Adaptive Control of Virtualized Resources in Utility Computing Environments

1
Adaptive Control of Virtualized Resources in
Utility Computing Environments
Pradeep Padala, Kang G. Shin University of
Michigan

• HP Labs Xiaoyun Zhu, Mustafa Uysal, Zhikui Wang,
  Sharad Singhal
• University of Waterloo Kenneth Salem

2
A typical scenario in data centers
Customer A
Customer B
Run auction site
Run news site
Shared Data Center
3
Hosting applications
Data Center
Common idiom One-to-one mapping of applications
to nodes
4
Problem Poor utilization
Wasted Resources
Ad-hoc resource allocation schemes waste resources
5
Solution Virtual data center
Web server
Linux
Virtualization (Xen, OpenVZ, VMware)
Consolidate
Improved utilization using consolidation
6
Problem Provisioning
Wasted Resources
Bursty Load
Bad response time
Peak
Average
Provisioning for dynamic workloads is hard!
Solution Adaptive controller
7
Goals

• Good utilization
• Good performance
• QoS differentiation

Average CPU utilization 80
Average response time 100ms
Gold vs. Silver customers 21 resources
8
Outline

• Motivation
• Background
• Modeling
• Design
• Evaluation
• Conclusion

9
How do we provision the customers ?
Customer A
Auction Client
News Client
Virtualized Server I
Virtualized Server II
Customer B
10
What are we controlling ?

• Goals
• Good performance
• Good utilization
• QoS differentiation

VM I
50
80
Goals met ? NO
50
20
VM II
Policy
Mechanism
CPU Usage ?
Controller
Xen scheduler
Set CPU shares
Virtualized Server
11
Related work

• Existing research
• Cluster management
• Load balancing
• Resource allocation scheduling
• QoS differentiation
• Our contribution Adaptive resource control
• Quantitative model of system behavior
• Fine-grained, adaptive control
• No wastage of resources
• High throughput, low response time
• QoS differentiation

12
How do we design an adaptive resource controller?
Understand system variables Input
Output
Design controller PI, PID, I controller
Stress the controller
Goals met ?
A control theoretic approach to systems
13
Outline

• Motivation
• Background
• Modeling
• Design
• Evaluation
• Conclusion

14
Modeling a virtual data center
VM utilization
VM Shares
Throughput
Response time
Workload
QoS differentiation
Virtualized Server I
Virtualized Server II
How to differentiate between two multi-tiered
systems ?
15
Modeling two multi-tiered systemsQoS metric
Linear
Non-Linear
Response time ratio is more controllable than
loss ratio
16
Outline

• Motivation
• Background
• Modeling
• Design
• Evaluation
• Conclusion

17
Utilization controller an example
Controller
Utilization goal 80
Using 20
New Utilization 20/25100 80
Utilization 20/40100 50

• Problems
• Utilization is variable
• Delays and errors in sensing setting
• Stability concerns

Solution Self-tuning integral controller
18
Utilization controller
Workload
Error in utilization e(k-1)
Self-tuning controller
System
CPU allocation u(k)
Utilization goal
Measured utilization u(k-1)
-

• Adjusts to varying demand
• Maintains goal utilization
• Knobs to control aggression (Kp)
• Proven stable Wang DSOM05

19
Let there be controllers
110 Cant fit (Saturation)
Want 40
Want 70
Virtualized Server I
Container consumptions
Problem All controllers independent
Solution Arbiter controller enforcing QoS
differentiation
Virtualized Server II
20
Final controller
Requested CPU shares
Final CPU shares
Arbiter Controller
Virtualized Server I
Container consumptions
Desired response time ratio
Virtualized Server II
21
Outline

• Motivation
• Background
• Modeling
• Design
• Evaluation
• Conclusion

22
Evaluation

• Multi-tiered systems
• 2 HP Proliant servers
• Apache MySQL
• Xen 3.0 with SEDF scheduler
• Clients
• RUBiS auction client
• 2 RUBiS clients 500 1000 threads
• Can we maintain 70 QoS ratio ?

23
Varying load - throughput
1000 threads
500 threads
24
Varying load - control
Saturation
Buffer to maintain good performance
Penalized to maintain QoS ratio
25
Varying load QoS ratio
Goal
Goal ratio of 70 maintained!
26
Conclusion

• Adaptive control of virtual data center
• Good application performance
• High throughput
• Low response time
• Good utilization
• Maintain goal CPU utilization
• QoS differentiation
• Maintain goal QoS ratio
• Project page http//kabru.eecs.umich.edu/twiki/bi
  n/view/Main/DynamicControl
• E-mail ppadala_at_eecs.umich.edu
• Questions ?

27
Backup and old slides
28
Enterprise data centers

• Large data centers
• 100s/1000s of nodes
• Shared infrastructure
• Run critical applications
• Should meet service levels
• Problems
• Power costs
• Management costs
• Poor utilization
• Unmet service levels

29
Solution Consolidate !
30
Hosting two multi-tiered systems
Auction Client
Customer A
News Client
Virtualized Server II
Virtualized Server I
Customer B
31
Varying load
1000 clients
500 clients
Workload I
Workload II
Time
32
(No Transcript)
33
(No Transcript)
34
Modeling results - throughput
Dom0 effect
Web share
Throughput
Saturation causes Real throughput lt Offered
throughput
35
Arbiter controller features

• Is an integral controller
• Decides final shares based on QoS differentiation
  goals
• Integral gain knobs for aggression
• Stable gain value based on model

36
Modeling a multi-tiered system
Workload
Web usage
Web share
DB usage
DB share
QoS metrics
Virtual Server

• Stress the system in various scenarios
• Observe all variables

37
Modeling results response time
Dom0 effect
Web share
Response time
38
Questions

• ?