Time-Domain Approach to Energy: Network Elements in Design If printed, please share or recycle this slide deck after use dkh@juniper.net 1
Energy Efficiency Basics Energy Efficiency (def.) : Using less energy to provide the same service (source: Wikipedia) Network devices provide service in form of moving data Therefore, network energy efficiency is a relation between effective throughput (in bits) and consumed energy (in Joules) This relation exists at various levels network element, network layer, end-to-end network service For more information, see ECR initiative (www.ecrinitiative.org) 2
Network Element Level SECTION TITLE WITH IMAGE 3
Energy Efficiency (Network Element) Measuring energy efficiency at the element (device) level requires amp meters and test profile from packet generators. System under test (SUT) forwards test load and consumes energy, while being actively measured by test equipment. Note: system subcomponents (ie power supplies) can have their own efficiency targets; We do not measure or report them separately because they are included in the system (network device) Question: What affects system energy efficiency? 4
Energy Efficiency vs Operations per Second On platforms with hardware forwarding planes, packet rate and packet policy together determine the number of lookup operations per second required to maintain line-rate forwarding ECR, W/Gbps Source: T1600 ECR stats for IPv4 lookups, 23x effective lookup speed change from left to right Lookup speed does not affect efficiency significantly 5
Element Efficiency as function of chassis fill On modular platforms, common infrastructure (power, fabric, control plane) is shared ECR, W/Gbps Source: MX960 chassis measured for progressively less linecards installed (11 down to 1) Traffic is forwarded at 100% line rate (256B packet size) across all linecards in the test Chassis utilization below 30% significantly affects ECR 6
Element Efficiency as function of utilization In the field, a network device may have variable utilization levels, dependent on service type and network activity. ECR, W/Gbps Source: T1600 ECR vs effective utilization (offered traffic load varies from 100 to 5 percent) Traffic utilization below 40% significantly affects ECR 7
Element Efficiency as function of device role On platforms with hardware forwarding planes, packet size and packet policy determines the number of lookup operations per second (in Mpps) ECR, W/Gbps Sample efficiency ratings for various network product classes (projected data) Efficiency improves with scale and lower packet touch 8
Element Efficiency as function of technology M40 M160 T640 T1600 Year 1998 2000 2002 2007 Slot Bandwidth 3Gbps 10Gbps 40Gbps 100Gbps (full-duplex) System Throughput 40Gbps 160Gbps 640Gbps 1600Gbps Fabrication 180nm 180nm 130nm 90nm ECR, Watts/Gbps 76.9 33.3 14.08 9.34 Notes World s 1 st 40G router World s 1 st 10G/slot World s 1st 40G/slot World s 1 st 100G/slot Historical improvements in parallel with silicon progress (Dennard s Scaling Theory) 9
Current diversity in efficiency is ~50% Competing vendors with independent technology sets often split ways in energy efficiency. Juniper case: custom ASICs (no RISC arrays), TCAM-less services, custom fabric crossbar But traffic growth still outruns projected energy limits 10
Basic idea: Slowdown at low activity Typical link-layer utilization (source: Portland State) Example: IEEE 802.3az Slowdown trades energy consumption for delay. In most cases, this requires additional hardware/buffering. 11
Is link-level energy saving enough? Need to look much deeper into the system 12
Typical carrier equipment infrastructure Linecard breakdown 7.55 Watts/ Gbps 4.61 Watts/Gbps C1.1: Core C3.1: Carrier Ethernet C3.3: Desktop Ethernet Identifying places to save energy 13
Application Tolerance to system events MC = mission critical apps, BFD = control plane liveness protocol Most system transitions are N/A for carrier networks 14
Crossing delay boundaries: is there hope? Slow-scale network events Packet loss is inevitable, so enabling slow energy modes is not a technical but administrative challenge 15
Conclusions: In production networks, it is not sufficient to classify energy-saving algorithms merely in sleep and slowdown categories, it makes much more sense to use delay boundaries to rate technoogy against impact Different time domains dictate very different technical solutions; technologies for peak, variable and slowspeed energy optimization are orthogonal to each other Slow energy transition technologies are still the lowest hanging fruit for equipment vendors; the drawback to poor man's energy savings is administrative responsibility for risk of service degradation 16
17 Thank You For Saving Energy