand is Low with probability of 0.996)
CLOUDNET 2013, 13th November 2013

13. Summary & Conclusions
Offline modeling of BNSC is achievable and practically implementable
Scaling decisions were found to be coherent and plausible
BNSC solution demonstrated successful predictive and diagnostic decision making for scaling up/down of Cloud Computing resources
Novelty of BNSC solution is to provide autonomic decision making
Future work involves incorporating more performance metrics in the BNSC model for more realistic resource scaling decisions
Another aspect worth researching is to model multiple distributed datacenter performance behavior
To make BNSC a comprehensive online learning and decision support system
CLOUDNET 2013, 13th November 2013

14. Acknowledgement
The author would like to acknowledge the support of Prince Mohammad Bin Fahd University, KSA for performing this research work.
CLOUDNET 2013, 13th November 2013

15. THANK YOU
CLOUDNET 2013, 13th November 2013