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

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 route? Datacenter Latency Electricity Price Carbon footprint DC1 (Texas) Low High High DC2 (Washington) High Low Low Texas Washington Other Other Nuclear 6% 8% Coal 10% 8% Gas 10% Coal Nuclear 37% 9% Hydro Gas 67% 45% 4

Where to route? Datacenter Latency Electricity Price Carbon footprint DC1 (Texas) Low High High A.M. DC2 (Washington) High Low Low Datacenter Latency Electricity Price Carbon footprint DC1 (Texas) Low High Low P.M. DC2 (Washington) High Low High 450 Carbon Footprint (g/kWh) 400 350 300 250 200 150 100 50 0 5 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 Hours of a day

How to split? DC 1 DC 2 6

FORTE and its Contributions FORTE: F low O ptimization based framework for R equest- routing and T raffic E ngineering 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

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 only about 5% of electricity price 8 Electricity Cost "Carbon Cost"

Outline • Model • P1: Assigning users to datacenters • P2: Assigning data objects to datacenters • P3: Datacenter upgrade • Evaluation 9

Model Datacenters: n j Data Objects: d k User Groups: u i Carbon emission: c(n j ) Electricity price: e(n j ) Capacity: cap(n j ) NY LA DC Requests r(u i , d k ) 10

Model Datacenters: n j Data Objects: d k User Groups: u i Carbon emission: c(n j ) Electricity price: e(n j ) Capacity: cap(n j ) NY P3 P2 P1 LA DC Requests r(u i , d k ) 11

12

Outline • Model • P1: Assigning users to datacenters • P2: Assigning data objects to datacenters • P3: Datacenter upgrade • Evaluation 13

Assigning Users to Datacenters Datacenters: n j Data Objects: d k User Groups: u i P1 14

15

16

Datacenter Capacity Constraints Datacenters: n j Data Objects: d k User Groups: u i u2 n4 n4 u3 Capacity: cap(n j ) 17

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

Fast-FORTE • Greedy Heuristic • Running time O(N logN) vs Simplex O(~N 6 ) • Reduces running time from 2 minutes to 6 seconds • Ratio between approximated and optimal objective value is 1.003 19

Outline • Model • P1: Assigning users to datacenters • P2: Assigning data objects to datacenters • P3: Datacenter upgrade • Evaluation 20

Assigning Data Objects to Datacenters Datacenters: n j Data Objects: d k User Groups: u i P2 21

22

23

24

25

Outline • Model • P1: Assigning users to datacenters • P2: Assigning data objects to datacenters • P3: Datacenter upgrade • Evaluation 26

Using FORTE for upgrading datacenters Datacenters: n j Data Objects: d k User Groups: u i P3 27

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

29

30

31

32

33

Outline • Model • P1: Assigning users to datacenters • P2: Assigning data objects to datacenters • P3: Datacenter upgrade • Evaluation 34

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

36

37

Three-way Tradeoff Tradeoff between carbon emissions, average distance, and electricity costs. 38

Two-way Tradeoff between Carbon Emission and Electricity Cost 7 6,8 6,6 Carbon Emission ( ton/hour) 6,4 (987, 6.5) 6,2 6 (987, 5.83) 5,8 (1010, 5.73) 5,6 5,4 5,2 5 850 900 950 1000 1050 Electricity Cost ($/hour) 39

Will Carbon Taxes or Credits Work? Akamai uses ~2 * 10 8 kWh per year • Electricity cost of 2 * 10 8 kWh: 2 * 10 8 kWh * 11.2c/kWh = $22.4 M • “Carbon cost” of 2 * 10 8 kWh : 2 * 10 8 kWh * 500g/kWh = 10 5 t Electricity Cost "Carbon Cost" 10 5 t * $10/t = $1 M 40

Green Upgrades • Low Electricity Price • Use Green Energy WA1 NY1 • Use Green Energy • Reduce Access Latency NJ1 CA1 • Reduce Access Latency NJ2 CA2 TX1 Year 2 Year 3 Year1 Reduces carbon emission by 25% compare to carbon oblivious plan 41

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

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

25% carbon reduction FORTE performs upgrades every June, hence the drop in carbon emissions every 12 months. 44

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

Is electricity price and carbon footprint correlated? 46

PUE • Our experiment: PUE = 1.2 • Google 2012: Average PUE = 1.13, Best PUE = 1.08 • Industrial Average 2011: PUE = 1.8 47

Energy and Emergy (Embedded Energy) Embedded energy accounts for about 20% of the total energy over a datacenter’s lifetime. 1 __________________________ 1. Chang et. at., Green Server Design: Beyond Operational Energy to Sustainability, HotPower2010 48

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

Three-way tradeoff 50

Average Carbon Footprint in U.S. 51

Population Density Map 52

Residential Electricity Price Map 53

Data Source Details 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 54