sebd2015_presentationvitali
TRANSCRIPT
Managing Energy Efficiency and Quality of Service in Cloud Applications Using a Distributed Monitoring System
Monica Vitali
Dip. Elettronica Informazione e BioingegneriaPolitecnico di Milano, Italy
MOTIVATION
monitor and improve Energy Efficiency in a cloud environment from an application perspective, while respecting the constraints established in the Service Level Agreement (QoS).
ISSUEMonitoring data are too BIG (volume and velocity) to be collected and analyzed in a centralized way.
GOALproposing a distributed and efficient methodology for monitoring the application and analyze the data.
The Model
Goal to Goal relations
influence existing between indicators (KPIs and GPIs)
Actions to Goal relationsthe effect of an action over a goal (positive or negative)
Monica Vitali, Barbara Pernici, and Una-May O’Reilly. "Learning a goal-oriented model for energy efficient adaptive applications in data centers." Information Sciences (2015).
Enables “what if” analysis and indirect improvement.
The Model (Goal Layer)
Goal to goal (metrics) relations are represented by an automatically learned Bayesian Network (BN).
Learn relations (correlation)
Learn directions (MMHC)
Learn Parameters (MAPE)
The Model (Goal Layer)
If executed in a centralized node, computation time is exponential with the number of monitored indicators
Example of BNClusters of relations
The Model (Goal Layer)
If executed in a centralized node, computation time is exponential with the number of monitored indicators
Example of BNClusters of relations
The Model (Goal Layer)
If executed in a centralized node, computation time is exponential with the number of monitored indicators
Example of BNClusters of relations
The Approach
Strengths the approach is dynamic, flexible, automatic
Weaknesses the sample time of the collection of monitoring data is limited by the amount of data, time for the analysis grows exponentially with the number of monitored indicators
Solution employ a distributed monitoring system and implement a distributed algorithm for data analysis
Distributed Monitoring
Data are saved where they are collectedNetwork usage reduction, improvement in the effectiveness of data collection
Server StorageVirtual Machine
Application activity
Monitoring Agent
CLOUD
DATA CENTER 1 DATA CENTER 2 Indicators collected and computed at three layers:VM, Host, Application
Two steps:Local analysis in each host and global analysis in a centralized node
Distributed LearningThe BN can be learned with a distributed algorithm exploiting the clusters of variables
Apply the BN learning for each VM
Learn relations between and with the hosts
metrics
Learn relations between and with the
application metrics
Distributed LearningMonitored indicators:4 indicators for each VM,2 indicators for each server
Initial configuration:2 servers,2 VMs on each server
Step:Increase the number of VMs
RESULTS:RT reduction between 62(81)% and 92(96)%
From exponential to linear behavior → scalability improvement