CloudMonitor: Profiling Power Usage

38
University of St Andrews School of Computer Science 1 or can we charge for Energy? Profiling Power at Cloud scale @jws7 James Smith Ali Khajeh-Hosseini Jonathan Ward Ian Sommerville

description

A presentation I gave at IEEE Cloud 2012 – one of the premier conferences on Cloud Computing for a paper by Ali Khajeh-Hoessini and myself.The paper is entitled “CloudMonitor: Profiling Power Usage” and discusses the CloudMonitor software and how it could be used to create an energy tariff for Cloud providers.I’m very grateful to my supervisor Prof Ian Sommerville for encouraging me to go to the conference, to LSCITS for funding the trip and for Google for giving me the time off!More information on CloudMonitor can be found here: http://jws7.net/cloudmonitor

Transcript of CloudMonitor: Profiling Power Usage

Page 1: CloudMonitor: Profiling Power Usage

University of St Andrews

School of Computer Science

1

or can we charge for Energy?

Profiling Powerat Cloud scale

@jws7

James SmithAli Khajeh-HosseiniJonathan WardIan Sommerville

Page 2: CloudMonitor: Profiling Power Usage

University of St Andrews

School of Computer Science

2

or can we charge for Energy?

Profiling Power

CloudMonitorat Cloud scale

@jws7

James SmithAli Khajeh-HosseiniJonathan WardIan Sommerville

Page 3: CloudMonitor: Profiling Power Usage

University of St Andrews

School of Computer Science

3

Operational costs exceeding purchase costs

Page 4: CloudMonitor: Profiling Power Usage

University of St Andrews

School of Computer Science

4

Even over a relatively short lifespan (3-5 years). Mainly driven by energy costs

Operational costs exceeding purchase costs

Page 5: CloudMonitor: Profiling Power Usage

University of St Andrews

School of Computer Science

5

Even over a relatively short lifespan (3-5 years). Mainly driven by energy costs

Operational costs exceeding purchase costs

We need realistic Energy estimations

Page 6: CloudMonitor: Profiling Power Usage

University of St Andrews

School of Computer Science

6

Even over a relatively short lifespan (3-5 years). Mainly driven by energy costs

Operational costs exceeding purchase costs

We need realistic Energy estimationsnot just included in “infrastructure”

Page 7: CloudMonitor: Profiling Power Usage

University of St Andrews

School of Computer Science

Monitoring

7

“then you cannot improve it...”

If you cannot measure something

- Lord Kelvin

Page 8: CloudMonitor: Profiling Power Usage

University of St Andrews

School of Computer Science

Monitoring

8

“then you cannot improve it...”

If you cannot measure something

- Lord Kelvin

13.4% of organisations monitor their energy consumption!

Page 9: CloudMonitor: Profiling Power Usage

University of St Andrews

School of Computer Science

Datacentres

9

Page 10: CloudMonitor: Profiling Power Usage

University of St Andrews

School of Computer Science

Datacentres

10

Page 11: CloudMonitor: Profiling Power Usage

University of St Andrews

School of Computer Science

Datacentres

11

Page 12: CloudMonitor: Profiling Power Usage

University of St Andrews

School of Computer Science

Datacentres

12

especially when you need additional hardware!

Monitoring at Scale is hard

Page 13: CloudMonitor: Profiling Power Usage

University of St Andrews

School of Computer Science

CloudMonitor

13

Page 14: CloudMonitor: Profiling Power Usage

University of St Andrews

School of Computer Science

CloudMonitor

14

Use SIGAR library to gather resource usage

Page 15: CloudMonitor: Profiling Power Usage

University of St Andrews

School of Computer Science

CloudMonitor

15

Use SIGAR library to gather resource usage

Correlate with Power Meter

readings

Page 16: CloudMonitor: Profiling Power Usage

University of St Andrews

School of Computer Science

CloudMonitor

16

Use SIGAR library to gather resource usage

Correlate with Power Meter

readings

Generate Power Model

Page 17: CloudMonitor: Profiling Power Usage

University of St Andrews

School of Computer Science

17

So what?

