Internet energy efficiency

From WiL

There is much research to be done to reduce energy use by the internet. This page is a collection of preliminary thoughts on both reducing (energy use by the internet) and (reducing energy use) by the internet.

Contents 1 Trends in energy use by the internet 2 The numbers 3 Reducing the internet's energy consumptions 3.1 Reduce use 3.2 Low-power states 3.3 Low-power electronics 3.4 Life cycle costs 3.5 Parts of the network 3.5.1 Network core 3.5.2 Access network: wired vs wireless 3.5.3 Data centres 3.5.4 End users 4 Using the Internet to reduce energy use 5 Current players 5.1 Research 5.2 Industry 6 Papers 6.1 Network-wide 6.2 Interfaces 6.3 Routers 6.3.1 Switching fabrics 6.3.2 Other 6.4 Servers 7 Open questions

[edit] Trends in energy use by the internet

• Big and getting bigger
  • Cellular network infrastructure doubling every 4-5 years [1]
• Google is placing data centres near power stations
• Cooling is limiting data centre density

[edit] The numbers

• European commission estimates that home broadband will consume 50TWh per year by 2012 [2]
• Cellular network infrastructure used 60TWh in 2007, of which 80% is for the radio access network [3]
• Energy is 2-3% of a telco's opex ($8e9 in Europe/Middle-East/Africa, $6e9 in Asia-Pacific)[4]. Telecoms is 1% of greenhouse gas emission.
• AIST estimates routers would be 9% of Japan's electricity by 2015, if no technology changes. NTT estimates PON CPE takes four times the power of the network itself [5]
• Server power supplies typically 70% efficient in June 2006 [6]
• Data centers used 61 billion kilowatt-hours in 2006, or 1.5% of all power consumed in the United States. The cost: $4.5 billion, or about as much as was spent by 5.8 million average households. [7]. (About 30e9 kWh in 2001.) Set to double by 2011 unless action is taken
• Hubs, switches, and routers consumed 6.15 TWh/yr in 1999. The difference between 100baseT and GbE is about 4W [8].
• AC units can draw as much as 60% of the power required by the systems they're meant to cool [9]
• Well designed data centres can operate at 78F (26C); most operate below 70F (21C). A few percentage points of energy use per degree.

[edit] Reducing the internet's energy consumptions

[edit] Reduce use

• Reduce spam
• Improve efficiency of routing

[edit] Low-power states

• Put interfaces into low-rate modes when load is low
  • IEEE 802.3az
  • Protocols able to tolerate rapid large changes in BDP
  • (Does GbE use power control? Does it continually send a carrier?)
• Put circuitry into sleep mode when not needed
  • memory, main processor, bus, line card processor/ASIC, and the switching fabric. [10]
  • Current memory can go into low-power mode. Need memory management which leaves entire banks of memory idle.
    • Intel speedstep already dynamically adjusts L2 cache "size"
  • Prediction/coordination of how long to sleep. Must exploit the heavy tailed distribution of interarrival times.
• Put entire devices into sleep mode when not needed
  • Protocols able to tolerate sporadic connectivity
  • Human factors encouraging users to turn off end systems.
• models of speed vs power
• Suspend to RAM can be made more efficient using custom BIOS which boots in 3 seconds.

[edit] Low-power electronics

• Generic low-voltage / low-power circuitry
• Heat recovery using the thermoelectric effect
• Systems that can operate at higher temperatures reduce need for air conditioning
• Increased parallelism with reduced clock rates
• Using ASICs or one-time FPGAs instead of SRAM-based FPGAs

[edit] Life cycle costs

• Turning off servers to save energy reduces MTBF.
• Use of simpler devices
• Recyclability of electronics
  • Re-use of cases and mounting hardware

[edit] Parts of the network

[edit] Network core

• Routers: Ball-park breakdown of power consumed by a router. (Photonic Switching and the Energy Bottleneck J. Baliga, R. Ayre, K. Hinton and R. S. Tucker, IEEE Photonics in Switching 2007.

Subsystem	Percentage
Power supply and fans	35
Forwarding engine	33.5
Switching Fabric	10
Control plane	11
I/O	7
Buffers	3.5

[edit] Access network: wired vs wireless

• Wired links typically use less power
• For sparse areas, the energy cost of installing wire may outweigh this
• The energy cost for satellite "installation" is huge.

