Combined Smart Sleeping and Power Scaling for Energy Efficiency in Green Data Center Networks

Transcription

1 (May 14 th 17 th 2013, Krabi, Thailand) Combined Smart Sleeping and Power Scaling for Energy Efficiency in Green Data Center Networks Nguyen Huu Thanh Department of Communications Engineering School of Electronics and Telecommunications Hanoi University of Science and Technology

2 Implementation Results Conclusions Contents Motivations Why going green? Energy-aware data centers Ideas Implementations Analytical models Testbed architecture Results Conclusions 2

3 Implementation Results Conclusions Today s Network Status and Trend 40% per annum growth in network traffic 10% per annum growth in number of users Data center traffic increases quickly Distributed data center is becoming more common New applications lead to increased traffic volume in the core of the network Cloud applications: Dropbox, Sky Drive, Google Docs etc. YouTube, Flickr etc. Social networking: Facebook, Twitter etc. P2P applications: Bittorrent ICT is one of the fastest growing sector in terms of energy consumption 3

4 Implementation Results Conclusions Today s Network Status and Trend (cont ) Energy-consumption of cloud services [7] Download per hour 4

5 Implementation Results Conclusions Why Going Green? ICT has been considered as a key objective to reduce energy wastes and achieve higher levels of efficiency Until recently, ICT has not applied the efficiency concepts, not even in fast growing sectors like telecommunications, datacenters and the Internet Two main motivations for green ICT: Environmentally, it is related to the reduction of wastes, in order to impact on CO 2 emission; Economically, it stems from the reduction of operating costs (OPEX) of ICT services. 5

6 Power consumption Energy per transmitted bit Motivations Implementation Results Conclusions Gaps Between Theory and Practice ideal Traffic load (b/s) Traffic load (b/s) How to make energy consumption proportionally to network load? How to reduce energy consumption per transmitted bit? 6

7 Probability Motivations Implementation Results Conclusions And the Reasons? Internet is designed to be extremely overdimensioned and available 24/7. Links and devices are provisioned for rush hour load. The overall power consumption remains more or less constant even in the presence of fluctuating traffic loads. 40% 35% 30% Total Telit GrNet Nask 25% 20% 15% 10% 5% 0% 0% 5% 10% 15% 20% 25% 30% 35% 40% 45% 50% 55% 60% 65% 70% 75% 80% 85% 90% 95% 100% Offered Load Distribution of offered load Power Consumption Source: Franco Davoli, Green Networking: A Tutorial, ICCE

8 Implementation Results Conclusions Energy Consumption in Data Centers 5% focus 1% 1% 1 2 Cooling 33% Damp absorption 3% Air-conditioned PC room 9% IT equipments 30% PDU 5% UPS 18% Electricity generator 1% Lighting 1% 8

9 Implementation Results Conclusions Facts E-commerce website: 292 production webservers over 5 days. Traffic varies by day/ weekend, power doesn't. 9

10 Implementation Results Conclusions Energy Traffic Energy proportional to traffic load 10

11 Implementation Results Conclusions Main Activities Defining suited network architecture and network management Defining energy-aware methods and algorithms Building testbed and evaluation methods 11

12 Implementation Results Conclusions Network Architecture Requirements for the new DC architecture: Flexibility: new schemes and mechanisms in support of energy awareness should easily be integrated, managed and investigated. Controllability: more control abilities on routing, flow classification, load balancing, changing switch clock frequencies, on/off switch, migrating servers etc. Deployment of Software-Defined Networking (SDN) 12

13 Pod 1 Pod 2 Motivations Implementation Results Conclusions Network Architecture (cont ) OpenFlow ( SDN technology developed by Stanford [2] OpenFlow-based FatTree architecture for data center networks [1] Core OpenFlow Controller Aggregation Traffic and Power Monitoring Optimizer Top-of-Rack Power Control Servers Energy-aware routing 13

