Topology design, Embodied Energy and Peer-to-Peer Networks

Prof. Jaafar Elmirghani, University of Leedsj.m.h.elmirghani@leeds.ac.uk

Contributors: X. Dong, Ahmed Lawey, T. El-Gorashi

• Previous work

• Physical Topology Optimization Considering Embodied Energy• Network Devices Embodied Energy• Optimized Topologies Considering Operational and Embodied Energies

• Energy-Efficient Data Compression for Optical Networks

Outline

2

Energy-Efficient Data Compression for Optical Networks• Power Consumption of Data Compression • MILP Model for Data compression in IP over WDM networks • Energy-Efficient Data Compression and Routing Heuristic

• Energy-Efficient BitTorrent• MILP Model for Energy-Efficient BitTorrent • Energy-Efficient BitTorrent Heuristic

WP1: Energy efficient network architecture

Renewable energy Topology optimisation

3

Data centresContent distribution networks

• Introducing additional devices to minimize the operational energy mightincrease the embodied energy and consequently the total Carbon footprint ofthe network.

• The average commercial lifetime (LT) of network devices is estimated as 10years and the maintenance adds 10% of the device production embodiedenergy EEMB-p annually.

• Objective: MILP Minimize:

Physical Topology Optimization Considering Embodied Energy

• The embodied energy of most network devices is mainly composed of: PrintedCircuit Boards (PCB), semiconductor devices, bulk materials and metal.

Materials/Processing Embodied EnergyMJ/kg

Densityg/m2

Semiconductor device 120000 400 (on PCBs)Metals 100-400 Various

Bulk materials 20-400 VariousPCB 300-500 2000-4500

The Embodied Energy and the Density of the Different Materials of Network Devices

Network Devices Embodied Energy

The Embodied Energy of CRS-1 16 Slots Chassis Routing SystemCRS-1 16 Slots Chassis Routing System

Module Dimension (cm) Weight (kg)

Embodied energy (MJ)Units

Total(GJ)

PCB Semiconductor Bulk Materials

Metals

IPPort

PLIM H52.3, D47.2, W4.6 3.8 555 9480 144 900 16 177.3

MSC H52.3, D47.2, W4.6 6.68 555 8280 200 2000 16 176.6

Power H50,D46,W90 (estimate)

35 980 1440 1300 11900 1 15.6

RP H52.3, D28.4, W7.1 5.8 335 7080 228 1800 2 18.9

FC H52.2, D28.5, W7.1 5.6 223 4920 224 1820 2 14.4

SM H52.3, D28.5, W3.6 5.4 335 6960 182 1690 8 73.3

Fan Tray N/A 20 0 0 0 8000 2 16

System Chassis

N/A 486 0 0 19440 174960 1 194.4

Total embodied energy of a full load CRS-1 16 Slots Chassis Routing System 686.5

The Embodied Energy of Active Network Devices

Device Dimension (cm) Weight (kg)Embodied energy (MJ) Total

(GJ)PCB Semiconductor Bulk Materials Metals

Transponder H32.1, D22.8, W2.3 1.4 164 3480 40 380 4.1

EDFA H4.5, D25.9, W48.3 3.08 135 3393 224 899 4.7

Regenerator H4.4, D30, W43.9 4.4 197 4425 320 1100 6

Multi/Demultiplexer H32.1, D22.8, W2.3 1.5

(Estimated) 164 2446 225 414 3.2

PLIM: Physical layer interface module; MSC: Module service card; RP: router processor; FC: Fan controller; SM: Switch module

Network Devices Embodied EnergyThe Embodied Energy of the 192x192 Glimmerglass Optical Switch

Materials/Processing Embodied Energy (MJ) Weight (g)

Total Embodied Energy (GJ)

SCS processing 30.3 0.253

11Semiconductor device 4116 34.3

Metals 5440 13600Bulk materials 1200 (Estimated) 3000

PCB 220.5 (Estimated) 490

The Embodied Energy per km of the GYTY53 Optical CableComponent Material Thickness or