[edit] Data centres

• Work on low power interconnects, speed scaling (DVS) and scheduling to allow servers to power down
• Processors themselves: A lot of work on low power network-on-chip architectures. Can this be applied to local/wide area networks?

[edit] End users

[edit] Using the Internet to reduce energy use

• Increased telecommuting / teleconferencing
• Does on-line shopping reduce total energy for shipping / warehousing / HVAC etc?

[edit] Current players

[edit] Research

• Berkeley Labs' Environmental Energy Technologies Division, 802.3az
  • Mike Bennett, EEE Chair: mjbennett at lbl.gov
  • Bruce Nordman (BNordman at lbl.gov) and Alan Meier (AKMeier at lbl.gov) (Draft Principles for Energy Efficient Digital Networks and Network-connected Devices)
• Rod Tucker / CUBIN
• Global Research into Environmentally Efficient Networking (GREEN) at Swinburne University of Technology, Australia
• Nick Bambos -- power aware packet scheduling, tradeoff of delay vs low-power modes
• Hugh Barrass of Cisco, Howard Frazier of Broadcom (adaptive link rate [11])
• http://www.cs.pitt.edu/PARTS/
• Thousands of wireless / sensor network researchers

[edit] Industry

• Green ISP Solar powered offices
• web hosting
• Phone co-op
• IBM's [12] Calibrated vector cooling, and ExecutivePower, flash memory instead of HDD
• Nokia's energy efficient base stations (GSM BS typically 800W, WCDMA typically 500W. By 2010, 650W and 300W respectively)
• http://www.green500.org/ List of 500 most energy-efficient supercomputers

[edit] Papers

[edit] Network-wide

• Greening the Internet, M. Gupta and S. Singh, (Portland State University) SIGCOMM 2003
  • Rerouting to allow routers to sleep. What impact on higher-layer protocols?
• Enabling an Energy-Efficient Future Internet Through Selectively Connected End Systems, Mark Allman, Ken Christensen, Bruce Nordman, Vern Paxson, HotNets'07
  • Position paper. Architectural changes: DTN-style intermittent connectivity, but without delays.
• [?? Dynamic Power Management for Power Optimization of Interconnection Networks Using On/Off Links], Vassos Soteriou and Li-Shiuan Peh, (Princeton) HOTI’03
  • Power-down some links on lattice interconnection network (not Internet). New routing alg.
  • Decision on buffer occupancy over a window. Considers time (but not power) used for on/off transition
  • Simulation. Up to 37.5% improvement.

[edit] Interfaces

• Ethernet Adaptive Link Rate (ALR): Analysis of a Buffer Threshold Policy C Gunaratne, K Christensen, S Suen. Uni South Florida. GLOBECOM 2006
  • ALR. Markov analysis for bi-level hysteretic switching. Compares with data from traces. (Assumes arrival process would be unchanged, neglecting TCP.)
• Reducing Network Energy Consumption via Rate-Adaptation and Sleeping Sergiu Nedevschi, Lucian Popa, Gianluca Iannaccone, Sylvia Ratnasamy, David Wetherall (Berkeley, Intel) Technical Report No. UCB/EECS-2007-128
  • For given static/dynamic power, when is it better to sleep than go low rate?
• A Feasibility Study for Power Management in LAN Switches, M. Gupta, S. Grover, S. Singh, ICNP'04
  • LAN traffic measurements, simulation of different levels of sleep. Model includes time/power for on/off transitions
• Dynamic Ethernet Link Shutdown for Energy Conservation on Ethernet Links, M. Gupta, S. Singh, ICC'07
  • Simulation of link shutdown with long-range dependent traffic
• Power Awareness in Network Design and Routing, Joseph Chabarek, Joel Sommers, Paul Barford, Cristian Estan, David Tsiang, Steve Wright, INFOCOM'08
  • Measurement of router power consumption.
  • Integer programming to minimise number of cards / chassis needing to be active to achieve given capacity and connectivity
• The next frontier for communications networks: power management, Kenneth J. Christensen, Chamara Gunaratne, Bruce Nordman, Alan D. George, Comp. Commun '04
  • Always-on NIC Proxy handles DHCP etc, allowing host PC to sleep.