14 Implementation Results Conclusions Energy-Efficient Algorithms in Fat Tree-based DC (1/2) k-port switch k Performance Optimized Data Center (POD) 2 k servers/pod 2 4 æ k ö ç k-port core switches è 2 ø k Support 3 servers in the whole network 4 2 k 4 14

15 Implementation Results Conclusions Energy-Efficient Algorithms (2/2) Proposed Rate-Adaptive Topology-Aware Heuristic (RATAH) algorithm: combination of Smart sleeping: Optimizing actual topology by turning off unused switches and servers based on current traffic load Re-route traffic on the actual topology Load balancing and resilient algorithms to maintain certain QoS Dynamic adaptation: Adaptively changing clock frequencies of switches or network interfaces to optimize energy consumption 15

16 Implementation Results Conclusions Smart Sleeping vs. Power Scaling Standard operations Smart sleeping Wakeup and sleeping times Power scaling Increased service times Smart sleeping + power scaling Wakeup and sleeping Increased service times 16

17 Implementation Results Conclusions Example 1Gbps Full Fat-tree 10Mbps 1Gbps Elastic tree with power scaling 17

18 Implementation Results Conclusions RATAH Algorithm Begin Read in/out traffic statistic on switch s port using Monitoring component Limit Bandwitdh 10 Mbps No No 0 T uth. 10 uth. R10 T uth.r100 uth. R100 T uth.r1000 u. R R th Yes Yes Yes Limit Bandwitdh 100 Mbps Limit Bandwitdh 1 Gbps No Calculate number of links and switchs need to be turned on by using TAH On/Off switchs Caculate saved power 18

19 Implementation Results Conclusions Algorithm Analysis If traffic utilization is low enough to maintain a Minimum Spanning Tree, then the energy saving of RATAH is upper bounded by: S RA k TAH k 1 k ( k 2 2k) P 4 If traffic utilization is high enough to set up a fully meshed Fat-Tree topology then the energy saving of the RATAH algorithm is upper bounded by: 3 S RATAH ( k 4k 8k)( ) P 2 19

20 Implementation Results Conclusions Algorithm Analysis (cont ) Low traffic case: (Minimum spanning tree) (K=50 ~ servers) 20

21 Implementation Results Conclusions Algorithm Analysis (cont ) High traffic case: (Full fat tree) (K=50 ~ servers) 21

22 Implementation Results Conclusions Algorithm Analysis (cont ) Some remarks: In case of low traffic utilization (minimum spanning tree): Smart sleeping can save up to 60% of energy (for DC with 2000 servers) Power scaling can save only 2% more In case high traffic utilization (full fat tree): Power scaling is more advantageous with 15% of energy saving for DC with 2000 servers. Smart sleeping saves 0% of energy It is beneficial to combine these two mechanisms for better energy utilization 22

23 Source: [8,9] Truong Thu Huong, et al.,"ecodane Reducing Energy Consumption in Data Center Networks based on Traffic Engineering, EuroView2011, Würzburg, Germany Motivations Implementation Results Conclusions Testbed Architecture Optimizer Calculate optimal topology based on current traffic and energy conditions Subset Routing Concentrate traffic on a minimum number of link Traffic state Bit rate, packet rate per port etc. Monitoring Power state Port & switch power consump-tion Topology Link/port/ switch state Power Control Adjust link, port, switch state Port/linecard/box on/off Flow routes. NOX Traffic stat. Virtual Testbed: Mininet Data Center Network (switches, links) Traffic Generators lognormal IAT and pkt. length with average rates based on real traces Mininet OpenFlow Node Realistic energy model Real data center network Measurement Traffic modeling & characterization Energy measurement and modeling Hardware testbed based on NetFPGA 23

