1 = ~1M servers CO 2 of 280,000 cars 2 Datacenters and Request - - PowerPoint PPT Presentation

▶
1
SMART_READER_LITE
LIVE PREVIEW

1 = ~1M servers CO 2 of 280,000 cars 2 Datacenters and Request - - PowerPoint PPT Presentation

May 23, 2023 267 likes •817 views

Peter X. Gao, Andrew R. Curtis, Bernard Wong, S. Keshav Cheriton School of Computer Science University of Waterloo August 15, 2012 1 = ~1M servers CO 2 of 280,000 cars 2 Datacenters and Request Routing DC 2 Dynamic DNS DC 1 3 Where to

slide-1
SLIDE 1

Peter X. Gao, Andrew R. Curtis, Bernard Wong, S. Keshav

Cheriton School of Computer Science University of Waterloo August 15, 2012

1

slide-2
SLIDE 2

2

=

CO2 of 280,000 cars

~1M servers

slide-3
SLIDE 3

Datacenters and Request Routing

DC 2 DC 1 Dynamic DNS

3

slide-4
SLIDE 4

Where to route?

Datacenter Latency Electricity Price Carbon footprint DC1 (Texas) Low High High DC2 (Washington) High Low Low

Hydro 67% Nuclear 9% Gas 10% Coal 8% Other 6%

Washington

Nuclear 10% Gas 45% Coal 37% Other 8%

Texas

4

slide-5
SLIDE 5

Where to route?

Datacenter Latency Electricity Price Carbon footprint DC1 (Texas) Low High High DC2 (Washington) High Low Low

5

Datacenter Latency Electricity Price Carbon footprint DC1 (Texas) Low High Low DC2 (Washington) High Low High

A.M. P.M.

50 100 150 200 250 300 350 400 450 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24

Carbon Footprint (g/kWh) Hours of a day

slide-6
SLIDE 6

How to split?

6

DC 1 DC 2

slide-7
SLIDE 7

FORTE and its Contributions

FORTE: Flow Optimization based framework for Request- routing and Traffic Engineering Contributions:

– Principled framework for managing the three-way trade-off between access latency, electricity cost, and carbon footprint

  • Green datacenter upgrade plans

– Impact of carbon taxes on datacenter carbon footprint reduction

7

slide-8
SLIDE 8

Surprising Results

  • FORTE can reduce datacenter carbon footprint

by 10% with no increase in electricity cost and access latency

  • Carbon Tax is not effective because taxes are
  • nly about 5% of electricity price

8

Electricity Cost "Carbon Cost"

slide-9
SLIDE 9

Outline

  • Model
  • P1: Assigning users to datacenters
  • P2: Assigning data objects to datacenters
  • P3: Datacenter upgrade
  • Evaluation

9

slide-10
SLIDE 10

Model

User Groups: ui Datacenters: nj Data Objects: dk Requests r(ui, dk)

NY LA DC

10

Carbon emission: c(nj) Electricity price: e(nj) Capacity: cap(nj)

slide-11
SLIDE 11

P3 P2

P1

Model

User Groups: ui Datacenters: nj Data Objects: dk Requests r(ui, dk)

NY LA DC

11

Carbon emission: c(nj) Electricity price: e(nj) Capacity: cap(nj)

slide-12
SLIDE 12

12

slide-13
SLIDE 13

Outline

  • Model
  • P1: Assigning users to datacenters
  • P2: Assigning data objects to datacenters
  • P3: Datacenter upgrade
  • Evaluation

13

slide-14
SLIDE 14

P1

User Groups: ui Datacenters: nj Data Objects: dk 14

Assigning Users to Datacenters

slide-15
SLIDE 15

15

slide-16
SLIDE 16

16

slide-17
SLIDE 17

17

Datacenter Capacity Constraints

n4

User Groups: ui Datacenters: nj Data Objects: dk

u2 u3 Capacity: cap(nj)

n4

slide-18
SLIDE 18

Scale of Linear Program

  • Evaluation problem size:

– Over 1 million variables – FORTE can solve it in approximately 2 min

  • Actual problem:

– Can be over 1 billion variables

18

slide-19
SLIDE 19

Fast-FORTE

  • Greedy Heuristic
  • Running time O(N logN) vs Simplex O(~N6)
  • Reduces running time from 2 minutes to 6

seconds

  • Ratio between approximated and optimal
  • bjective value is 1.003

19

slide-20
SLIDE 20

Outline

  • Model
  • P1: Assigning users to datacenters
  • P2: Assigning data objects to datacenters
  • P3: Datacenter upgrade
  • Evaluation

20

slide-21
SLIDE 21

P2

Assigning Data Objects to Datacenters

21 User Groups: ui Datacenters: nj Data Objects: dk

slide-22
SLIDE 22

22

slide-23
SLIDE 23

23

slide-24
SLIDE 24

24

slide-25
SLIDE 25

25

slide-26
SLIDE 26

Outline

  • Model
  • P1: Assigning users to datacenters
  • P2: Assigning data objects to datacenters
  • P3: Datacenter upgrade
  • Evaluation

26

slide-27
SLIDE 27

P3

Using FORTE for upgrading datacenters

27 User Groups: ui Datacenters: nj Data Objects: dk

slide-28
SLIDE 28

Using FORTE for upgrading datacenters

Datacenter operators need to decide:

– Which datacenters should be upgraded? – How many servers in that datacenter should be upgraded?