[edit] Routers

[edit] Switching fabrics

• Power Aware Management of Packet Switches Lykomidis Mastroleon, Daniel O'Neill, Benjamin Yolken and Nick Bambos (Stanford University) High-Performance Interconnects, 2007
  • Uses linear quadratic regulation to clear current backlog. (x² power).
  • Needn't starve flows, as it optimizes after an arbitrary matching rule.

• Energy-Efficient Scheme for Multiprocessor-Based Router Linecards M. Mandviwalla, Nian-Feng Tzeng Louisiana Univ. SAINT 2006.
  • All processors in a multiprocessor should run at same DVS voltage. (Intuitive, as power is convex.)

• Analysis of Power Consumption on Switch Fabrics in Network Routers TT Ye, L Benini, G De Micheli - Design Automation Conference 2002. Stanford/UBologna.
  • Models of energy use by cross-bar, fully connected, banyan and Batcher-banyan networks. Not sure how plausible the models are. Most are linear with throughput, but Banyan is superlinear.

• Optical switching technology comparison: optical MEMS vs. other technologies Xiaohua Ma, Geng-Sheng Kuo
  • Overall comparison of optical technologies, passing mention of energy use.

[edit] Other

• Low-Power Network-Packet-Processing Architecture Using Process-Learning Cache for High-End Backbone Router Michitaka OKUNO, Shin-ichi ISHIDA Hiroaki NISHI, IEICE TRANS. ELECTRON., VOL.E88–C, NO.4 APRIL 2005
  • Implements "fast path" packet processing by caching table look-ups etc.
• An Ultra High Throughput and Power Efficient TCAM-Based IP Lookup Engine Kai Zheng, Chengchen Hu, Hongbin Lu, Bin Liu, Tsinghua, INFOCOM 2004
  • Minimizes number of TCAM entries triggered on each lookup

[edit] Servers

• Kirk Pruhs has done lots of theoretical work on speed scheduling.
• Sandy Irani has worked on online algorithms for sleep schedules
• Analysis of Energy Reduction on Dynamic Voltage Scaling-Enabled Systems
  • Limited slew rate of speed. lpARM can continue operating during the 25us taken to change speeds
• Intraprogram dynamic voltage scaling: Bounding opportunities with analytic modeling FEN XIE, MARGARET MARTONOSI, and SHARAD MALIK, Princeton University
  • 12us switching time, 1.2uJ switching cost. Depends on regulator capacitor
• Managing Server Energy and Operational Costs in Hosting Centers Y. Chen, A. Das, W. Qin, A. Sivasubramaniam, Q. Wang, N. Gautam, Penn. State University, SIGMETRICS 2005.
  • Balances lower rates vs turning off. Estimates delay using G/G/m model based on time-varying predicted values, and optimizes heuristically. Also heuristically designs LQG controller. (x³ power)
• SLA based profit maximization in multi-tier systems D. Ardagna, M. Trubian and Li Zhang, IEEE Int Conf. Network Computing and Applications, 2005.
  • On-off server model. Heuristic for solving NP-hard optimization problem offline. Markov model of number of jobs of each application type.
• A Hybrid Disk-Aware Spin-Down Algorithm with I/O Subsystem Support Timothy Bisson, Scott A. Brandt, Darrell D.E. Long, UC Santa Cruz, IPCCC'07.
  • 30% of general purpose computer power is HDD. ~70% in storage systems.
  • Algorithms using NV cache to minimize spin-up of disks, and associated wear.
• System-level power optimization: techniques and tools L Benini, G de Micheli - ACM Trans. Design Automation of Electronic Systems, 2000
• I. Hong, M. Potkonjak and M.Srivastava. "On-line scheduling of hard real-time tasks." Proceedings of International conference on computer-aided design, 1998
• T.Ishihara and H. Yasuura. "Voltage scheduling problem for dynamically variable voltage processors." Proceedings of International symposium on low power electronics and design, 1998
• A. Manzak and C. Chakrabarti. "Variable voltage task scheduling algorithms for minimizing energy." International Symposium on low power electronics and design, 2001

[edit] Open questions

• How does burstiness of TCP affect the (energy) effectiveness of caching of table lookups in routers?
• How much extra power is required to provide "QoS"? How does that compare with the extra power required for increase bitrate?
• What impact does the heavy-tailed packet inter-arrival time distribution have on the effectiveness of ALR?
• Can energy efficiency in RPRs with ALR be improved by scheduling isochronous traffic in contiguous "bursts", allowing rate reduction between bursts?