24 Implementation Results Conclusions Testbed Architecture ECODANE testbed: Hybrid real testbed and emulation Mininet-based emulation environment [4] OpenFlow-based testbed: NOX as OpenFlow Controller [3] Real NetFPGA OpenFlow switches [9] Software traffic generator with real DC traffic patterns: D-ITG [5] NOX Controller Minninet Traffic Generator D-ITG 24

25 Implementation Results Conclusions Test Environment (cont ) 25

26 Implementation Results Conclusions Test Scenarios Traffic pattern: lognormal inter-arrival packet distribution (top-of-rack traffic) Traffic load: various load conditions Size: K=4 ~ 16 servers K=6 ~ 54 servers K=8 ~ 128 servers Test cases: Near traffic Mid traffic Far traffic Mixed traffic Algorithms: Energy efficiency and optimization: RATAH (proposed Rate-Adaptive Topo- Aware Heuristic algorithm) Routing: ECMP 26

27 Implementation Results Conclusions Traffic Scenarios Near Traffic traffic within edge switches Middle Traffic - Traffic within PODs Far Traffic - inter-pod traffic 27

28 Implementation Results Conclusions Power Saving vs. Number of Active Components 60% 50% 40% 30% 20% 10% k=6 k=4 k=8 Power saving level 0% Near Traffic 1 Near Traffic 2 Middle Traffic Far Traffic Mix Traffic k=4 k=6 k=8 Number of active components 0 Near Traffic 1 Near Traffic 2 Middle Traffic Far Traffic Mix Traffic 28

29 Network power consumption (%) Network Utilization (%) Network power consumption (%) Network Utilization (%) Motivations Implementation Results Conclusions Energy Proportionality k=6 Near %Power %Power LSA %Pmax Network Utilization Time (s) Middle %Power %Power LSA %Pmax Network Utilization Time (s) 29

30 Network power consumption (%) Network Utilization (%) Network power consumption (%) Network Utilization (%) Motivations Implementation Results Conclusions Energy Proportionality k=6 (cont ) Far %Power %Power LSA %Pmax Network Utilization Time (s) Mix %Power %Power LSA %Pmax Network Utilization Time (s) 30

31 Implementation Results Conclusions Our contributions: Conclusions Propose combined smart sleep and dynamic adaptation algorithm to improve energy proportionality. Introduce a novel testbed platform for the indepth analysis of energy-aware data centre networks based on OpenFlow technology. 31

32 Implementation Results Conclusions Thank you! 32

33 Implementation Results Conclusions References [1]. Elastic tree: Saving energy in Data Center Networks. B. Heller, S. Seetharaman, P. Mahadevan, Y. Yiakoumis, P. Sharma, S. Banerjee, N. McKeown. USENIX NSDI, April, 2010 [2]. OpenFlow: Enabling Innovation in Campus Networks. Nick McKeown, Guru Parulkar, Tom Anderson, Larry Peterson, Hari Balakrishnan, Jennifer Rexford, Scott Shenker, Jonathan Turner, March 14, [3]. [4]. [5]. [6] Raffaele Bolla et al., Energy Efficiency in the Future Internet: A Survey of Existing and Trends in Energy-Aware Fixed Network Infrastructures, IEEE Communications Survey and Tutorials, Second Quarter 2011 [7] J. Baliga, R. W. A. Ayre, K. Hinton, R.S. Tucker, Green Cloud Computing: Balancing Energy in Processing, Storage, and Transport, Proc. IEEE, vol. 99, no. 1, pp , Jan [8] Nguyen Huu Thanh, et al. "Enabling Experiments for Energy-Efficient Data Center Networks on OpenFlow-based Platform", ICCE 2012, Hue, Vietnam [9] Tran Hoang Vu et al., Power Aware OpenFlow Switch Extension for Energy Saving in Data Centers, in ATC 2012, Hanoi Vietnam [10] Truong Thu Huong, et al.,"ecodane Reducing Energy Consumption in Data Center Networks based on Traffic Engineering, in EuroView2011, August 1st - 2nd 2011, Würzburg, Germany 33