Low-Carbon Routing Algorithms For Cloud Computing Services in IP-over-WDM Networks Achille Pattavina To be presented at ICC 2012, Ottawa, Canada

Achille Pattavina Summary  Motivations  Power consumption models  Low-carbon routing algorithms  Case study  Conclusions 2

Achille Pattavina Summary  Motivations  ICT and emissions  Data centers and renewable energy  Power consumption models  Low-carbon routing algorithms  Case study  Conclusions 3

Achille Pattavina ICT and emissions  ICT responsable of 8% of global energy consumption  Up to 20% in 2020  Data centers count for 2% of electrical consumption  ICT is the key for a Low-Carbon Economy  Great benefits in reduction of greenhouse gas emissions using renewable energy sources (sun and wind)  Difference between CO2 reduction and energy efficiency 4

Achille Pattavina Data center and renewable energy  Renewable energy sites are usually in remote locations, hard to connect to electrical grid  Data centers delocalization  Data centers moved nearby renewable energy sources (ex. Google, Microsoft)  Avoid power transmissions lines losses (up to 15 %)  Connected through optical network  Our aim is to do a dynamic-low CO2 emissions routing  «Follow the Wind, Follow the Sun» architectures  Trade-off between transport power and power needed to elaborate each connection inside a DC 5

Achille Pattavina 6 Renewable energy distribution  Solar powered data centers  Power profile during 24 hours of the day  Shifted according to the time zone  Variations mainly do to diurnal cycle of the sun  power from 6 to 22  Wind powered data centers  Power profile during 24 hours of the day  Shifted according to the time zone  Variations due to the waxing and waning of the wind  hard to predict

Achille Pattavina Summary  Motivations  Power consumption models  Data center power supply  Transport power  Processing power  Low-carbon routing algorithms  Case study  Conclusions 7

Achille Pattavina Renewable energy production 8

Achille Pattavina Transport power  Transport cost of a lightpath is function of the hop number H  IP-over-WDM basic configuration (Switching Lv. 3) 9 P t,IP = 2*H*P tr1 + (H-1)*P IP Source: F. Musumeci, M. Tornatore, and A. Pattavina, “A power consumption analysis for IP-Over-WDM core network architectures”

Achille Pattavina Transport power  Transport cost of a lightpath is function of the hop number H  IP-over-WDM basic configuration (Switching Lv. 3)  IP-over-SDH-over-WDM configuration (Switching Lv. 2) 10 Source: F. Musumeci, M. Tornatore, and A. Pattavina, “A power consumption analysis for IP-Over-WDM core network architectures” P t,SDH = 2*H*P tr2 + (H+1)*P SDH +4*P SR P t,IP = 2*H*P tr1 + (H-1)*P IP

Achille Pattavina Transport power  Transport cost of a lightpath is function of the hop number H  IP-over-WDM basic configuration (Switching Lv. 3)  IP-over-SDH-over-WDM configuration (Switching Lv. 2)  IP-over-WDM Opaque configuration (Switching Lv. 1) 11 Source: F. Musumeci, M. Tornatore, and A. Pattavina, “A power consumption analysis for IP-Over-WDM core network architectures” P t,SDH = 2*H*P tr2 + (H+1)*P SDH +4*P SR P t,IP = 2*H*P tr1 + (H-1)*P IP P t,OP = 2*H*P tr2 + (H+1)*P o +2*P SR

Achille Pattavina Transport power  Transport cost of a lightpath is function of the hop number H  IP-over-WDM basic configuration (Switching Lv. 3)  IP-over-SDH-over-WDM configuration (Switching Lv. 2)  IP-over-WDM Opaque configuration (Switching Lv. 1) 12 Source: F. Musumeci, M. Tornatore, and A. Pattavina, “A power consumption analysis for IP-Over-WDM core network architectures” P t,OP = 2*H*P tr2 + (H+1)*P o +2*P SR P t,SDH = 2*H*P tr2 + (H+1)*P SDH +4*P SR P t,IP = 2*H*P tr1 + (H-1)*P IP Network Element Power Consumption *per 10 Gbit/s Trasponder (P tr1 ) 34.5 (W) Trasponder (P tr2 ) (W) Short-reach (P SR ) (W) DXC (P SDH ) (W) IP router (P IP ) 145 (W) Optical switch (P O ) 1.5 (W)

Achille Pattavina Processing power 13 *Baliga et al., “Green cloud computing: Balancing energy in processing, storage and transport”.

