1 intelligent management of virtualized resources for database systems in cloud environment...
TRANSCRIPT
1
Intelligent Management of Virtualized Resources for
Database Systems in Cloud Environment
Pengcheng Xiong (Georgia Tech); Yun Chi (NEC Labs America); Shenghuo Zhu (NEC Labs America); Hyun
Moon (NEC Labs America); Calton Pu (Georgia Tech); Hakan Hacigumus (NEC Labs America)
Overview
Motivation Background System modeling Resource Allocation Conclusions
2
Overview
Motivation Background System modeling Resource Allocation Conclusions
3
Hosting database management systems
Common practice: One-to-one mapping of DBMSs to nodes
Problem: Poor utilization
(1)Both nodes are utilized under 10% most of the time. (2) The maximum CPU usage is much higher than the average CPU usage.
Pradeep Padala et al, Adaptive control of virtualized resources in utility computing environments(EuroSys '07)
Solution: consolidation based on virtualization
6
Solution: consolidation based on virtualization
3 CPUs
2 CPUs
4 is enough!
Problem: Provisioning under variable workload
8
Solution: Adaptive provisioning under variable
workload
9
GoalsGood performance
Maintain SLAService differentiation
Good resource utilization
10
Overview
Motivation Background System modeling Resource Allocation Conclusions
11
Background
1212
Virtualized Server I Virtualized Server II
VM 1
VM 2
VM 3
VM 4
MySQLMaster
MySQLSlave
MySQLMaster
MySQLSlave
Gold Client
Silver Client
Gold Client
Silver Client
13
What are we controlling ?
Inner level controller
GoalsSLA penalty cost =>Good
performanceInfrastructure cost=>Good
utilizationAction cost+)------------------------------------Total costMinimized? NO
Set CPU/Memory
shares
Outer level controller
Set number of
replicas
14
Related work Existing research
Resource allocation & scheduling Service differentiation
Our contribution: Non-linear model of DBMS behavior Two level controller
High resource utilization Low query response time, less SLA penalty cost,
Service differentiation Action cost
15
System overview
Overview
Motivation Background System modeling Resource Allocation Conclusions
16
System modeling How is the average SLA penalty
cost correlated with the system configuration? Statistic analysis, draw marginal
distribution How can we accurately predict
the average SLA penalty cost? Machine learning techniques,
linear and non-linear models17
System modeling: statistical analysis
18
Looks like straight line Non-straight line
System modeling: machine learning
19
Overview
Motivation Background System modeling Resource Allocation Conclusions
20
Resource allocation: two level controller
21
Find the next direction which can minimize the total costPerformance cost (related to SLA)Infrastructure cost (related to replicas)Action cost
22
Evaluation Experiment Environment
MySQL v5.0, MySQL replication Xen hypervisor
Workload Generator Dynamic arrival rate follows Poisson
distribution TPC-W Ordering (the browsing requests
and the ordering requests are 50%, respectively.)
TPC-W 100 EBs, 10K items. The whole database size is about 280MB.
Baseline
23
The total SLA penalty cost is 2802
Inner Level Controller
24
The total SLA penalty cost is 2363
Inner Level Controller Total SLA penalty cost under different
number of replicas
25
Outer level controller without action cost
26
<-EC2, Small instance, $0.085 per hour
Action cost modeling
27
Amortization factor Factor=0:the action cost will
be distributed along the infinity intervals. (optimistic)
Factor=1:the action cost can be compensated in the next interval. (pessimistic)
Factor between (0,1):the action cost can be compensated in several intervals.
28
Outer level controller with action cost
29
30
Conclusion Virtual resource management for
database management systems in Cloud computing Good performance
Maintain SLA Service differentiation
Good utilization SmartSLA
Non-linear model of DBMS behavior Two level controller which takes in to
consideration of SLA penalty cost, infrastructure cost and also action cost.
31
Thank you!
ELBA project page: http://www.cc.gatech.edu/systems/projects/Elba/index.html
E-mail: [email protected]