Diameter(estimation)

Weightkg/km

EmbodiedEnergyMJ/km

PE outer sheath PE 3mm 122.46(analysis)

9907

Steel tape steel 0.5 mm 37.5 (analysis) 1200 MJ/kmPE inner sheath PE 1 mm 25.12 (analysis) 2302 MJ/kmStrength member steel 2 mm 24.8 (analysis) 793 MJ/kmFibers glass 125 μm 1.73 (analysis) 123 MJ/kmLoose tube (6 items) PBT 1 mm 25.2 (estimated) 2245 MJ/kmFilling compound Polymers -- 14.9 (estimated) 1490 MJ/kmTotal embodied energy 18.059 GJ/km

The Embodied Energy per km of the GYTY53 Optical Cable

SCS: Single crystal silicon; PE: PolyEthylene

Network and trafficLifetime (10 year) energy, original NSFNET

7

Distance between two neighbouring EDFAs 80 (km)

Capacity of each wavelength (B) 40 (Gb/s)

Power consumption of a router port (PR) 1000 (W)

Power consumption of a transponder (PT) 73 (W)

Power consumption of an EDFA (PE) 8 (W)

Power consumption of an optical switch (PO) 85 (W)

Power consumption of a multiplexer/demultiplexer (PMD) 16 (W)

Optimized Physical Topologies Considering Operational and Embodied energies

Symmetric traffic, non-bypass Asymmetric traffic,

non-bypass

• Large embodied energy of theoptical cableà shorter links.

• The embodied energy is themajor contributor to the totalnetwork energy consumption non-bypass

• Significant embodied energy savings of 20% and 59% are achieved compared to the original NSFNET topology and the operational-power-optimized topology, respectively resulting in a total energy saving of 47% and 13%.

Energy-Efficient Data Compression for Optical Networks

• Data compression is becoming a widely used technique to savebandwidth which will consequently result in energy savings.

• Trade-off between the energy consumption of computational resourcesand memory required to compress and decompress data and the networkenergy savings.

• Cisco forecasts that 90% of the Internet traffic will be video by 2015.

• In [1], the authors considered semantic compression to reduce the videostorage space.

• YouTube videos can be compressed by a ratio of 20:1 compared toordinary histogram representations.

_______________________________________________________________________________1. Jörn Wanke et. al,”Topic Models for Semantics-preserving Video Compression,” ACM

International Conference on Multimedia Information Retrieval (ACM MIR), Philadelphia, PA,2010.

RC is the data compression ratio A and β are parameters

Power Consumption of Data Compression• In [2], the data compression energy

consumption per bit is given as:

_________________________________________________________________________________________________________2. Dan Kilper et. al ”insights on coding and transmission energy in optical networks”, E-energy 2011

A and β are parametersA is given as:

ε is a scaling parameteris the maximum data compression ratio

β represents the efficiency of the data compression algorithm. ENet is the energy consumption of the network.Therefore:

MILP Model for Data compression in IP over WDM networks

Subject to:Including:

Objective: minimize

Including:

Flow conservation constraintin the IP layer

Linear approximation of the relationship between power consumption of data

compression and data compression ratio

Limit on the maximum data compression ratio

Results

Traffic type

Compression algorithm

Compression ratio

Text bzip2, ppmd (lossless) 4:1

Image JPEG, GIF, PNG (lossy) 10:1

Video MPEG-4, H.264(lossy) 20:1

• The power consumption ofdecompression is equal to thepower consumption ofcompression.

• We consider a mixture of traffic(video, images, text) to reflect theglobal Internet traffic where 91% ofthe global Internet is video.

• Average power savings of 29% and

Algorithms and Compression Ratios for Different types of data

Power consumption under the bypass approach

• Average power savings of 29% and39% are achieved by the MILPmodel under the bypass approachfor β=1 and β=2, respectively.