Page 18: CloudMonitor: Profiling Power Usage

University of St Andrews

School of Computer Science

18

Generate power model on one machine. Roll out to 1,000

So what?

Page 19: CloudMonitor: Profiling Power Usage

University of St Andrews

School of Computer Science

19

Generate power model on one machine. Roll out to 1,000

So what?

of the same configuration

Page 20: CloudMonitor: Profiling Power Usage

University of St Andrews

School of Computer Science

Previous work

• by Bohra, Ata & Chaundary - VMeter

• linear power model is workload-specific

• our work is workload-independent

• calculate once - valid for all machines of the same configuration

20*Husain Bohra, Ata E, and Vipin Chaudhary. “VMeter: power modelling for virtualized clouds.” 2010 IEEE International Symposium on Parallel & Distributed Processing, Workshops and Phd Forum (IPDPSW) (April 2010): 1-8.

Page 21: CloudMonitor: Profiling Power Usage

University of St Andrews

School of Computer Science

21

Page 22: CloudMonitor: Profiling Power Usage

University of St Andrews

School of Computer Science

22

Page 23: CloudMonitor: Profiling Power Usage

University of St Andrews

School of Computer Science

23

Not perfect

Page 24: CloudMonitor: Profiling Power Usage

University of St Andrews

School of Computer Science

24

Not perfect

but pretty close...

Page 25: CloudMonitor: Profiling Power Usage

University of St Andrews

School of Computer Science

25

Experiments show 96% accuracy of Software Predictions against PDU Measurements

The lesson?

Page 26: CloudMonitor: Profiling Power Usage

University of St Andrews

School of Computer Science

An Energy Tariff for Cloud

26

Page 27: CloudMonitor: Profiling Power Usage

University of St Andrews

School of Computer Science

An Energy Tariff for Cloud

27

Energy rates are rising at 15% pa

Page 28: CloudMonitor: Profiling Power Usage

University of St Andrews

School of Computer Science

An Energy Tariff for Cloud

28

36 month cost of test system:

Energy rates are rising at 15% pa

Page 29: CloudMonitor: Profiling Power Usage

University of St Andrews

School of Computer Science

An Energy Tariff for Cloud

29

$2,76736 month cost of test system:

Energy rates are rising at 15% pa

Page 30: CloudMonitor: Profiling Power Usage

University of St Andrews

School of Computer Science

Shop for Cloud

30

Page 31: CloudMonitor: Profiling Power Usage

University of St Andrews

School of Computer Science

Amazon AWS EU Cloud

31

Category Cost

Data Transfer $58,084

VM hours $40,568

Storage $2,325

Storage I/O $2,293

Total $103, 270

Page 32: CloudMonitor: Profiling Power Usage

University of St Andrews

School of Computer Science

Hypothetical Cloud thatcharges for Energy

32

Category Cost

Data Transfer $58,084

VM hours $40,568

Energy $2,767

Storage $2,325

Storage I/O $2,293

Total $106, 037

Page 33: CloudMonitor: Profiling Power Usage

University of St Andrews

School of Computer Science

33

Energy would account for 2.6% of the total costs

The lesson?

Page 34: CloudMonitor: Profiling Power Usage

University of St Andrews

School of Computer Science

34

Energy would account for 2.6% of the total costs

The lesson?

and is likely to rise in the future

Page 35: CloudMonitor: Profiling Power Usage

University of St Andrews

School of Computer Science

Summary

35

CloudMonitorDistributed Monitoring Platform

Power Models

Resource Usage

Page 36: CloudMonitor: Profiling Power Usage

University of St Andrews

School of Computer Science

36

Page 37: CloudMonitor: Profiling Power Usage

University of St Andrews

School of Computer Science

37Everybody benefits!

Page 38: CloudMonitor: Profiling Power Usage

University of St Andrews

School of Computer Science

Thanks for listening

38

@jws7

jws7.net/cloudmonitor