1 economical and robust provisioning of n-tier cloud workloads: a multi-level control approach...
TRANSCRIPT
1
Economical and Robust Provisioning of N-Tier Cloud
Workloads: A Multi-level Control Approach
Pengcheng Xiong1, Zhikui Wang2, Simon Malkowski1, Qingyang Wang1, Deepal
Jayasinghe1, Calton Pu1
1Georgia Institute of Technology2HP Labs
Email: [email protected]
Overview
MotivationBackgroundResource partition
controllerApplication controllerConclusions
2
Overview
MotivationBackgroundResource partition
controllerApplication controllerConclusions
3
Applications in a typical Cloud environment
Different feedback controller design for a single/multi-tiered application (1)
5Zhu et al, ACC 2006
Different feedback controller design for a single/multi-tiered application (2)
6Wang et al, FeBID 2007
C3
C2
E1
MRT (Mean Response Time)
+
RTref
N (Transaction Mix)
Feedback Controller
Client Feed-Forward Controller
Umdl
U U
-
Utilization Controller 1
+
WWW Server
App. Server
DB server
RT
AutoParam
Utilization Controller 2
Utilization Controller 3
E2
E3
C1
U1 ref
U2 ref
U3 ref
TUC
TFB
TFF
Different controllability under different workload generator (1)
7Schroeder et al, NSDI 2006
Different controllability under different workload generator (2)
8Xiong et al, NOMS 2010
Goals Economical
– We want to meet the performance requirement for the N-tier web application with the minimum total resources.
Robust– We want to be robust to different
time-varying workload types, e.g., open, closed, semi-open.
9
Overview
MotivationBackgroundResource partition
controllerApplication controllerConclusions
10
Control Architecture
11
Test bed
Experiment Environment– Apache, Tomcat, Mysql– Xen hypervisor
Workload Generator– RUBiS “Browsing mix” workload that has 10
transaction types, e.g., Home, Browse, ViewItem. (just like eBay)
– Workload types (open, closed, semi-open)– Workload intensity
12
Overview
MotivationBackgroundResource partition
controllerApplication controllerConclusions
13
System modeling
14
Optimal resource partition
Solution 1(Shares)
Solution 2(Util.)
Our solution(Opt.)
Evaluation of resource partition controller
16
Overview
MotivationBackgroundResource partition
controllerApplication controllerConclusions
17
Application controller design
18
System model between the RTT and S– System identification method based on ARMA
model
Controller design– Root-locus method based on control theory
System identification
19
Controller design
20
ARX01 model
Proportional-integral (PI) controller
The closed model transfer function
Performance controller(setting=35ms)
21
Util has MORE fluctuation than Opt.
Performance controller(setting=200ms)
22
Conclusions
We propose economical and robust provisioning for Cloud resources for N-tier web applications through a multi-level control approach.
Experimental results show that our
solution outperforms other existing approaches– Almost the same performance but save up to
20% CPU resources. – Robust to deal with different workload styles.
23
24
Thanks