Upload
sivadon-chaisiri
View
290
Download
1
Embed Size (px)
DESCRIPTION
The presentation of my accepted paper in 2009 IEEE/ACM International Symposium on Cluster Computing and the Grid, in Shanghai, China, May 18-21, 2009
Citation preview
Optimal Power Management for
Server Farm to Support Green Computing
Dusit Niyato , Sivadon Chaisiri, and Lee Bu Sung, Francis
[email protected], [email protected], [email protected] of Computer Engineering
Nanyang Technological University, Singapore
IEEE/ACM International Symposium on
Cluster Computing and the Grid (CCGrid 09)
May 19, 2009
Outline
• Introduction• System Model• Challenges• Optimization Formulation• Performance Evaluation• Summary
2
Introduction
• Green Computing is the study and practice of using computing resource efficiently, not only the performance but energy
• We firstly propose an optimal power management (OPM) used by a batch scheduler in a server farm
• This OPM observes the state of the server farm, then make the decision to switch the operation mode (active / sleep) of the server
• An optimal decision of OPM is obtained by the constrained Markov decision process (CMDP)
• We consider the system with a job broker to assign users to multiple server farms while the cost is minimized
3
Power Management
• Power management is a major approach of green computing• Power management is applied to control power consumption and
operation of computing resources• Two levels of power management
– Machine level (e.g., some components can be suspended)– Network level (e.g., a node in a server farm can be turned to
sleep mode)• Our work is based on the network level power management
4
System Model
5
OPM is a part of a batch scheduler in a server farm
Incoming jobs
Server in active mode
Server in sleep mode
Batch scheduler
Job
bro
ker
Serverfarm
Serverfarm
users
Challenges
• Uncertainty
– Job arrival is random; users generate job randomly
– Job size and thus processing time is random
6
Incoming jobs
Server in active mode
Server in sleep mode
Batch scheduler
Challenges
• Questions to Be Answer
– When and how many servers to be switched between active and
sleep modes ?– Which server farm should be chosen for a user ?
7
Incoming jobs
Server in active mode
Server in sleep mode
Batch scheduler
Power and Workload Management
• OPM is a part of a batch scheduler• An optimal decision of OPM is obtained by formulating and solving
the constrained Markov decision process (CMDP)• Markov decision process (MDP)
– a discrete time stochastic control process characterized by a set of states; in each state there are several actions from which the decision maker must choose
– For state s and action a, a state transition function Pa(s) determines the transition probabilities to the next state
– The decision maker earns a reward (incursa cost) for each state transition
8
Optimization Formulation of OPM
• State Space
– (X,S): Composite state of server farm
– X: Number of jobs in queue
– S: Number of servers in active mode
• Decision Epoch
– Time slot
• Action
– Us: Number of servers to be switched between active and sleep modes
9
Time Slots and Actions
10
Time slot t-2
Time
Time slot t-1
Action: -2Switch two servers
to sleep mode
Five servers are active Three servers
are active,two servers are
switching to sleep mode
Action: 0Do nothing
Action: +1Switch one server
to active mode
Time slot t
Time slot t+1
Three servers are active,
one server isswitching toactive mode
Three servers are active
Time slot t+2
Four servers are active,
one server isswitching toactive mode
Action: +1Switch one server
to active mode
Action: 0Do nothing
Action: …
Formulation of OPM
11
Minimize power consumption
Waiting time requirement
Bellman’s equation
Loss requirement
Job Broker• The system with multiple server farms and multiple users are
considered
• A job broker assigns the user to the appropriate server farm such that the power consumption cost and network cost of a system is minimized
12
Formulation of Assignment Problem for Job Broker
13
Minimize total cost
Total delay requirement
Performance Evaluation
An individual server farm with batch schedule + OPM• A discrete-time simulation is used to verify the correctness of an
analytical model• The optimal decision (or policy) is made after the state of the system
is observed• Parameter Setting
– Power consumption watts– The job dropping probability requirement – The size of time slot seconds– The waiting time requirement seconds
– The min and max number of servers to be mode switched
14
Pact=400 ,P slp=40Bmax=10
−3
T=20W max=150
Amax=2
Performance of Individual Server Farm
15
A set of actions given the number of jobs in a queue and the number of servers in active mode
0
5
10
150
5
10
2
1
0
1
2
Number of servers in active modeNumber of jobs in queue
Act
ion
Performance of Individual Server Farm
16
6 7 8 9 10 11 12 13 14 151600
1700
1800
1900
2000
2100
Total number of servers (S)
Ave
rage
pow
er c
onsu
mpt
ion
(wat
ts)
Wmax=150
Wmax=200
Simulation
Average power consumption under different total number of servers
Performance of Individual Server Farm
17
The minimum power consumption given different waiting time and job blocking probability requirements
150 200 250 3001200
1250
1300
1350
1400
Maximum waiting time (Wmax)
Pow
er c
onsu
mpt
ion
(wat
ts)
Bmax=0.001
Bmax=0.01
Bmax=0.02
Bmax=0.03
Performance Evaluation
The system with multiple server farms and a job broker• Two server farms are evaluated ( F = 2 )• Multiple users ( U = 20 ) are coming to the job broker• The network cost is represented by a distance between location of
user and location of server farm• A number of servers per server farm is 10 ( )• Two different scenarios
– Identical power consumption cost ( / Wh)– Different power consumption costs ( / Wh )
18
C1 p =C 2
p=0 .01
S1=S2=10
C1 p =0 .01 ,C2
p =0.012
Performance of Multiple Server Farms
19
The assignment of the users to the server farms under different power consumption costs
UserServer farm 1Server farm 2Assignment of user to server farm
Summary
• We have considered the power management issue in green computing
• We have first proposed OPM, formulated as CMDP, for an individual server farm
• Our OPM can dynamically reduce the power consumption of servers in a farm by switching them to sleep mode
• The system with multiple server farms has been also considered in which the job broker has been optimized to assign the user to the server farm
• Future work– Computing resource planning under uncertainty
– Virtualization + Cloud ...
20