• Comparable power savings areachieved by the energy-efficientdata compression and routingheuristic.

• High power savings of 45% and55% for β=1 and β=2, respectivelyare achieved under the non-bypassapproach.

Results

• The optimal data compression ratio for most of the node pairs variesslightly (between 70%-80%) under both the maximum and minimumtraffic demands.

Low Traffic Demand (6 am), Bypass High Traffic Demand (10 pm), Bypass

• We have also analysed the impact of compression on the BER

Energy-Efficient BitTorrent

• The two content distribution schemes, Client/Server (C/S) and Peer-to-Peer(P2P), account for a high percentage of the Internet traffic.

• We investigate the energy consumption of BitTorrent in IP over WDM networks.

• We show, by mathematical modelling (MILP) and simulation, that peers’ co-location awareness, known as locality, can help reduce BitTorrent’s cross trafficand consequently reduces the power consumption of BitTorrent on the networkside.

Energy-Efficient BitTorrent

• The file is divided into small pieces.

• A tracker monitors the group of users currently downloading.

• Downloader groups are referred to as swarms and their members as peers. Peers aredivided into seeders and leechers.

• As a leecher finishes downloading a piece, it selects a fixed number (typically 4) ofinterested leechers to upload the piece to, ie unchoke, (The choke algorithm).

• Tit-for-Tat (TFT) ensures fairness by not allowing peers to download more than they upload.

• We consider 160,000 groups of downloaders distributed randomly over the NSFNETnetwork nodes.network nodes.

• Each group consists of 100 members.

• File size of 3GB.

• Homogeneous system where all the peers have the same upload capacity of 1Mbps.

• Optimal Local Rarest First pieces dissemination where Leechers select the least replicatedpiece in the network to download first.

• BitTorrent traffic is 50% of total traffic.

• Flash crowd where the majority of leechers arrive soon after a popular content is shared.

• We compare BitTorrent to a C/S model with 5 data centers optimally located at nodes 3, 5,8, 10 and 12 in NSFNET.

• The upload capacity and download demands are the same for BitTorrent and C/Sscenarios (16Tbps).

MILP Model for Energy-Efficient BitTorrentObjective: Maximize

Subject To: Including:

Setting β=0 gives the original BitTorrent

Subject To: Including:

Fairness constraint, Tit-For-TAT (TFT)

Peers upload rate constraints

Peers download rate constraint

Original BitTorrent (Random Selection)

Peer Selection (100 Peer: 30 Seeders and 70 Leechers in Swarm 1)

17

Energy Efficient BitTorrent (Optimized Selection)

Energy-Efficient BitTorrent Heuristic

• Energy-Efficient BitTorrent modelperforms peer selection based onthe co-location of peers within thesame nodes to minimize energyconsumption.

• The heuristic tries to mimic thisbehavior by:

• Seeders span the neighboringnodes only.

• Leechers are limited to theirlocal nodes as long as thereare sufficient number of peers(5 at least), otherwise theyspan to neighboring nodes.

Results

Energy-efficient heuristic achieves a13% lower download rate.

Average Download Rate

Non-bypass:MILP avg Power Saving=36%Heuristic avg Power Saving =36%

Bypass:MILP avg Power Saving=30%Heuristic avg Power Saving =28%

BitTorrent Power Consumption

Results

Energy Consumption

Non-bypass:MILP average Energy Saving=36%

Heuristic average Energy Saving =25%

Bypass:MILP average Energy Saving=30%

Heuristic average Energy Saving =15%

Conclusions

• It is essential to consider embodied as well as operational energy if the goal is to minimise the network’s carbon footprint.

• Significant embodied energy savings of 20% and 59% are achieved compared to the original NSFNET topology and the operational-power-optimized topology, respectively resulting in a total energy saving of 47% and 13%.

• Shown that power savings can be achieved through the use of data

21

• Shown that power savings can be achieved through the use of data compression and appropriate routing heuristics.

