Cloud computing has enabled a wide array of applications to be exposed as elastic cloud services. While the number of such services has rapidly increased, there is a lack of techniques for supporting cross-layered multi-level monitoring and analysis of elastic service behavior. In this paper we introduce novel concepts, namely elasticity space and elasticity pathway, for understanding elasticity of cloud services, and techniques for monitoring and evaluating them. We present MELA, a customizable framework that enables service providers and developers to analyze cross-layered, multi-level elasticity of cloud services, from the whole cloud service to service units, based on service structure dependencies. Besides support for real-time elasticity analysis of cloud service behavior, MELA provides several customizable features for extracting functions and patterns that characterize that behavior. To illustrate the usefulness of MELA, we conduct several experiments with a realistic data-as-a-service in an M2M cloud platform.
Prototype and Demos at http://tuwiendsg.github.io/MELA/
Paper DOI: http://dx.doi.org/10.1109/CloudCom.2013.18
MELA: Monitoring and Analyzing Elasticity of Cloud Services -- CloudCom 2013
1. MELA: Monitoring and Analyzing Elasticity of Cloud
Services
Daniel Moldovan,
Georgiana Copil, Hong-Linh Truong, Schahram Dustdar
Distributed Systems Group (http://dsg.tuwien.ac.at/)
Vienna University of Technology (http://www.tuwien.ac.at/)
Work partially supported by the European Commission in terms of the
CELAR FP7 project (http://www.celarcloud.eu/)
2. Motivation
Elastic Cloud Service
Data-as-a-Service for Machine to Machine platforms
Load balancer distributes incoming requests to Event Processing instances
Distributed Data Store: Controller and Nodes
Start with an initial lighter configuration
2
3. Motivation
Elastic Cloud Service
Data-as-a-Service for Machine to Machine platforms
Load balancer distributes incoming requests to Event Processing instances
Distributed Data Store: Controller and Nodes
Add service unit instance when load increases
2
4. Motivation
Elastic Cloud Service
Data-as-a-Service for Machine to Machine platforms
Load balancer distributes incoming requests to Event Processing instances
Distributed Data Store: Controller and Nodes
2
Remove service unit instance when load decreases
5. Motivation
Elastic Cloud Service
Data-as-a-Service for Machine to Machine platforms
Load balancer distributes incoming requests to Event Processing instances
Distributed Data Store: Controller and Nodes
2
Add service unit instance and data node instance
when load increases too much
6. Motivation
Insufficient Cloud Service Monitoring and Analysis Support
Service Level Monitoring
Response time
Number of clients
Other specific metrics
Controlling the service’s elasticity
User-Defined Requirements violation:
- Cost per client too high
Reasons:
- Too much logging? Monitoring chatter?
- Too expensive VMs? Which one can be downsized?
- Not enough clients? Why?
System Level Monitoring
Ganglia, Nagios, etc.
CPU usage
Memory usage
Network transfer
3
7. Approach and Challenges
Structure Monitoring Data
How to map system data to service level?
How to derive higher level information?
Monitoring Data
Service Structure
Impose service structure over collected monitoring data
4
15. Approach and Challenges
Evaluate Service’s Elasticity
How to characterize service elasticity?
How to derive service‘s behavior limits?
How to characterize and predict elasticity behavior?
6
16. Runtime Properties of Elastic Cloud Services
Background
Elastic process: cost, quality and resources elasticity
Extend concept to cloud services
Elasticity Space
Collection of monitoring snapshots
I.e. the space in which an elastic service moves
Elasticity Boundary
Elasticity Space boundaries in which service’s requirements are respected
Elasticity Pathway
Characterizes service evolution trough elasticity space
16
Elasticity Dimensions
17. Multi-Level Elasticity Space
Event Processing Topology
Service requirement
COST<= 0.0034$/client/h
2.5$ monthly subscription for each service client (sensor)
Elasticity Space “Clients/h” Dimension
Elasticity Space Snapshot
Elasticity Space “Response Time” Dimension
8
18. Multi-Level Elasticity Space
Event Processing Topology
Service requirement
COST<= 0.0034$/client/h
2.5$ monthly subscription for each service client (sensor)
Determined Elasticity Space Boundaries
Clients/h > 148
300ms ≤ ResponseTime ≤ 1100 ms
Elasticity Space “Clients/h” Dimension
Elasticity Space “Response Time” Dimension
8
19. Multi-Level Elasticity Pathway
Service requirement
COST<= 0.0034$/client/h
2.5$ monthly subscription for each
service client (sensor)
9
20. Multi-Level Elasticity Pathway
Service requirement
COST<= 0.0034$/client/h
2.5$ monthly subscription for each
service client (sensor)
Cloud Service Elasticity Pathway
9
21. Multi-Level Elasticity Pathway
Service requirement
COST<= 0.0034$/client/h
2.5$ monthly subscription for each
service client (sensor)
Cloud Service Elasticity Pathway
9
Event Processing service unit Elasticity Pathway
22. Conclusions
Concepts
Elasticity Space and Elasticity Boundary
Elasticity Pathway
Mechanisms
Constructing cross-layer monitoring snapshots
Determining elasticity space and boundary
Determining elasticity pathway
MELA
Customizable framework for monitoring and analyzing elasticity of
cloud services
MELA: Monitoring and Analyzing Elasticity of
Cloud Services
http://dsg.tuwien.ac.at/research/viecom/mela/
Distributed Systems Group(http://dsg.tuwien.ac.at/)
Vienna University of Technology (http://www.tuwien.ac.at/)
Work partially supported by the European Commission in terms of the
CELAR FP7 project (http://www.celarcloud.eu/)
10