The upgrade decisions are based on:

– Estimation of future traffic demands – Annual budget on upgrading – Trade-off between cost and benefit

28

slide-29
SLIDE 29

29

slide-30
SLIDE 30

30

slide-31
SLIDE 31

31

slide-32
SLIDE 32

32

slide-33
SLIDE 33

33

slide-34
SLIDE 34

Outline

  • Model
  • P1: Assigning users to datacenters
  • P2: Assigning data objects to datacenters
  • P3: Datacenter upgrade
  • Evaluation

34

slide-35
SLIDE 35

Datasets

Akamai traffic data – Akamai delivers about 15% - 20% Internet traffic – 3 weeks coarse-grained data in U.S. – Aggregated every 5 minutes U.S. Energy Information Administration – Carbon footprint – Electricity cost Data Objects: Synthetic with long-tail popularity, 10% latency tolerant

35

slide-36
SLIDE 36

36

slide-37
SLIDE 37

37

slide-38
SLIDE 38

Three-way Tradeoff

Tradeoff between carbon emissions, average distance, and electricity costs.

38

slide-39
SLIDE 39

Two-way Tradeoff between Carbon Emission and Electricity Cost

39

5 5,2 5,4 5,6 5,8 6 6,2 6,4 6,6 6,8 7 850 900 950 1000 1050

Carbon Emission ( ton/hour)

(987, 6.5) (987, 5.83) (1010, 5.73)

Electricity Cost ($/hour)

slide-40
SLIDE 40

Will Carbon Taxes or Credits Work?

Akamai uses ~2 * 108 kWh per year

  • Electricity cost of 2 * 108 kWh:

2 * 108 kWh * 11.2c/kWh = $22.4 M

  • “Carbon cost” of 2 * 108 kWh :

2 * 108 kWh * 500g/kWh = 105 t 105 t * $10/t = $1 M

40

Electricity Cost "Carbon Cost"

slide-41
SLIDE 41

Green Upgrades

41

WA1 CA1 CA2 TX1 NY1 NJ1 NJ2 Year1 Year 2 Year 3

Reduces carbon emission by 25% compare to carbon oblivious plan

  • Use Green Energy
  • Use Green Energy
  • Reduce Access Latency
  • Reduce Access Latency
  • Low Electricity Price
slide-42
SLIDE 42

Conclusions

  • FORTE is a request routing framework that can

reduce carbon emissions by 10% without affecting latency and electricity cost

  • Surprisingly, carbon taxes do not provide sufficient

incentives to reduce carbon emissions

  • A green upgrade plan can further reduce carbon

emissions by 25% over 3 years

42

slide-43
SLIDE 43

Acknowledgement

  • We thank Prof. Bruce Maggs for providing us

access to Akamai traces

  • We thank our shepherd Prof. Fabian

Bustamante and the reviewers for their insightful comments

43

slide-44
SLIDE 44

25% carbon reduction

44

FORTE performs upgrades every June, hence the drop in carbon emissions every 12 months.

slide-45
SLIDE 45

Impact on Application Service Providers

There are three types of service providers:

– Integrated Service Provider. e.g.: Google, Facebook – Infrastructure Service Provider. e.g.: Amazon Cloudfront, Microsoft Azure – Content Distribution Network. e.g.:Akamai, Limelight

Service providers that own their infrastructure can use FORTE to reduce their carbon footprint.

45

slide-46
SLIDE 46

Is electricity price and carbon footprint correlated?

46

slide-47
SLIDE 47

PUE

  • Our experiment: PUE = 1.2
  • Google 2012: Average PUE = 1.13, Best PUE = 1.08
  • Industrial Average 2011: PUE = 1.8

47

slide-48
SLIDE 48

Energy and Emergy (Embedded Energy)

48 Embedded energy accounts for about 20% of the total energy

  • ver a datacenter’s lifetime.1

__________________________

  • 1. Chang et. at., Green Server Design: Beyond Operational Energy to

Sustainability, HotPower2010

slide-49
SLIDE 49

Datacenter electricity consumption and carbon footprint

  • Datacenters consume 1.3% of worldwide

electricity and it is expected to grow to 8% in

  • 2020. [Analytics Press]
  • Datacenters carbon emission is about 0.6% of

world total and it is expected to grow to 2.6% in 2020, exceeding Germany. [McKinsey Quarterly, Nov 2008]

49

slide-50
SLIDE 50

Three-way tradeoff

50

slide-51
SLIDE 51

Average Carbon Footprint in U.S.

51

slide-52
SLIDE 52

Population Density Map

52

slide-53
SLIDE 53

Residential Electricity Price Map

53

slide-54
SLIDE 54

Data Source Details

54

Data Source User Groups Akamai Datacenters Location Akamai Datacenters Capacity Akamai User Requests Akamai Datacenters Carbon footprint U.S. Energy Information Administration Datacenters Electricity Price U.S. Energy Information Administration Data Objects Simulated Access Latency Approximated by geographical distance