• Average power savings of 29% and 39% are achieved by the MILP model under the bypass approach for β=1 and β=2, respectively.

• High power savings of 45% and 55% for β=1 and β=2, respectively are achieved under the non-bypass approach.

• Introducing locality to BitTorrent can lead to a more efficient content distribution scheme compared to C/S with power savings of 30% (bypass) and 36% (nonbypass).

Related Publications1. Dong, X., El-Gorashi, T.E.H. and Elmirghani, J.M.H., “IP Over WDM Networks Employing Renewable Energy

Sources,” IEEE/OSA Journal of Lightwave Technology, vol. 27, No. 1, pp. 3-14, 2011.2. Dong, X., El-Gorashi, T.E.H. and Elmirghani, J.M.H., “Green IP over WDM Networks with Data Centres,”

IEEE/OSA Journal of Lightwave Technology, vol. 27, 2011.3. Dong, X., El-Gorashi, T.E.H. and Elmirghani, J.M.H., “On the Energy Efficiency of Physical Topology Design for IP

over WDM Networks,” IEEE/OSA Journal of Lightwave Technology, vol. 28, 2012.4. Dong, X., El-Gorashi, T.E.H. and Elmirghani, J.M.H., “Renewable Energy in IP Over WDM Networks,” Proc IEEE

12th International Conference on Transparent Optical Networks ICTON 2010, June 27 - July 1, 2010, Munich, Germany, invited paper.

5. Dong, X., El-Gorashi, T.E.H. and Elmirghani, J.M.H., “hybrid-power IP over WDM network,” Proc IEEE Seventh International Conference on Wireless and Optical Communications Networks WOCN2010, September 2010.

6. Dong, X., El-Gorashi, T.E.H. and Elmirghani, J.M.H., “An Energy Efficient IP over WDM Network,” Proc. IEEE/ACMInternational Conference on Green Computing and Communications, GREENCOM, Hangzhou, China, Dec. 2010.International Conference on Green Computing and Communications, GREENCOM, Hangzhou, China, Dec. 2010.

7. Dong, X., El-Gorashi, T.E.H. and Elmirghani, J.M.H., “Renewable Energy for Low Carbon Emission IP over WDM networks,” Proc. 15th IEEE Optical Network Design and Modelling conference (ONDM’11), Bologna, Italy, 8-10 Feb 2011.

8. Dong, X., El-Gorashi, T.E.H. and Elmirghani, J.M.H., “Low Carbon Emission IP over WDM network,” IEEE International Conference on Communications (ICC’11), Koyoto, Japan, June 2011.

8. Audzevich, Y., Moore, A., Rice, A., Sohan, R., Timotheou, S., Crowcroft, J., Akoush, S., Hopper, A., Wonfor, A., Wang, H., Penty, R., White, I., Dong, X., El-Gorashi, T. and Elmirghani, J., “Intelligent energy aware networks,” book chapter, published in Handbook of Energy-Aware and Green Computing, Taylor and Francis, invited, 2011.

9. Dong, X., El-Gorashi, T.E.H. and Elmirghani, J.M.H., “Energy-Efficient IP over WDM Networks with Data Centres,” Proc IEEE 12th International Conference on Transparent Optical Networks ICTON 2011, 26 – 30 June, 2011, Stockholm, Sweden, invited paper.

10. Osman, N. I., El-Gorashi, T.E.H. and Elmirghani, J.M.H., “Reduction of Energy Consumption of Video-on-Demand Services using Cache Size Optimization,” Proc IEEE Eighth International Conference on Wireless and Optical Communications Networks WOCN2011, May 2011.

11. Dong, X., Lawey, A.Q., El-Gorashi, T.E.H. and Elmirghani, J.M.H., “Energy Efficient Core Networks,” Proc 16th IEEE Conference on Optical Network Design and Modelling (ONDM’12), 17-20 April, 2012, UK.