November 15 - 19, 2009 SERVICE COMPUTATION 2009

Analysis of Energy Efficiency in Clouds

H. AbdelSalam K. Maly (maly@cs.odu.edu)

R. Mukkamala M. ZubairDepartment of Computer Science,

Old Dominion University

D. KaminskyIBM, Raleigh, North Carolina

1

Outline

• Cloud Computing• Change Management• Power Management

– Pro-active approach– Minimize total power consumption– Constraints:

• SLAs• Prior change management commitments

– Compute possible time slots for change management task

November 15 - 19, 2009 SERVICE COMPUTATION 2009 2

Cloud Computing

• A cloud can be defined as:– a pool of computer resources that can host a variety

of different workloads, including batch-style back-end jobs and interactive user applications.

• A cloud computing platform dynamically provisions, configures, reconfigures, and deprovisions servers as needed.

• Servers in the cloud can be physical machines or virtual machines.

• Customers have Service Level Agreements to buy computing services from cloud manager

November 15 - 19, 2009 SERVICE COMPUTATION 2009 3

Change Management

• Managing large IT environments such as computing clouds is expensive and labor intensive.

• Servers go through several software and hardware upgrades.

• IT organizations handle change management through human group interactions and coordination.

November 15 - 19, 2009 SERVICE COMPUTATION 2009 4

Pro-active Approach

• We proposed earlier and implemented an infrastructure-aware autonomic manager for change management– scheduler that computes possible open time slots in

which changes can be applied without violating any of SLAs reservations.

• Here we propose pro-active energy-aware technique for change management in a cloud computing environment.

November 15 - 19, 2009 SERVICE COMPUTATION 2009 5

November 15 - 19, 2009 SERVICE COMPUTATION 2009 6

Cloud Computing Architecture

Job distribution

• applications in a cloud computing: – intensive compute processing, non-

interactive applications– user interactive: Web applications and Web

services are typical examples.

November 15 - 19, 2009 SERVICE COMPUTATION 2009 7

Non-interactive applications

• dedicate one or more servers to each of these applications, number of dedicated servers depends on the underlying SLA and the availability of servers in the cloud

• servers should be run at their top speed (frequency) so the application can finish as soon as possible

November 15 - 19, 2009 SERVICE COMPUTATION 2009 8

Job distribution

• Assume that, based on its SLA, Job X requires s seconds response time for u users.

• From the historical data for Job X, we estimate the average processing required for a user query to be l instructions.

• Assume that job X is to be run on a server that runs on frequency f and on the average requires CPI clock ticks (CPU cycles) to execute an instruction.

• the server can execute q=(s*f)/(l*CPI) user queries within s seconds.

• If q<u , then the remaining (u-q) user requests should be routed to another server.

November 15 - 19, 2009 SERVICE COMPUTATION 2009 9

System model

• estimate the computing power (MIPS) needed to achieve the required response time

• client provides a histogram that shows the frequency of each expected query

• replace the minimum average response time constraint in SLA by the minimum number of instructions that the application is allowed to execute every second

November 15 - 19, 2009 SERVICE COMPUTATION 2009 10

November 15 - 19, 2009 11Distribution of jobs onto serversSERVICE COMPUTATION 2009

System model

• Conversion of response time to MIPS

– If user query has average response time of t1 seconds when it runs solely on a server configuration with x MIPS (million instructions per second), this can be benchmarked for each server configuration), then

– to have an average response time of t2 seconds, it is required to run the query such that it can execute a minimum of (t1*x)/t2 million instructions per second.

• Power management of server

– Minimum Fmin

– Maximum Fmax

– Discrete values in between

• Power – frequency relation

November 15 - 19, 2009 SERVICE COMPUTATION 2009 12

3BfAP

Mathematical analysis

• given k servers that should run on frequencies respectively, such that total compute load is:

the total energy consumption is given by,

November 15 - 19, 2009 SERVICE COMPUTATION 2009 13

kk ffff ,, 121

kT fffL 21 .

].)([** 31

1

31

32

31

k

jjTk fLfffBAkP

Mathematical analysis

• the number of servers k, that should run to optimize power consumption, is (assuming continuous frequency spectrum):

• Each server should run at frequency

November 15 - 19, 2009 SERVICE COMPUTATION 2009 14

TLA

Bk 3 *

2

kLf T /

Sample cloud load

November 15 - 19, 2009 SERVICE COMPUTATION 2009 15

Actual and Approximated Load due to several SLAs.

Servers available for change management

• in each time segment, – the number of idle servers in the cloud equals the difference

between the total number of cloud servers and kt.

– idle server is a candidate for change management.

November 15 - 19, 2009 SERVICE COMPUTATION 2009 16

)(*2

3 tLA

Bk Tt

November 15 - 19, 2009 SERVICE COMPUTATION 2009 17

Servers Available for changes as a function of time

Scenario comparison

• Total energy consumption during one period (one day) using the pro-active approach is 37305 Watt-Hour, for an average of 1554 Watt.

• Total and the average energy consumption when using 5 % over-provisioning at various frequencies:

November 15 - 19, 2009 SERVICE COMPUTATION 2009 18

Frequency Total (Watt.Hour) Average (Watt) 1.0 GHZ 64861 2703 2.0 GHZ 39324 1639 2.4 GHZ 42742 1781 3.0 GHZ 58246 2427

November 15 - 19, 2009 SERVICE COMPUTATION 2009

Conclusion

• Pro-active management is the computation of when servers will be idle so they can be scheduled for change maintenance.

• Pro-active power management leads to considerable saving in total energy consumed, for specific examples ranging from 5-75%.

• Can be modified to include compute intensive jobs

• Can be modified to include hardware failure rates

19