Achille Pattavina 14 Data center consumption and design  Server consumption Cooling factor + OH Conversion time

Achille Pattavina Summary  Motivations  Power consumption models  Low-carbon routing algorithms  Anycast routing  Energy-aware routing algorithms  Case study  Conclusions 15

Achille Pattavina Routing problem 16 0 Dummy Node Anycast link CO 2 Is it better to route towards a remote DC with renewable energy ? Anycast Routing Problem Trade-off between transport and processing power

Achille Pattavina Algorithms  Routing algorithms: aim to minimize the total non-renewable (brown) energy of the network (emissions  1 kwh=228 gCO2)  Sun&Wind Energy-Aware Routing (SWEAR): chooses for each connection between two paths  1 - Maximum usage of renewable energy  2 - Lowest transport energy  Green Energy-Aware Routing (GEAR): finds for each connection the path with  Lowest non-renewable (brown) energy  Trade-off bbetween transport power and (green) processing power 17

Achille Pattavina SWEAR A simple example CO 2 Routing towards Green Sources Load Balancing when the traffic on a link overcome the fixed threshold P lg -P ls < P p ?

Achille Pattavina SWEAR Flow chart 19 Weight Assignment for anycast links T1-T2< S Weight Assignment for transport links Save the transport cost of the path in T1 Restore to 1 all weights and calculate the shortest path Save the transport cost of the path in T2 Choose Path 1 Choose Path 2 NO YES Exist? NO YES Route the connection and update the residual capacity on auxiliary graph Block the connection Calculate Shortest Path with these weights

Achille Pattavina 20 GEAR A simple example CO 2 Transport links weight equal to transport power Anycast links weight equal to brown processing power Homogeneous weights asseignment (brown power) for all network links

Achille Pattavina 21 GEAR Flow chart Weight assignment for anycast links Weight assignment for transport links Calculate Shortest Path with these weights NO YES Exist? Route the connection and update the residual capacity on auxiliary graph Block the connection Weight on transport links is equal to the power needed to transport the connection on the link Weight on anycast links is equal to the brown power needed to elaborate the connection in the DC With this assignment, we have homogeneous weights (brown power) on all network links

Achille Pattavina Summary  Motivations  Power consumption models  Low-carbon routing algorithms  Case study  Conclusions 22

Achille Pattavina 23 Case study  Poisson interarrival traffic  Holding time with negative exponential distribution Average duration 1 s  Each connection requires an entire 10 Gbit/s lighpath  Traffic uniformly distributed  24 nodes, links  16 wavelenghts/link  6 Data centers: 1 with Wind energy (1) 2 with Solar energy (14,15) 3 with Non-Renewable energy (5,11,17)  21 nodes, 36 links  16 wavelenghts/link  6 Data centers: 1 with Wind energy (15) 2 with Solar energy (7,20) 3 with Non-Renewable energy (1,11,13)

Achille Pattavina Case study  GEAR and SWEAR compared with two reference algorithms:  Shortest Path (SP)  Best Green Data Center (BGD): always routing towards the DC with more renewable energy 24

Achille Pattavina 25 Results Total Brown Power Reduction in total emissions vs. SP: 21% con SWEAR 24% con GEAR Reduction in total emissions vs. BGD: 25% with SWEAR 27% with GEAR Results for ItalyNet in Opaque architecture configuration

Achille Pattavina Results Other contributions 26 Increase in renewable energy consumption of 120% vs. SP Reduction in processing energy consumption vs. SP 75% with SWEAR 79% with GEAR Increase in renewable energy consumption of 10% vs. BGD Reduction in processing energy consumption vs. BGD 22% with SWEAR 35% with GEAR Increase in transport energy consumption of 28% vs. SP Reduction in transport energy consumption of 26% vs. BGD

Achille Pattavina 27 Results Total Brown Power Reduction in total emissions vs. SP: 8% con SWEAR 11% con GEAR Reduction in total emissions vs. BGD: 20% with SWEAR and GEAR Results for USA24 in Opaque architecture configuration

Achille Pattavina  Comparison with different IPoWDM configurations  Opaque configuration obtains best results  Evolution toward an «all- optic» solution fits better with our approach 28 Results Architecture comparison Switching Lv.3 Switching Lv.2 Switching Lv.1 IPbasic reduction in total emissions of 1% vs. SP IPoSDH reduction in total emissions of 2,7% and 5% vs. SP

Achille Pattavina 29 Results Blocking probability

Achille Pattavina Summary  Motivations  Power consumption models  Low-carbon routing algorithms  Case study  Conclusions 30

Achille Pattavina Conclusions and Future Works 31

Achille Pattavina 32 Thank You!!